Upload folder using huggingface_hub

iclr26/6FyYLNnDL6/appendix_chunks.jsonl

@@ -0,0 +1,38 @@
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0068", "section": "[Instruction] Answer the question.", "page_start": 14, "page_end": 14, "type": "Text", "text": "IMPORTANT: Respond ONLY with the exact answer in the same format as the examples. Do NOT add any extra text, explanations, or punctuation. Do NOT include \"Answer:\" or any similar prefix in your response.", "source": "marker_v2", "marker_block_id": "/page/13/Text/7"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0069", "section": "[Examples]", "page_start": 14, "page_end": 14, "type": "Text", "text": "Question: Which major Russian city borders the body of water in which Saaremaa is located?", "source": "marker_v2", "marker_block_id": "/page/13/Text/9"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0070", "section": "[Examples]", "page_start": 14, "page_end": 14, "type": "Text", "text": "Answer: Saint Petersburg", "source": "marker_v2", "marker_block_id": "/page/13/Text/10"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0071", "section": "[Examples]", "page_start": 14, "page_end": 14, "type": "Text", "text": "Question: Who was thee first president of the association that wrote the code of ethics for psychology?", "source": "marker_v2", "marker_block_id": "/page/13/Text/11"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0072", "section": "[Examples]", "page_start": 14, "page_end": 14, "type": "Text", "text": "Answer: G. Stanley Hall", "source": "marker_v2", "marker_block_id": "/page/13/Text/12"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0073", "section": "[Examples]", "page_start": 14, "page_end": 14, "type": "Text", "text": "Question: Where did the Baldevins bryllup director die?", "source": "marker_v2", "marker_block_id": "/page/13/Text/13"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0074", "section": "[Examples]", "page_start": 14, "page_end": 14, "type": "Text", "text": "Answer: Copenhagen", "source": "marker_v2", "marker_block_id": "/page/13/Text/14"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0075", "section": "[Current Question]", "page_start": 14, "page_end": 14, "type": "Text", "text": "{ Summary of the retrieved documents }", "source": "marker_v2", "marker_block_id": "/page/13/Text/16"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0076", "section": "[Current Question]", "page_start": 14, "page_end": 14, "type": "Text", "text": "Question: { Question }", "source": "marker_v2", "marker_block_id": "/page/13/Text/17"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0077", "section": "[Current Question]", "page_start": 14, "page_end": 14, "type": "Text", "text": "Answer:", "source": "marker_v2", "marker_block_id": "/page/13/Text/18"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0078", "section": "[Current Question]", "page_start": 14, "page_end": 14, "type": "Caption", "text": "Figure 6: Prompt template for LLM QA.", "source": "marker_v2", "marker_block_id": "/page/13/Caption/19"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0079", "section": "A USE OF LLMS", "page_start": 14, "page_end": 14, "type": "Text", "text": "Large language models (LLMs) were employed solely as writing assistants to enhance the language, improve clarity, and check grammatical correctness. They were not used to generate research ideas, design or implement methodologies, conduct data analysis, or produce any of the results presented in this work. The authors assume full responsibility for the entire content of the paper.", "source": "marker_v2", "marker_block_id": "/page/13/Text/21"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0080", "section": "B PROMPT TEMPLATES", "page_start": 14, "page_end": 14, "type": "Text", "text": "Figure 5 displays the prompt employed to train the compressor model for generating a summary of the retrieved documents, conditioned on the given query. This prompt is notably concise. For end-task answer generation, the prompt provided to the LLM M is illustrated in Figure 6, which incorporates few-shot in-context examples, the (generated summary of) retrieved documents, and the question.", "source": "marker_v2", "marker_block_id": "/page/13/Text/23"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0081", "section": "C CASE STUDY", "page_start": 14, "page_end": 14, "type": "Text", "text": "To conduct an in-depth analysis of the advantages of our compressor, we performed case studies on one single-hop QA dataset (NQ) and one multi-hop QA dataset (2Wiki), with the results presented in Table 4 and Table 5, respectively. For each case, we compared the summaries generated by off-the-shelf Qwen2.5-1.5B-Instruct, RECOMP, and our method CORE based on the same set of documents, as well as the predicted answers generated by the LLM after prepending these summaries. As shown in the tables, although the summaries produced by off-the-shelf Qwen2.5-1.5B", "source": "marker_v2", "marker_block_id": "/page/13/Text/25"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0082", "section": "Top-5 documents", "page_start": 15, "page_end": 15, "type": "Text", "text": "last men conscripted, who were born in 1952 and who reported for duty in June 1973. On February 2, 1972, a drawing was held to determine draft priority numbers for men born in 1953, but in early 1973 it was announced by Secretary of Defense Melvin Laird that no further draft orders would be issued. In March 1973, 1974, and 1975, the Selective Service assigned draft priority numbers for all men born in 1954, 1955, and 1956, in case the draft was extended, but it never was. Command Sergeant Major Jeff Mellinger, believed to be the last drafted enlisted ranked.", "source": "marker_v2", "marker_block_id": "/page/14/Text/6"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0083", "section": "Top-5 documents", "page_start": 15, "page_end": 15, "type": "Text", "text": "The Gates Commission issued its report in February 1970, describing how adequate military strength could be maintained without having conscription. The existing draft law was expiring at the end of June 1971, but the Department of Defense and Nixon administration decided the draft needed to continue for at least some time. In February 1971, the administration requested of Congress a two-year extension of the draft, to June 1973. Senatorial opponents of the war wanted to reduce this to a one-year extension, or eliminate the draft altogether, or tie the draft renewal to a timetable for troop withdrawal from Vietnam; Senator.", "source": "marker_v2", "marker_block_id": "/page/14/Text/7"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0084", "section": "Top-5 documents", "page_start": 15, "page_end": 15, "type": "Text", "text": "draft operated from 1940 until 1946 when further inductions were suspended, and its legislative authorization expired without further extension by Congress in 1947. During this time, more than 10 million men had been inducted into military service. However, the Selective Service System remained intact. Scattered opposition was encountered especially in the northern cities where some African-Americans protested the system. The tiny Nation of Islam was at the forefront, with many Black Muslims jailed for refusing the draft, and their leader Elijah Muhammed was sentenced to federal prison for 5 years for inciting draft resistance. Organized draft resistance also developed in.", "source": "marker_v2", "marker_block_id": "/page/14/Text/8"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0085", "section": "Top-5 documents", "page_start": 15, "page_end": 15, "type": "Text", "text": "was the first time any form of student deferment was used. During the Korean War a student carrying at least twelve semester hours was spared until the end of his current semester. The United States breathed easier with the Korean War Armistice on July 27, 1953; however, technology brought new promises and threats. U.S. air and nuclear power fueled the Eisenhower doctrine of \"massive retaliation\". This strategy demanded more machines and fewer foot soldiers, so the draft slipped to the back burner. However, the head of the SSS, Maj. Gen. Hershey, urged caution fearing the conflict looming in Vietnam. In.", "source": "marker_v2", "marker_block_id": "/page/14/Text/9"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0086", "section": "Top-5 documents", "page_start": 15, "page_end": 15, "type": "Text", "text": "policy of using the draft as force to compel \"voluntary\" enlistment was unique in U.S. history. Previous drafts had not aimed at encouraging individuals to sign up in order to gain preferential placement or less dangerous postings. However, the incremental buildup of Vietnam without a clear threat to the country bolstered this. Some estimates suggest conscription encompassed almost one-third of all eligible men during the period of 1965–69. This group represented those without exemption or resources to avoid military service. During the active combat phase, the possibility of avoiding combat by selecting their service and military specialty led as many.", "source": "marker_v2", "marker_block_id": "/page/14/Text/10"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0087", "section": "Top-5 documents", "page_start": 15, "page_end": 15, "type": "Table", "text": "Compressor Generated summary Predicted answer Off-the-shell Qwen2.5- 1.5B-Instruct The US stopped drafting for the Vietnam War in late 1972 after the draft expired at the end of June 1971. late 1972 RECOMP The U.S. stopped drafting for the Vietnam War after the Selective Service System was officially abolished in December 1972 due to growing public opposition and the expiration of the draft's legal authority in 1971. While the draft was formally ended, some individuals were still consripted into the military December 1972 Ours The U.S. stopped drafting for the Vietnam War in 1973 after announcing the decision by Secretary of Defense Melvin Laird earlier that year. Although the Selective Service System was later updated to include draft priority numbers, this decision marked the end of the draft's use for national service. 1973", "source": "marker_v2", "marker_block_id": "/page/14/Table/11"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0088", "section": "Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? ´ Gold answer: [Charles Armand Rene de La Tr ´ emoille] ´", "page_start": 16, "page_end": 16, "type": "Text", "text": "Top-5 documents:", "source": "marker_v2", "marker_block_id": "/page/15/Text/5"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0089", "section": "Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? ´ Gold answer: [Charles Armand Rene de La Tr ´ emoille] ´", "page_start": 16, "page_end": 16, "type": "Text", "text": "as at Versailles: he was brigadier of cavalry (January 1709), first gentleman of the King's chamber (June 1709), governor of Thouars (July 1709), and Marechal de camp (February 1719). His sister ´ Marie Armande Victoire de La Tremoille married Emmanuel Th ´ eodose de La Tour d'Auvergne. ´ On 13 April 1706 he married Marie-Madeleine Motier de La Fayette (1691–1717), the daughter of Rene-Armand, marquis de La Fayette and Marie-Madeleine de Marillac, and granddaughter of the author Marie-Madeleine Pioche de la Vergne, comtesse de la Fayette. They had one child, Charles Armand Rene de La Tr ´ emoille, born in 1708. Charles Louis Bretagne de La ´", "source": "marker_v2", "marker_block_id": "/page/15/Text/6"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0090", "section": "Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? ´ Gold answer: [Charles Armand Rene de La Tr ´ emoille] ´", "page_start": 16, "page_end": 16, "type": "Text", "text": "Charles Bretagne Marie de La Tremoille Charles Bretagne Marie de La Tr ´ emoille (24 March ´ 1764 – 10 November 1839), 8th duc de Thouars, was a French soldier and the son of Jean Bretagne Charles de La Tremoille and his wife, Marie-Maximilienne, princess of Salm-Kyrburg. La ´ Tremoille married Louise-Emmanuelle de Ch ´ atillon in 1781. She was a grand daughter of Louis ˆ Cesar de La Baume Le Blanc, the famous writer. The couple had one daughter: At the outbreak ´ of the French Revolution, he sided with King Louis XVI, as did the rest of his family. In 1789, La Tremoille and his parents ´", "source": "marker_v2", "marker_block_id": "/page/15/Text/7"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0091", "section": "Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? ´ Gold answer: [Charles Armand Rene de La Tr ´ emoille] ´", "page_start": 16, "page_end": 16, "type": "Text", "text": "Jean Bretagne Charles de La Tremoille Jean-Bretagne-Charles de La Tr ´ emoille (5 February 1737 ´ – 19 May 1792), 7th duc de Thouars, was a French soldier and the son of Charles Armand Rene´ de La Tremoille and his wife, Marie Hortense de La Tour d'Auvergne. La Tr ´ emoille began his ´ military career in the Seven Years' War as the head of the regiment of Aquitaine-Cavalry. He was seriously wounded and afterwards promoted to the rank of sergeant, and later of brigadier, of the armies of the king. During the French Revolution, La Tremoille remained faithful to King Louis ´ XVI. He emigrated in", "source": "marker_v2", "marker_block_id": "/page/15/Text/8"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0092", "section": "Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? ´ Gold answer: [Charles Armand Rene de La Tr ´ emoille] ´", "page_start": 16, "page_end": 16, "type": "Text", "text": "Charles Louis Bretagne de La Tremoille Charles Louis Bretagne de La Tr ´ emoille (15 March ´ 1683 – 9 October 1719), 6th duke of Thouars, was the son of Charles Belgique Hollande de La Tremoille and Madeleine de Cr ´ equy, daughter and heiress of Charles III de Cr ´ equy. He became ´ duke of Thouars upon his father's death in 1709. He was also count of Laval and Montfort and inherited the ancient French claim to the Kingdom of Naples. The Neapolitan claim gave the family the rank of \"princes etrangers\" at the French court. The Duke made a career in the ´ military as well", "source": "marker_v2", "marker_block_id": "/page/15/Text/9"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0093", "section": "Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? ´ Gold answer: [Charles Armand Rene de La Tr ´ emoille] ´", "page_start": 16, "page_end": 16, "type": "Text", "text": "de Saint-Didier. They had two daughters: Following her death, La Tremoille married for a third ´ time in 1830 to Valentine Eugenie Jos ´ ephine Walsh de Serrant. They had two children: La ´ Tremoille died shortly after his son's birth, in 1839. Charles Bretagne Marie de La Tr ´ emoille ´ Charles Bretagne Marie de La Tremoille (24 March 1764 – 10 November 1839), 8th duc de ´ Thouars, was a French soldier and the son of Jean Bretagne Charles de La Tremoille and his wife, ´ Marie-Maximilienne, princess of Salm-Kyrburg. La Tremoille married Louise-Emmanuelle de ´ Chatillon in 1781. She was a grand daughter of Louis C ˆ esar de ´", "source": "marker_v2", "marker_block_id": "/page/15/Text/10"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0094", "section": "Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? ´ Gold answer: [Charles Armand Rene de La Tr ´ emoille] ´", "page_start": 16, "page_end": 16, "type": "Table", "text": "Compressor Generated summary Predicted answer Off-the-shell Qwen2.5- 1.5B-Instruct Charles Bretagne Marie de La Tremoille's paternal grandfather ´ was Jean Bretagne Charles de La Tremoille. ´ Jean Bretagne Charles de La Tremoille ´ RECOMP Charles Bretagne Marie de La Tremoille's paternal grandfather ´ was Charles Louis Bretagne de La Tremoille, who was the son ´ of Charles Belgique Hollande de La Tremoille and Madeleine ´ de Crequy. This makes him the direct paternal grandfather of ´ Charles Bretagne Marie de La Tremoille. ´ Charles Louis Bretagne de La Tremoille ´ Ours ´ Charles Bretagne Marie de La Tremoille's paternal grandfather is Charles Armand Rene de La Tr ´ emoille, her father's father, the ´ 7th duc de Thouars. Charles Armand Rene´ de La Tremoille ´", "source": "marker_v2", "marker_block_id": "/page/15/Table/11"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0095", "section": "Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? ´ Gold answer: [Charles Armand Rene de La Tr ´ emoille] ´", "page_start": 17, "page_end": 17, "type": "TableGroup", "text": "Table 6: Failure Cases on NQ dataset. Question Generated Summary Predicted Answer Gold Answer who is the president of usa right now The current president of the USA is **Donald Trump**, who was elected on January 20, 2017, and is the 45th president. He is a businessman and reality TV personality, and his inauguration marked a shift in the country's political landscape. Joe Biden Donald Trump who has been ranked no. 1 in the latest football rankings announced by fifa The latest FIFA football rankings, announced in 2018, have ranked France as the top team, after winning the 2018 World Cup. This change was implemented using the Elo rating system, and the rankings were introduced to better reflect football team strengths. France Germany", "source": "marker_v2", "marker_block_id": "/page/16/TableGroup/346"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0096", "section": "Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? ´ Gold answer: [Charles Armand Rene de La Tr ´ emoille] ´", "page_start": 17, "page_end": 17, "type": "Text", "text": "are concise, they largely fail to capture key information relevant to answering the question. In contrast, RECOMP demonstrates better summarization capability but is prone to being overwhelmed by lengthy documents, resulting in misjudgments and even generating misleading information—such as the statement in Table 4: \" The U.S. stopped drafting for the Vietnam War after the Selective Service System was officially abolished in December 1972 \"—which leads the downstream LLM to produce the incorrect answer \"1972\". Our method, CORE, accurately extracts answer-critical information from lengthy documents, exemplified by the summary: \" The U.S. stopped drafting for the Vietnam War in 1973 after announcing the decision by Secretary of Defense Melvin Laird earlier that year \", thereby enabling the LLM to generate the correct answer \"1973\". This indicates that our compressor, trained with an end-task target-oriented reinforcement learning optimization strategy, can produce document summaries that are most helpful for answering the given question while effectively filtering out irrelevant information.", "source": "marker_v2", "marker_block_id": "/page/16/Text/3"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0097", "section": "Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? ´ Gold answer: [Charles Armand Rene de La Tr ´ emoille] ´", "page_start": 17, "page_end": 17, "type": "Text", "text": "To further understand the limitations of our approach, we present two failure cases from the NQ dataset where the model provided incorrect answers based on our generated summaries. As shown in Table 6, the first case reveals that although the summary contained the key information required for the correct answer, the downstream LLM still produced an error, potentially due to its over-reliance on parametric knowledge. In the second case, the summary itself omitted critical information needed to answer the question, which likely led to the incorrect response.", "source": "marker_v2", "marker_block_id": "/page/16/Text/4"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0098", "section": "D IMPACT OF DIFFERENT COMPRESSORS ON PERFORMANCE", "page_start": 17, "page_end": 17, "type": "Text", "text": "In our previous experiments, we employed Qwen2.5-1.5B as the initial model to train our compressor. In this section, we utilize two additional models—Llama3.2-1B and Llama3.2-3B—as starting points to train our compressor and the baseline compressor, respectively. The experimental results are presented in Table 7 and Table 8. As shown in the results, our method CORE continues to achieve lossless compression with both models, maintaining a high token compression ratio while exhibiting no performance degradation in terms of Exact Match (EM) and F1 score compared to uncompressed RAG. Furthermore, under both new model configurations, our approach consistently outperforms the baseline methods, indicating that its superiority is not dependent on a specific model architecture and thus demonstrates strong robustness.", "source": "marker_v2", "marker_block_id": "/page/16/Text/6"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0099", "section": "D IMPACT OF DIFFERENT COMPRESSORS ON PERFORMANCE", "page_start": 17, "page_end": 17, "type": "Text", "text": "We also observe that our method adheres to a form of scaling law: the compressor trained using the 3B model outperforms the one trained with the 1B model. Specifically, the 1B compressor improves performance by 1–4 EM points over the uncompressed baseline, while the 3B compressor yields gains of 3–9 EM points.", "source": "marker_v2", "marker_block_id": "/page/16/Text/7"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0100", "section": "D IMPACT OF DIFFERENT COMPRESSORS ON PERFORMANCE", "page_start": 18, "page_end": 18, "type": "TableGroup", "text": "Table 7: Open-domain QA results using Qwen2.5-14B-Instruct as the downstream LLM (M). The reported token counts represent the length of in-context documents, excluding few-shot examples. RECOMP and our method CORE are both trained using llama3.2-1B-Instruct . NQ TriviaQA HotpotQA 2WikiMultihopQA EM F1 # tok EM F1 # tok EM F1 # tok EM F1 # tok No Retrieval 0.2136 0.3097 0 0.5323 0.5998 0 0.2105 0.2948 0 0.2611 0.2951 0 RAG without compre ession Top1 Document 0.3446 0.4441 142 0.6082 0.6870 143 0.2920 0.3893 147 0.2679 0.3187 153 Top3 Documents 0.3778 0.4845 427 0.6260 0.7102 430 0.3164 0.4187 442 0.2789 0.3358 460 Top5 Documents 0.3803 0.4916 712 0.6410 0.7248 715 0.3299 0.4369 737 0.2964 0.3521 766 Top10 Documents 0.3867 0.5003 1425 0.6440 0.7292 1428 0.3395 0.4488 1471 0.3104 0.3675 1531 Compression of top 3 5 docs llama3.2-1B 0.3147 0.4227 64 0.5552 0.6415 60 0.2648 0.3639 58 0.2498 0.3003 61 Deepseek-V3 (671B) 0.3773 0.5039 54 0.6528 0.7433 51 0.3359 0.4483 48 0.2507 0.3031 45 RECOMP (1B) 0.3410 0.4655 57 0.6071 0.6880 48 0.2987 0.4121 49 0.3045 0.3653 33 CORE (1B) 0.3947 0.4923 47 0.6483 0.7287 43 0.3344 0.4454 45 0.3378 0.3969 34 Compression of top 10 docs (w vith the co mpresso r trained on top 5 d locs) llama3.2-1B 0.3141 0.4228 62 0.5651 0.6512 5 8 0.2663 0.3661 56 0.2493 0.3006 61 Deepseek-V3 (671B) 0.3779 0.5107 56 0.6529 0.7445 53 0.3462 0.4569 50 0.2900 0.3464 40 RECOMP (1B) 0.3421 0.4661 59 0.6095 0.6917 52 0.2982 0.4105 55 0.3072 0.3681 44 CORE (1B) 0.4033 0.5033 47 0.6521 0.7296 45 0.3412 0.4500 48 0.3586 0.4162 42", "source": "marker_v2", "marker_block_id": "/page/17/TableGroup/149"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0101", "section": "D IMPACT OF DIFFERENT COMPRESSORS ON PERFORMANCE", "page_start": 18, "page_end": 18, "type": "TableGroup", "text": "Table 8: Open-domain QA results using Qwen2.5-14B-Instruct as the downstream LLM (M). The reported token counts represent the length of in-context documents, excluding few-shot examples. RECOMP and our method CORE are both trained using llama3.2-3B-Instruct . NQ TriviaQA HotpotQA 2WikiMultihopQA EM F1 # tok EM F1 # tok EM F1 # tok EM F1 # tok No Retrieval 0.2136 0.3097 0 0.5323 0.5998 0 0.2105 0.2948 0 0.2611 0.2951 0 RAG without compre ession Top1 Document 0.3446 0.4441 142 0.6082 0.6870 143 0.2920 0.3893 147 0.2679 0.3187 153 Top3 Documents 0.3778 0.4845 427 0.6260 0.7102 430 0.3164 0.4187 442 0.2789 0.3358 460 Top5 Documents 0.3803 0.4916 712 0.6410 0.7248 715 0.3299 0.4369 737 0.2964 0.3521 766 Top10 Documents 0.3867 0.5003 1425 0.6440 0.7292 1428 0.3395 0.4488 1471 0.3104 0.3675 1531 Compression of top : 5 docs llama3.2-3B 0.3252 0.4334 60 0.5650 0.6521 59 0.2772 0.3809 58 0.2485 0.2995 60 Deepseek-V3 (671B) 0.3773 0.5039 54 0.6528 0.7433 51 0.3359 0.4483 48 0.2507 0.3031 45 RECOMP (3B) 0.3657 0.4912 55 0.6183 0.6920 47 0.3025 0.4238 52 0.3274 0.3806 42 CORE (3B) 0.4310 0.5234 32 0.6650 0.7306 38 0.3507 0.4736 51 0.3905 0.4474 40 Compression of top 10 docs (w vith the co mpresso or trained on top 5 a locs) llama3.2-3B 0.3318 0.4359 61 0.5720 0.6588 57 0.2791 0.3854 60 0.2491 0.3011 59 Deepseek-V3 (671B) 0.3779 0.5107 56 0.6529 0.7445 53 0.3462 0.4569 50 0.2900 0.3464 40 RECOMP (3B) 0.3682 0.4963 52 0.6205 0.6973 44 0.3077 0.4261 54 0.3312 0.3869 50 CORE (3B) 0.4526 0.5467 33 0.6736 0.7404 37 0.3693 0.4926 51 0.4071 0.4633 48", "source": "marker_v2", "marker_block_id": "/page/17/TableGroup/150"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0102", "section": "D IMPACT OF DIFFERENT COMPRESSORS ON PERFORMANCE", "page_start": 19, "page_end": 19, "type": "TableGroup", "text": "Table 9: Zero-Shot Evaluation on HotpotQA of Models Trained on Natural Questions. EM F1 #tok No Retrieval 21.05 29.48 0 Full Documents 32.99 43.69 737 BM25 24.18 35.73 71 NoiseFilter-IB 27.97 38.62 38 RECOMP 28.96 39.95 56 CORE 33.67 45.06 36 RECOMP-Transfer 26.68 37.29 58 CORE-Transfer 31.25 42.84 35 Table 10: Evaluation on Noisy Natural Questions.", "source": "marker_v2", "marker_block_id": "/page/18/TableGroup/304"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0103", "section": "D IMPACT OF DIFFERENT COMPRESSORS ON PERFORMANCE", "page_start": 19, "page_end": 19, "type": "Table", "text": "EM F1 #tok full documents 35.21 45.38 1427 RECOMP 33.29 43.90 59 CORE 38.19 48.85 48", "source": "marker_v2", "marker_block_id": "/page/18/Table/4"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0104", "section": "E CROSS-DATASET GENERALIZATION PERFORMANCE", "page_start": 19, "page_end": 19, "type": "Text", "text": "To verify the universal compression capability of our method, we directly transfer the model trained on the single-hop question answering dataset NQ to the multi-hop dataset HotpotQA for evaluation. The results, shown in Table 9, indicate that our transferred model achieves nearly lossless performance compared to using full documents without compression, while substantially outperforming the transfer results of the RECOMP baseline. Moreover, although both our method and the baseline underperform relative to models trained directly on the target HotpotQA dataset, our approach exhibits a smaller performance drop and demonstrates greater robustness compared to the baseline.", "source": "marker_v2", "marker_block_id": "/page/18/Text/6"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0105", "section": "F ROBUSTNESS AGAINST NOISY CONTEXTS", "page_start": 19, "page_end": 19, "type": "Text", "text": "To evaluate the robustness of our approach against adversarial retrievals and noisy contexts, we constructed a noisy version of the NQ dataset. For each question, we constructed the input context by combining the top-3 passages retrieved by the DPR retriever with 7 randomly selected passages from the Wikipedia corpus to serve as irrelevant/noisy information. This resulted in a context of 10 passages, which were then shuffled to randomize the order. We then compared the performance of our method against the full-document baseline. Experimental results are presented in the table 10. In the \"full documents\" setting, the downstream LLM directly uses all these 10 passages to answer the question, whereas in our method, the compressor first summarizes the context, and the LLM then generates an answer based on the compressed content. The model we used was trained on the standard NQ dataset without any such noise augmentation. Our method not only matches but slightly surpasses the performance of using all documents, demonstrating its strong noise resistance and ability to extract key information from cluttered contexts. In addition, we compared our approach with the RECOMP baseline, and our method consistently outperforms it, reaffirming the superior compression capability and robustness of our model. Furthermore, our method achieves a high compression rate, condensing the source content from 1,427 tokens to just 48.", "source": "marker_v2", "marker_block_id": "/page/18/Text/8"}

iclr26/6FyYLNnDL6/appendix_text_v3.txt

@@ -0,0 +1,113 @@
+[p. 14 | section: [Instruction] Answer the question. | type: Text]
+IMPORTANT: Respond ONLY with the exact answer in the same format as the examples. Do NOT add any extra text, explanations, or punctuation. Do NOT include "Answer:" or any similar prefix in your response.
+
+[p. 14 | section: [Examples] | type: Text]
+Question: Which major Russian city borders the body of water in which Saaremaa is located?
+
+[p. 14 | section: [Examples] | type: Text]
+Answer: Saint Petersburg
+
+[p. 14 | section: [Examples] | type: Text]
+Question: Who was thee first president of the association that wrote the code of ethics for psychology?
+
+[p. 14 | section: [Examples] | type: Text]
+Answer: G. Stanley Hall
+
+[p. 14 | section: [Examples] | type: Text]
+Question: Where did the Baldevins bryllup director die?
+
+[p. 14 | section: [Examples] | type: Text]
+Answer: Copenhagen
+
+[p. 14 | section: [Current Question] | type: Text]
+{ Summary of the retrieved documents }
+
+[p. 14 | section: [Current Question] | type: Text]
+Question: { Question }
+
+[p. 14 | section: [Current Question] | type: Text]
+Answer:
+
+[p. 14 | section: [Current Question] | type: Caption]
+Figure 6: Prompt template for LLM QA.
+
+[p. 14 | section: A USE OF LLMS | type: Text]
+Large language models (LLMs) were employed solely as writing assistants to enhance the language, improve clarity, and check grammatical correctness. They were not used to generate research ideas, design or implement methodologies, conduct data analysis, or produce any of the results presented in this work. The authors assume full responsibility for the entire content of the paper.
+
+[p. 14 | section: B PROMPT TEMPLATES | type: Text]
+Figure 5 displays the prompt employed to train the compressor model for generating a summary of the retrieved documents, conditioned on the given query. This prompt is notably concise. For end-task answer generation, the prompt provided to the LLM M is illustrated in Figure 6, which incorporates few-shot in-context examples, the (generated summary of) retrieved documents, and the question.
+
+[p. 14 | section: C CASE STUDY | type: Text]
+To conduct an in-depth analysis of the advantages of our compressor, we performed case studies on one single-hop QA dataset (NQ) and one multi-hop QA dataset (2Wiki), with the results presented in Table 4 and Table 5, respectively. For each case, we compared the summaries generated by off-the-shelf Qwen2.5-1.5B-Instruct, RECOMP, and our method CORE based on the same set of documents, as well as the predicted answers generated by the LLM after prepending these summaries. As shown in the tables, although the summaries produced by off-the-shelf Qwen2.5-1.5B
+
+[p. 15 | section: Top-5 documents | type: Text]
+last men conscripted, who were born in 1952 and who reported for duty in June 1973. On February 2, 1972, a drawing was held to determine draft priority numbers for men born in 1953, but in early 1973 it was announced by Secretary of Defense Melvin Laird that no further draft orders would be issued. In March 1973, 1974, and 1975, the Selective Service assigned draft priority numbers for all men born in 1954, 1955, and 1956, in case the draft was extended, but it never was. Command Sergeant Major Jeff Mellinger, believed to be the last drafted enlisted ranked.
+
+[p. 15 | section: Top-5 documents | type: Text]
+The Gates Commission issued its report in February 1970, describing how adequate military strength could be maintained without having conscription. The existing draft law was expiring at the end of June 1971, but the Department of Defense and Nixon administration decided the draft needed to continue for at least some time. In February 1971, the administration requested of Congress a two-year extension of the draft, to June 1973. Senatorial opponents of the war wanted to reduce this to a one-year extension, or eliminate the draft altogether, or tie the draft renewal to a timetable for troop withdrawal from Vietnam; Senator.
+
+[p. 15 | section: Top-5 documents | type: Text]
+draft operated from 1940 until 1946 when further inductions were suspended, and its legislative authorization expired without further extension by Congress in 1947. During this time, more than 10 million men had been inducted into military service. However, the Selective Service System remained intact. Scattered opposition was encountered especially in the northern cities where some African-Americans protested the system. The tiny Nation of Islam was at the forefront, with many Black Muslims jailed for refusing the draft, and their leader Elijah Muhammed was sentenced to federal prison for 5 years for inciting draft resistance. Organized draft resistance also developed in.
+
+[p. 15 | section: Top-5 documents | type: Text]
+was the first time any form of student deferment was used. During the Korean War a student carrying at least twelve semester hours was spared until the end of his current semester. The United States breathed easier with the Korean War Armistice on July 27, 1953; however, technology brought new promises and threats. U.S. air and nuclear power fueled the Eisenhower doctrine of "massive retaliation". This strategy demanded more machines and fewer foot soldiers, so the draft slipped to the back burner. However, the head of the SSS, Maj. Gen. Hershey, urged caution fearing the conflict looming in Vietnam. In.
+
+[p. 15 | section: Top-5 documents | type: Text]
+policy of using the draft as force to compel "voluntary" enlistment was unique in U.S. history. Previous drafts had not aimed at encouraging individuals to sign up in order to gain preferential placement or less dangerous postings. However, the incremental buildup of Vietnam without a clear threat to the country bolstered this. Some estimates suggest conscription encompassed almost one-third of all eligible men during the period of 1965–69. This group represented those without exemption or resources to avoid military service. During the active combat phase, the possibility of avoiding combat by selecting their service and military specialty led as many.
+
+[p. 15 | section: Top-5 documents | type: Table]
+Compressor Generated summary Predicted answer Off-the-shell Qwen2.5- 1.5B-Instruct The US stopped drafting for the Vietnam War in late 1972 after the draft expired at the end of June 1971. late 1972 RECOMP The U.S. stopped drafting for the Vietnam War after the Selective Service System was officially abolished in December 1972 due to growing public opposition and the expiration of the draft's legal authority in 1971. While the draft was formally ended, some individuals were still consripted into the military December 1972 Ours The U.S. stopped drafting for the Vietnam War in 1973 after announcing the decision by Secretary of Defense Melvin Laird earlier that year. Although the Selective Service System was later updated to include draft priority numbers, this decision marked the end of the draft's use for national service. 1973
+
+[p. 16 | section: Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? ´ Gold answer: [Charles Armand Rene de La Tr ´ emoille] ´ | type: Text]
+Top-5 documents:
+
+[p. 16 | section: Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? ´ Gold answer: [Charles Armand Rene de La Tr ´ emoille] ´ | type: Text]
+as at Versailles: he was brigadier of cavalry (January 1709), first gentleman of the King's chamber (June 1709), governor of Thouars (July 1709), and Marechal de camp (February 1719). His sister ´ Marie Armande Victoire de La Tremoille married Emmanuel Th ´ eodose de La Tour d'Auvergne. ´ On 13 April 1706 he married Marie-Madeleine Motier de La Fayette (1691–1717), the daughter of Rene-Armand, marquis de La Fayette and Marie-Madeleine de Marillac, and granddaughter of the author Marie-Madeleine Pioche de la Vergne, comtesse de la Fayette. They had one child, Charles Armand Rene de La Tr ´ emoille, born in 1708. Charles Louis Bretagne de La ´
+
+[p. 16 | section: Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? ´ Gold answer: [Charles Armand Rene de La Tr ´ emoille] ´ | type: Text]
+Charles Bretagne Marie de La Tremoille Charles Bretagne Marie de La Tr ´ emoille (24 March ´ 1764 – 10 November 1839), 8th duc de Thouars, was a French soldier and the son of Jean Bretagne Charles de La Tremoille and his wife, Marie-Maximilienne, princess of Salm-Kyrburg. La ´ Tremoille married Louise-Emmanuelle de Ch ´ atillon in 1781. She was a grand daughter of Louis ˆ Cesar de La Baume Le Blanc, the famous writer. The couple had one daughter: At the outbreak ´ of the French Revolution, he sided with King Louis XVI, as did the rest of his family. In 1789, La Tremoille and his parents ´
+
+[p. 16 | section: Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? ´ Gold answer: [Charles Armand Rene de La Tr ´ emoille] ´ | type: Text]
+Jean Bretagne Charles de La Tremoille Jean-Bretagne-Charles de La Tr ´ emoille (5 February 1737 ´ – 19 May 1792), 7th duc de Thouars, was a French soldier and the son of Charles Armand Rene´ de La Tremoille and his wife, Marie Hortense de La Tour d'Auvergne. La Tr ´ emoille began his ´ military career in the Seven Years' War as the head of the regiment of Aquitaine-Cavalry. He was seriously wounded and afterwards promoted to the rank of sergeant, and later of brigadier, of the armies of the king. During the French Revolution, La Tremoille remained faithful to King Louis ´ XVI. He emigrated in
+
+[p. 16 | section: Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? ´ Gold answer: [Charles Armand Rene de La Tr ´ emoille] ´ | type: Text]
+Charles Louis Bretagne de La Tremoille Charles Louis Bretagne de La Tr ´ emoille (15 March ´ 1683 – 9 October 1719), 6th duke of Thouars, was the son of Charles Belgique Hollande de La Tremoille and Madeleine de Cr ´ equy, daughter and heiress of Charles III de Cr ´ equy. He became ´ duke of Thouars upon his father's death in 1709. He was also count of Laval and Montfort and inherited the ancient French claim to the Kingdom of Naples. The Neapolitan claim gave the family the rank of "princes etrangers" at the French court. The Duke made a career in the ´ military as well
+
+[p. 16 | section: Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? ´ Gold answer: [Charles Armand Rene de La Tr ´ emoille] ´ | type: Text]
+de Saint-Didier. They had two daughters: Following her death, La Tremoille married for a third ´ time in 1830 to Valentine Eugenie Jos ´ ephine Walsh de Serrant. They had two children: La ´ Tremoille died shortly after his son's birth, in 1839. Charles Bretagne Marie de La Tr ´ emoille ´ Charles Bretagne Marie de La Tremoille (24 March 1764 – 10 November 1839), 8th duc de ´ Thouars, was a French soldier and the son of Jean Bretagne Charles de La Tremoille and his wife, ´ Marie-Maximilienne, princess of Salm-Kyrburg. La Tremoille married Louise-Emmanuelle de ´ Chatillon in 1781. She was a grand daughter of Louis C ˆ esar de ´
+
+[p. 16 | section: Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? ´ Gold answer: [Charles Armand Rene de La Tr ´ emoille] ´ | type: Table]
+Compressor Generated summary Predicted answer Off-the-shell Qwen2.5- 1.5B-Instruct Charles Bretagne Marie de La Tremoille's paternal grandfather ´ was Jean Bretagne Charles de La Tremoille. ´ Jean Bretagne Charles de La Tremoille ´ RECOMP Charles Bretagne Marie de La Tremoille's paternal grandfather ´ was Charles Louis Bretagne de La Tremoille, who was the son ´ of Charles Belgique Hollande de La Tremoille and Madeleine ´ de Crequy. This makes him the direct paternal grandfather of ´ Charles Bretagne Marie de La Tremoille. ´ Charles Louis Bretagne de La Tremoille ´ Ours ´ Charles Bretagne Marie de La Tremoille's paternal grandfather is Charles Armand Rene de La Tr ´ emoille, her father's father, the ´ 7th duc de Thouars. Charles Armand Rene´ de La Tremoille ´
+
+[p. 17 | section: Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? ´ Gold answer: [Charles Armand Rene de La Tr ´ emoille] ´ | type: TableGroup]
+Table 6: Failure Cases on NQ dataset. Question Generated Summary Predicted Answer Gold Answer who is the president of usa right now The current president of the USA is **Donald Trump**, who was elected on January 20, 2017, and is the 45th president. He is a businessman and reality TV personality, and his inauguration marked a shift in the country's political landscape. Joe Biden Donald Trump who has been ranked no. 1 in the latest football rankings announced by fifa The latest FIFA football rankings, announced in 2018, have ranked France as the top team, after winning the 2018 World Cup. This change was implemented using the Elo rating system, and the rankings were introduced to better reflect football team strengths. France Germany
+
+[p. 17 | section: Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? ´ Gold answer: [Charles Armand Rene de La Tr ´ emoille] ´ | type: Text]
+are concise, they largely fail to capture key information relevant to answering the question. In contrast, RECOMP demonstrates better summarization capability but is prone to being overwhelmed by lengthy documents, resulting in misjudgments and even generating misleading information—such as the statement in Table 4: " The U.S. stopped drafting for the Vietnam War after the Selective Service System was officially abolished in December 1972 "—which leads the downstream LLM to produce the incorrect answer "1972". Our method, CORE, accurately extracts answer-critical information from lengthy documents, exemplified by the summary: " The U.S. stopped drafting for the Vietnam War in 1973 after announcing the decision by Secretary of Defense Melvin Laird earlier that year ", thereby enabling the LLM to generate the correct answer "1973". This indicates that our compressor, trained with an end-task target-oriented reinforcement learning optimization strategy, can produce document summaries that are most helpful for answering the given question while effectively filtering out irrelevant information.
+
+[p. 17 | section: Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? ´ Gold answer: [Charles Armand Rene de La Tr ´ emoille] ´ | type: Text]
+To further understand the limitations of our approach, we present two failure cases from the NQ dataset where the model provided incorrect answers based on our generated summaries. As shown in Table 6, the first case reveals that although the summary contained the key information required for the correct answer, the downstream LLM still produced an error, potentially due to its over-reliance on parametric knowledge. In the second case, the summary itself omitted critical information needed to answer the question, which likely led to the incorrect response.
+
+[p. 17 | section: D IMPACT OF DIFFERENT COMPRESSORS ON PERFORMANCE | type: Text]
+In our previous experiments, we employed Qwen2.5-1.5B as the initial model to train our compressor. In this section, we utilize two additional models—Llama3.2-1B and Llama3.2-3B—as starting points to train our compressor and the baseline compressor, respectively. The experimental results are presented in Table 7 and Table 8. As shown in the results, our method CORE continues to achieve lossless compression with both models, maintaining a high token compression ratio while exhibiting no performance degradation in terms of Exact Match (EM) and F1 score compared to uncompressed RAG. Furthermore, under both new model configurations, our approach consistently outperforms the baseline methods, indicating that its superiority is not dependent on a specific model architecture and thus demonstrates strong robustness.
+
+[p. 17 | section: D IMPACT OF DIFFERENT COMPRESSORS ON PERFORMANCE | type: Text]
+We also observe that our method adheres to a form of scaling law: the compressor trained using the 3B model outperforms the one trained with the 1B model. Specifically, the 1B compressor improves performance by 1–4 EM points over the uncompressed baseline, while the 3B compressor yields gains of 3–9 EM points.
+
+[p. 18 | section: D IMPACT OF DIFFERENT COMPRESSORS ON PERFORMANCE | type: TableGroup]
+Table 7: Open-domain QA results using Qwen2.5-14B-Instruct as the downstream LLM (M). The reported token counts represent the length of in-context documents, excluding few-shot examples. RECOMP and our method CORE are both trained using llama3.2-1B-Instruct . NQ TriviaQA HotpotQA 2WikiMultihopQA EM F1 # tok EM F1 # tok EM F1 # tok EM F1 # tok No Retrieval 0.2136 0.3097 0 0.5323 0.5998 0 0.2105 0.2948 0 0.2611 0.2951 0 RAG without compre ession Top1 Document 0.3446 0.4441 142 0.6082 0.6870 143 0.2920 0.3893 147 0.2679 0.3187 153 Top3 Documents 0.3778 0.4845 427 0.6260 0.7102 430 0.3164 0.4187 442 0.2789 0.3358 460 Top5 Documents 0.3803 0.4916 712 0.6410 0.7248 715 0.3299 0.4369 737 0.2964 0.3521 766 Top10 Documents 0.3867 0.5003 1425 0.6440 0.7292 1428 0.3395 0.4488 1471 0.3104 0.3675 1531 Compression of top 3 5 docs llama3.2-1B 0.3147 0.4227 64 0.5552 0.6415 60 0.2648 0.3639 58 0.2498 0.3003 61 Deepseek-V3 (671B) 0.3773 0.5039 54 0.6528 0.7433 51 0.3359 0.4483 48 0.2507 0.3031 45 RECOMP (1B) 0.3410 0.4655 57 0.6071 0.6880 48 0.2987 0.4121 49 0.3045 0.3653 33 CORE (1B) 0.3947 0.4923 47 0.6483 0.7287 43 0.3344 0.4454 45 0.3378 0.3969 34 Compression of top 10 docs (w vith the co mpresso r trained on top 5 d locs) llama3.2-1B 0.3141 0.4228 62 0.5651 0.6512 5 8 0.2663 0.3661 56 0.2493 0.3006 61 Deepseek-V3 (671B) 0.3779 0.5107 56 0.6529 0.7445 53 0.3462 0.4569 50 0.2900 0.3464 40 RECOMP (1B) 0.3421 0.4661 59 0.6095 0.6917 52 0.2982 0.4105 55 0.3072 0.3681 44 CORE (1B) 0.4033 0.5033 47 0.6521 0.7296 45 0.3412 0.4500 48 0.3586 0.4162 42
+
+[p. 18 | section: D IMPACT OF DIFFERENT COMPRESSORS ON PERFORMANCE | type: TableGroup]
+Table 8: Open-domain QA results using Qwen2.5-14B-Instruct as the downstream LLM (M). The reported token counts represent the length of in-context documents, excluding few-shot examples. RECOMP and our method CORE are both trained using llama3.2-3B-Instruct . NQ TriviaQA HotpotQA 2WikiMultihopQA EM F1 # tok EM F1 # tok EM F1 # tok EM F1 # tok No Retrieval 0.2136 0.3097 0 0.5323 0.5998 0 0.2105 0.2948 0 0.2611 0.2951 0 RAG without compre ession Top1 Document 0.3446 0.4441 142 0.6082 0.6870 143 0.2920 0.3893 147 0.2679 0.3187 153 Top3 Documents 0.3778 0.4845 427 0.6260 0.7102 430 0.3164 0.4187 442 0.2789 0.3358 460 Top5 Documents 0.3803 0.4916 712 0.6410 0.7248 715 0.3299 0.4369 737 0.2964 0.3521 766 Top10 Documents 0.3867 0.5003 1425 0.6440 0.7292 1428 0.3395 0.4488 1471 0.3104 0.3675 1531 Compression of top : 5 docs llama3.2-3B 0.3252 0.4334 60 0.5650 0.6521 59 0.2772 0.3809 58 0.2485 0.2995 60 Deepseek-V3 (671B) 0.3773 0.5039 54 0.6528 0.7433 51 0.3359 0.4483 48 0.2507 0.3031 45 RECOMP (3B) 0.3657 0.4912 55 0.6183 0.6920 47 0.3025 0.4238 52 0.3274 0.3806 42 CORE (3B) 0.4310 0.5234 32 0.6650 0.7306 38 0.3507 0.4736 51 0.3905 0.4474 40 Compression of top 10 docs (w vith the co mpresso or trained on top 5 a locs) llama3.2-3B 0.3318 0.4359 61 0.5720 0.6588 57 0.2791 0.3854 60 0.2491 0.3011 59 Deepseek-V3 (671B) 0.3779 0.5107 56 0.6529 0.7445 53 0.3462 0.4569 50 0.2900 0.3464 40 RECOMP (3B) 0.3682 0.4963 52 0.6205 0.6973 44 0.3077 0.4261 54 0.3312 0.3869 50 CORE (3B) 0.4526 0.5467 33 0.6736 0.7404 37 0.3693 0.4926 51 0.4071 0.4633 48
+
+[p. 19 | section: D IMPACT OF DIFFERENT COMPRESSORS ON PERFORMANCE | type: TableGroup]
+Table 9: Zero-Shot Evaluation on HotpotQA of Models Trained on Natural Questions. EM F1 #tok No Retrieval 21.05 29.48 0 Full Documents 32.99 43.69 737 BM25 24.18 35.73 71 NoiseFilter-IB 27.97 38.62 38 RECOMP 28.96 39.95 56 CORE 33.67 45.06 36 RECOMP-Transfer 26.68 37.29 58 CORE-Transfer 31.25 42.84 35 Table 10: Evaluation on Noisy Natural Questions.
+
+[p. 19 | section: D IMPACT OF DIFFERENT COMPRESSORS ON PERFORMANCE | type: Table]
+EM F1 #tok full documents 35.21 45.38 1427 RECOMP 33.29 43.90 59 CORE 38.19 48.85 48
+
+[p. 19 | section: E CROSS-DATASET GENERALIZATION PERFORMANCE | type: Text]
+To verify the universal compression capability of our method, we directly transfer the model trained on the single-hop question answering dataset NQ to the multi-hop dataset HotpotQA for evaluation. The results, shown in Table 9, indicate that our transferred model achieves nearly lossless performance compared to using full documents without compression, while substantially outperforming the transfer results of the RECOMP baseline. Moreover, although both our method and the baseline underperform relative to models trained directly on the target HotpotQA dataset, our approach exhibits a smaller performance drop and demonstrates greater robustness compared to the baseline.
+
+[p. 19 | section: F ROBUSTNESS AGAINST NOISY CONTEXTS | type: Text]
+To evaluate the robustness of our approach against adversarial retrievals and noisy contexts, we constructed a noisy version of the NQ dataset. For each question, we constructed the input context by combining the top-3 passages retrieved by the DPR retriever with 7 randomly selected passages from the Wikipedia corpus to serve as irrelevant/noisy information. This resulted in a context of 10 passages, which were then shuffled to randomize the order. We then compared the performance of our method against the full-document baseline. Experimental results are presented in the table 10. In the "full documents" setting, the downstream LLM directly uses all these 10 passages to answer the question, whereas in our method, the compressor first summarizes the context, and the LLM then generates an answer based on the compressed content. The model we used was trained on the standard NQ dataset without any such noise augmentation. Our method not only matches but slightly surpasses the performance of using all documents, demonstrating its strong noise resistance and ability to extract key information from cluttered contexts. In addition, we compared our approach with the RECOMP baseline, and our method consistently outperforms it, reaffirming the superior compression capability and robustness of our model. Furthermore, our method achieves a high compression rate, condensing the source content from 1,427 tokens to just 48.

iclr26/6FyYLNnDL6/assets.json

@@ -0,0 +1,57 @@
+[
+  {
+    "filename": "_page_1_Figure_1.jpeg",
+    "path": "/network/scratch/j/jianan.zhao/ReviewAgent/data/processed_papers/iclr26_20260429_1952_all/marker_raw/6FyYLNnDL6/marker_markdown/6FyYLNnDL6/_page_1_Figure_1.jpeg",
+    "bytes": 32800,
+    "width": 456,
+    "height": 440,
+    "aspect_ratio": 1.0363636363636364,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_1_Figure_1.jpeg"
+  },
+  {
+    "filename": "_page_1_Figure_2.jpeg",
+    "path": "/network/scratch/j/jianan.zhao/ReviewAgent/data/processed_papers/iclr26_20260429_1952_all/marker_raw/6FyYLNnDL6/marker_markdown/6FyYLNnDL6/_page_1_Figure_2.jpeg",
+    "bytes": 33372,
+    "width": 449,
+    "height": 436,
+    "aspect_ratio": 1.0298165137614679,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_1_Figure_2.jpeg"
+  },
+  {
+    "filename": "_page_2_Figure_1.jpeg",
+    "path": "/network/scratch/j/jianan.zhao/ReviewAgent/data/processed_papers/iclr26_20260429_1952_all/marker_raw/6FyYLNnDL6/marker_markdown/6FyYLNnDL6/_page_2_Figure_1.jpeg",
+    "bytes": 112584,
+    "width": 1092,
+    "height": 635,
+    "aspect_ratio": 1.7196850393700787,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_2_Figure_1.jpeg"
+  },
+  {
+    "filename": "_page_7_Figure_1.jpeg",
+    "path": "/network/scratch/j/jianan.zhao/ReviewAgent/data/processed_papers/iclr26_20260429_1952_all/marker_raw/6FyYLNnDL6/marker_markdown/6FyYLNnDL6/_page_7_Figure_1.jpeg",
+    "bytes": 62904,
+    "width": 971,
+    "height": 357,
+    "aspect_ratio": 2.719887955182073,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_7_Figure_1.jpeg"
+  },
+  {
+    "filename": "_page_8_Figure_3.jpeg",
+    "path": "/network/scratch/j/jianan.zhao/ReviewAgent/data/processed_papers/iclr26_20260429_1952_all/marker_raw/6FyYLNnDL6/marker_markdown/6FyYLNnDL6/_page_8_Figure_3.jpeg",
+    "bytes": 16270,
+    "width": 394,
+    "height": 318,
+    "aspect_ratio": 1.2389937106918238,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_8_Figure_3.jpeg"
+  }
+]

iclr26/6FyYLNnDL6/chunks_v3_anonymized.jsonl

@@ -0,0 +1,106 @@
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0000", "section": "ABSTRACT", "page_start": 1, "page_end": 1, "type": "Text", "text": "Retrieval-Augmented Generation (RAG) has emerged as a promising approach to enhance the timeliness of knowledge updates and the factual accuracy of responses in large language models. However, incorporating a large number of retrieved documents significantly increases input length, leading to higher computational costs. Existing approaches to document compression tailored for RAG often degrade task performance, as they typically rely on predefined heuristics in the absence of clear compression guidelines. These heuristics fail to ensure that the compressed content effectively supports downstream tasks. To address these limitations, we propose CORE, a novel method for lossless context compression in RAG. CORE is optimized end-to-end and does not depend on predefined compression labels, which are often impractical to obtain. Instead, it leverages downstream task performance as a feedback signal, iteratively refining the compression policy to enhance task effectiveness. Extensive experiments across four datasets demonstrate the effectiveness of CORE. With a high compression ratio of 3%, CORE not only prevents performance degradation compared to including full documents (i.e., without compression) but also improves the average Exact Match (EM) score by 3.3 points. The code for CORE is available at .", "source": "marker_v2", "marker_block_id": "/page/0/Text/4"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0001", "section": "1 INTRODUCTION", "page_start": 1, "page_end": 1, "type": "Text", "text": "Large language models (LLMs) have undergone rapid development in recent years, significantly enhancing performance across various language tasks due to their emergent capabilities in semantic understanding and reasoning. Nevertheless, LLMs still face challenges in updating knowledge and providing factual responses (Fan et al., 2024) . To address these issues, Retrieval-Augmented Generation (RAG) has emerged as a promising approach. By retrieving the most relevant documents from external knowledge bases and prepending them as contextual information to the original input, RAG substantially improves LLM performance on knowledge-intensive tasks (Ram et al., 2023) .", "source": "marker_v2", "marker_block_id": "/page/0/Text/6"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0002", "section": "1 INTRODUCTION", "page_start": 1, "page_end": 1, "type": "Text", "text": "While RAG enhances performance, its effectiveness is closely tied to the number of retrieved documents used, since a broader context increases the probability of encompassing critical evidence. As illustrated in Figure 1, performance was weakest without any retrieved documents (i.e., without RAG). Accuracy improved consistently as more documents were added to the context, ultimately exceeding the no-RAG baseline by over 10 Exact Match (EM) points. However, this performance gain came with two significant limitations: (1) a substantial increase in computational cost from processing a larger number of context tokens (Xu et al., 2024) , and (2) the model's difficulty in effectively leveraging all provided documents, often resulting in the omission of key information located in the middle of the context (Liu et al., 2023) .", "source": "marker_v2", "marker_block_id": "/page/0/Text/7"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0003", "section": "1 INTRODUCTION", "page_start": 1, "page_end": 1, "type": "Text", "text": "These limitations have motivated recent research efforts aimed at compressing the retrieved context (Jin et al., 2024b; Wu et al., 2025; Jin et al., 2024a; Zhang et al., 2024a) . Prominent approaches include document summarization (Xu et al., 2024) , key information extraction (Cao et al., 2024; Xu et al., 2024) , the construction of key supporting evidence (Jin et al., 2024b) , and noise filtering based on information theory (Zhu et al., 2024) . Despite recent progress, these methods have several notable shortcomings. First, compression often results in a performance trade-off. For instance, RECOMP", "source": "marker_v2", "marker_block_id": "/page/0/Text/8"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0004", "section": "1 INTRODUCTION", "page_start": 2, "page_end": 2, "type": "FigureGroup", "text": "Figure 1: Performance evolution with an increasing number of retrieved documents on two datasets. Traditional RAG requires more documents for better performance, while our method achieves comparable or superior results with significant token compression.", "source": "marker_v2", "marker_block_id": "/page/1/FigureGroup/293"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0005", "section": "1 INTRODUCTION", "page_start": 2, "page_end": 2, "type": "Text", "text": "(Xu et al., 2024) suffers a 3–5 point drop in EM score (Table 1) , making it unsuitable for accuracysensitive applications. Second, most compression methods are heuristic in nature. The models are typically trained to generate summaries that are generally good but not necessarily useful for the downstream answer-generation LLM. This limitation arises from the lack of an ideal supervisory signal that defines what an optimal summary should be for the downstream task. This fundamental gap hinders the end-to-end optimization of existing methods. Finally, some compression models (Zhu et al., 2024) have parameter counts comparable to the LLM that performs the end task, resulting in substantial computational costs that undermine the efficiency gains of compression.", "source": "marker_v2", "marker_block_id": "/page/1/Text/4"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0006", "section": "1 INTRODUCTION", "page_start": 2, "page_end": 2, "type": "Text", "text": "Addressing these critical shortcomings requires a method that aligns compression with downstream task requirements, thereby minimizing performance trade-offs. To bridge this gap, we propose CORE, a novel method designed to achieve lossless context compression for RAG. Unlike previous compression methods, CORE is optimized in an end-to-end manner. Since obtaining predefined summary labels for supervision is impractical, we instead use downstream task performance as a feedback signal to evaluate the compression model's output. This feedback enables iterative refinement of the compression policy, guiding the compression model toward improved downstream performance. To this end, we employ Group Relative Policy Optimization (GRPO), a technique particularly well-suited for this purpose (Liu et al., 2024; Shao et al., 2024; Chen et al., 2025a) . In our framework, the accuracy of the downstream QA task is formalized as a reward, with the compression policy optimized through group-wise relative comparisons. Furthermore, our compression model is substantially smaller than the downstream LLM, which significantly reduces the computational overhead associated with encoding retrieved documents.", "source": "marker_v2", "marker_block_id": "/page/1/Text/5"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0007", "section": "1 INTRODUCTION", "page_start": 2, "page_end": 2, "type": "Text", "text": "We evaluate CORE on four benchmark datasets: two single-hop QA datasets ( Natural Questions and TriviaQA ) and two multi-hop datasets ( HotpotQA and 2WikiMultihopQA ). As shown in Table 1, CORE achieves state-of-the-art performance across all baselines. With a compression ratio of 3%, our approach not only avoids performance degradation compared to prepending full documents but also improves the average EM score by 3.3 points. We further demonstrate two key advantages of our approach. First, the effectiveness of CORE is not tied to a specific model architecture, as demonstrated by the fact that lossless compression can be achieved when various models are trained as compressors (Figure 3) . Second, the compressor exhibits strong transferability: a compression model trained using feedback from one LLM generalizes effectively to other LLMs (Table 2) . Finally, an in-depth case study (Tables 4 and 5) provides a qualitative analysis of the benefits of CORE.", "source": "marker_v2", "marker_block_id": "/page/1/Text/6"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0008", "section": "2 CORE-RAG", "page_start": 2, "page_end": 2, "type": "Text", "text": "This section introduces our proposed method, COmpression via REinforcement learning (CORE), which is shown in Figure 2. First, we provide an overview of the entire workflow. Then, we detail the end-to-end training strategy for the compression model, which is designed to drastically reduce the number of document tokens while preserving task performance.", "source": "marker_v2", "marker_block_id": "/page/1/Text/8"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0009", "section": "2 CORE-RAG", "page_start": 3, "page_end": 3, "type": "FigureGroup", "text": "Figure 2: Overview of our method CORE. The upper section illustrates the inference pipeline. The lower section depicts the end-to-end training method for the compression model.", "source": "marker_v2", "marker_block_id": "/page/2/FigureGroup/349"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0010", "section": "2.1 Problem Formulation", "page_start": 3, "page_end": 3, "type": "Text", "text": "We adopt the same problem formulation as prior work (Xu et al., 2024). Given an input question q, a target output y, and a set of k retrieved documents D, our objective is to compress D with respect to q into a summary s that preserves the most useful information while using significantly fewer tokens than D. This summary s is then prepended to the original input q and fed into an LLM to generate the final answer for the downstream task. This overall pipeline is illustrated in the upper half of Figure 2. The process involves two key components: a compression model \\pi_{\\theta} \\colon (q, D) \\mapsto s and a large language model M \\colon (s, q) \\mapsto \\hat{y} , which generates the predicted answer \\hat{y} . We treat M as a black-box system and focus exclusively on training the compressor \\pi_{\\theta} . The compressor itself is also a language model, but it is intentionally designed to be significantly smaller than M to reduce the computational cost of encoding the retrieved documents.", "source": "marker_v2", "marker_block_id": "/page/2/Text/4"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0011", "section": "2.2 Training the Compressor", "page_start": 3, "page_end": 3, "type": "Text", "text": "Our compressor is designed to generate document summaries that are highly useful to the LLM (M) for downstream tasks. This objective is challenging because the criteria for an effective summary are task-dependent, and direct supervision is unavailable. We therefore formulate this as an end-to-end training problem and employ reinforcement learning to optimize the compressor without relying on pre-defined compression labels. The overall architecture of our training framework is illustrated in the lower portion of Figure 2. The following sections describe the key components of our approach: distillation warm-up, policy optimization, and reward calculation.", "source": "marker_v2", "marker_block_id": "/page/2/Text/6"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0012", "section": "2.2.1 DISTILLATION FOR WARM-START", "page_start": 3, "page_end": 3, "type": "Text", "text": "Due to the limited parameter size of our compression model, its capability for question-focused document summarization is constrained. To provide a strong initial policy for RL and ensure training stability, we employ knowledge distillation from a teacher model to initialize our compressor. Specifically, we first utilize a large-scale language model (DeepSeek-V3) as the teacher to generate summaries of retrieved documents related to the given question. We then evaluate the performance of the downstream LLM (M) on the QA task under two conditions: (1) with the teacher-generated summary \\hat{s} prepended to the input question q, and (2) with the original question alone. The corresponding performance scores are denoted as p_{\\text{summary}} and p_{\\text{original}} , respectively. By comparing these results, we retain instances where p_{\\text{summary}} > p_{\\text{original}} , indicating that the summary enhances RAG performance. We also retain cases where p_{\\text{original}} = 1 (i.e., the model produces a fully correct answer without the summary) and p_{\\text{summary}} < p_{\\text{original}} ; for these, we set the target summary \\hat{s} to an empty string. All other instances are discarded. The resulting filtered and modified dataset is denoted as", "source": "marker_v2", "marker_block_id": "/page/2/Text/8"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0013", "section": "2.2.1 DISTILLATION FOR WARM-START", "page_start": 4, "page_end": 4, "type": "Text", "text": "\\mathcal{X}_f , which is used for supervised fine-tuning of the compression model. The fine-tuning objective is defined as:", "source": "marker_v2", "marker_block_id": "/page/3/Text/1"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0014", "section": "2.2.1 DISTILLATION FOR WARM-START", "page_start": 4, "page_end": 4, "type": "Text", "text": "\\mathcal{L}_{\\text{distill}} = \\frac{1}{|\\mathcal{X}_f|} \\sum_{(q,D,\\hat{s}) \\in \\mathcal{X}_f} \\mathcal{L}_{\\text{CE}}(\\pi_{\\theta}(q,D),\\hat{s}), \\tag{1}", "source": "marker_v2", "marker_block_id": "/page/3/Text/2"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0015", "section": "2.2.1 DISTILLATION FOR WARM-START", "page_start": 4, "page_end": 4, "type": "Text", "text": "where \\pi_{\\theta}(q, D) denotes the output of the compression model and \\mathcal{L}_{CE} is the cross-entropy loss. This distillation process yields a robust initialization for RL and promotes stability in subsequent training.", "source": "marker_v2", "marker_block_id": "/page/3/Text/3"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0016", "section": "2.2.2 END-TO-END TRAINING WITH RL", "page_start": 4, "page_end": 4, "type": "Text", "text": "Following the distillation phase, the compressor possesses a preliminary compression capability. However, as summaries from even the largest teacher models are not guaranteed to be optimal for the downstream task, further end-to-end optimization is necessary. We therefore formulate this optimization as a reinforcement learning problem. In this framework, the compressor functions as a policy that generates a summary from an input question and its corresponding documents. A reward function, which directly reflects performance on the downstream task (e.g., question-answering accuracy), then evaluates the summary's quality. The objective is to optimize the compressor's parameters to maximize the expected cumulative reward, thereby directly aligning its outputs with the downstream task's objectives.", "source": "marker_v2", "marker_block_id": "/page/3/Text/5"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0017", "section": "2.2.2 END-TO-END TRAINING WITH RL", "page_start": 4, "page_end": 4, "type": "Text", "text": "Specifically, we employ Group Relative Policy Optimization (GRPO) (Shao et al., 2024) algorithm. Unlike Proximal Policy Optimization (PPO), which trains a separate critic model, GRPO estimates the baseline directly from a group of rollouts. Given an existing policy, \\pi_{\\theta_{\\text{old}}} , and a reference policy, \\pi_{\\theta_{\\text{ref}}} , the GRPO objective maximizes the compressor policy \\pi_{\\theta} using G rollouts \\tau = \\{s_i\\}_{i=1}^G \\sim \\pi_{\\theta_{\\text{old}}}(\\cdot|x) , for each input x \\sim \\mathcal{D} :", "source": "marker_v2", "marker_block_id": "/page/3/Text/6"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0018", "section": "2.2.2 END-TO-END TRAINING WITH RL", "page_start": 4, "page_end": 4, "type": "Equation", "text": "\\mathcal{J}(\\theta) = \\mathbb{E}_{x \\sim \\mathcal{D}, \\ \\{s_i\\}_{i=1}^G \\sim \\pi_{\\theta_{\\text{old}}}(\\cdot \\mid x)} \\frac{1}{G} \\sum_{i=1}^G \\left[ \\min \\left( \\frac{\\pi_{\\theta}(s_i \\mid x)}{\\pi_{\\theta_{\\text{old}}}(s_i \\mid x)} A_i, \\ \\text{clip} \\left( \\frac{\\pi_{\\theta}(s_i \\mid x)}{\\pi_{\\theta_{\\text{old}}}(s_i \\mid x)}, 1 - \\epsilon, 1 + \\epsilon \\right) A_i \\right) - \\beta \\mathbb{D}_{\\text{KL}} \\left( \\pi_{\\theta} \\parallel \\pi_{\\theta_{\\text{ref}}} \\right) \\right], (2)", "source": "marker_v2", "marker_block_id": "/page/3/Equation/7"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0019", "section": "2.2.2 END-TO-END TRAINING WITH RL", "page_start": 4, "page_end": 4, "type": "Text", "text": "where A_i = (r_i - \\text{mean}(\\{r_j\\}_{j=1}^G))/\\text{std}(\\{r_j\\}_{j=1}^G) represents the normalized advantage of the i-th rollout within the group, \\epsilon is the clipping ratio, and \\beta is the coefficient for the KL divergence penalty. The inclusion of the KL divergence term ensures that the updated policy does not deviate significantly from the reference policy.", "source": "marker_v2", "marker_block_id": "/page/3/Text/8"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0020", "section": "2.2.3 REWARD CALCULATION", "page_start": 4, "page_end": 4, "type": "Text", "text": "Generating End-Task Output. It is important to note that the reward is not computed directly from the compressor's output summary. Instead, the summary s generated by the compressor is prepended to the original input question q, and this combined input is fed into the LLM M:(s,q)\\mapsto \\hat{y} to produce a predicted answer \\hat{y} . The reward is then calculated by comparing \\hat{y} to the gold answer y. Throughout the training process, the parameters of M remain fixed and are not updated.", "source": "marker_v2", "marker_block_id": "/page/3/Text/10"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0021", "section": "2.2.3 REWARD CALCULATION", "page_start": 4, "page_end": 4, "type": "Text", "text": "Computing Rewards. We design simple rule-based rewards based on end-task performance to guide the compressor's improvement, which consists of two components:", "source": "marker_v2", "marker_block_id": "/page/3/Text/11"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0022", "section": "2.2.3 REWARD CALCULATION", "page_start": 4, "page_end": 4, "type": "Text", "text": "• EM Reward ( r_{\\rm EM} ). We employ EM as the main reward function, which is a widely adopted metric for evaluating the accuracy of QA tasks. The EM reward is binary: it yields a value of 1 if the generated answer perfectly matches the ground truth, and 0 otherwise.", "source": "marker_v2", "marker_block_id": "/page/3/Text/12"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0023", "section": "2.2.3 REWARD CALCULATION", "page_start": 4, "page_end": 4, "type": "Equation", "text": "r_{\\rm EM} = \\begin{cases} 1 & \\text{if } y = \\hat{y}, \\\\ 0 & \\text{otherwise.} \\end{cases} (3)", "source": "marker_v2", "marker_block_id": "/page/3/Equation/13"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0024", "section": "2.2.3 REWARD CALCULATION", "page_start": 4, "page_end": 4, "type": "Text", "text": "• F1 Reward (r_{\\rm F1}) . Since exact matches occur infrequently in practice, relying solely on EM rewards leads to sparse reward signals. Furthermore, the EM metric fails to distinguish between partially correct answers, as all non-exact matches receive zero reward. Therefore, we introduce F1 reward, which provides a finer-grained evaluation by measuring the degree of partial match.", "source": "marker_v2", "marker_block_id": "/page/3/Text/14"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0025", "section": "2.2.3 REWARD CALCULATION", "page_start": 4, "page_end": 4, "type": "Equation", "text": "r_{\\rm FI} = \\frac{2 \\times I_N}{P_N + R_N},\\tag{4}", "source": "marker_v2", "marker_block_id": "/page/3/Equation/15"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0026", "section": "2.2.3 REWARD CALCULATION", "page_start": 4, "page_end": 4, "type": "Text", "text": "where P_N denotes the number of tokens in the predicted answer, R_N denotes the number of tokens in the gold answer, and I_N is the number of intersecting tokens between the two answers.", "source": "marker_v2", "marker_block_id": "/page/3/Text/16"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0027", "section": "2.2.3 REWARD CALCULATION", "page_start": 5, "page_end": 5, "type": "Text", "text": "The final reward function combines these reward signals through weighted summation:", "source": "marker_v2", "marker_block_id": "/page/4/Text/1"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0028", "section": "2.2.3 REWARD CALCULATION", "page_start": 5, "page_end": 5, "type": "Equation", "text": "r = r_{\\rm EM} + \\alpha \\cdot r_{\\rm Fl},\\tag{5}", "source": "marker_v2", "marker_block_id": "/page/4/Equation/2"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0029", "section": "2.2.3 REWARD CALCULATION", "page_start": 5, "page_end": 5, "type": "Text", "text": "where α ∈ (0, 1] is a hyperparameter that controls the relative contribution of the F1 reward.", "source": "marker_v2", "marker_block_id": "/page/4/Text/3"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0030", "section": "2.2.4 TRAINING TEMPLATE", "page_start": 5, "page_end": 5, "type": "Text", "text": "Figure 5 displays the prompt employed to train the compressor model for generating a summary of the retrieved documents, conditioned on the given question. This prompt is notably concise. For end-task answer generation, the prompt provided to the LLM M is illustrated in Figure 6, which incorporates few-shot in-context examples, the (generated summary of) retrieved documents, and the question.", "source": "marker_v2", "marker_block_id": "/page/4/Text/5"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0031", "section": "2.2.5 EFFICIENCY ANALYSIS", "page_start": 5, "page_end": 5, "type": "Text", "text": "Training Efficiency. Since our method employs reinforcement learning for training, it incurs greater time and computational costs compared to approaches that do not utilize reinforcement learning (Xu et al., 2024; Cao et al., 2024) . However, our training process only optimizes a lightweight compressor model with relatively few parameters, while the larger generator LLM responsible for producing task answers remains fixed and is not updated during training. This design ensures high training efficiency—for instance, training one epoch takes approximately 2 hours using eight H20 GPUs, and convergence is typically achieved within just two epochs. In contrast, other reinforcement learning-based methods, such as ReSearch (Chen et al., 2025a) and R1-Searcher (Song et al., 2025) , require direct fine-tuning of the large generator LLM, leading to considerably higher training time and resource consumption. Furthermore, it is important to emphasize that our method exhibits strong generalization capability. As shown in Section 3.3, a model trained only once demonstrates broad applicability, thereby reducing the need for frequent retraining and further lowering the overall training cost.", "source": "marker_v2", "marker_block_id": "/page/4/Text/7"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0032", "section": "2.2.5 EFFICIENCY ANALYSIS", "page_start": 5, "page_end": 5, "type": "Text", "text": "Inference Efficiency. Our method significantly enhances inference efficiency. In contrast to RAG approaches that do not employ a compressor—and thus require the generator LLM to directly encode lengthy documents, often spanning thousands of tokens—our approach introduces a lightweight compressor that processes long documents and summarizes them into compact representations of only a few dozen tokens before feeding them to the generator LLM. Since the compressor is an order of magnitude smaller in parameter size than the generator LLM, it substantially reduces the encoding time that would otherwise be incurred by the generator, leading to notable gains in inference efficiency. It is also important to note that the use of reinforcement learning does not adversely affect inference efficiency, as it is only involved during the training phase.", "source": "marker_v2", "marker_block_id": "/page/4/Text/8"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0033", "section": "3.1 EXPERIMENTAL SETTINGS", "page_start": 5, "page_end": 5, "type": "Text", "text": "Datasets and Evaluation Metrics. We evaluate our method on four benchmark datasets: two single-hop question-answering datasets, Natural Questions (NQ) (Kwiatkowski et al., 2019) and TriviaQA (Joshi et al., 2017) , as well as two multi-hop question-answering datasets, HotpotQA (Yang et al., 2018) and 2WikiMultihopQA (Ho et al., 2020) . Results are reported on the test sets of Natural Questions and TriviaQA, as well as the development sets of HotpotQA and 2WikiMultihopQA. Following RECOMP (Xu et al., 2024) , the performance is measured using Exact Match and token-level F1 scores, while efficiency is assessed by the number of tokens provided in the context.", "source": "marker_v2", "marker_block_id": "/page/4/Text/11"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0034", "section": "3.1 EXPERIMENTAL SETTINGS", "page_start": 5, "page_end": 5, "type": "Text", "text": "Compression Model (πθ). We trained our compression model using Qwen2.5-1.5B-Instruct to generate summaries of the retrieved documents. To evaluate the effect of using different models as compressors, we also trained compressors using Llama3.2-1B-Instruct and Llama3.2-3B-Instruct (Section 3.3) .", "source": "marker_v2", "marker_block_id": "/page/4/Text/12"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0035", "section": "3.1 EXPERIMENTAL SETTINGS", "page_start": 5, "page_end": 5, "type": "Text", "text": "Large Language Model (M). We use Qwen2.5-14B-Instruct as the primary LLM model M to generate predicted answers which are used to guide the training of the compressor. To evaluate the generalization ability of our method, we also transfer to another LLM model, LLama3.1-8B-Instruct (Table 2) .", "source": "marker_v2", "marker_block_id": "/page/4/Text/13"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0036", "section": "3.1 EXPERIMENTAL SETTINGS", "page_start": 6, "page_end": 6, "type": "TableGroup", "text": "Table 1: Open-domain QA results using Qwen2.5-14B-Instruct as the downstream LLM (M). The reported token counts represent the length of in-context documents, excluding few-shot examples. RECOMP, NoiseFilter-IB and our method CORE are all trained using Qwen2.5-1.5B-Instruct. NQ 7 TriviaQA L F IotpotQ. A 2Wiki Multiho pQA EM F1 # tok EM F1 # tok EM F1 # tok EM F1 # tok No Retrieval 21.36 30.97 0 53.23 59.98 0 21.05 29.48 0 26.11 29.51 0 RAG without compression on Top1 Document 34.46 44.41 142 60.82 68.70 143 29.20 38.93 147 26.79 31.87 153 Top3 Documents 37.78 48.45 427 62.60 71.02 430 31.64 41.87 442 27.89 33.58 460 Top5 Documents 38.03 49.16 712 64.10 72.48 715 32.99 43.69 737 29.64 35.21 766 Top10 Documents 38.67 50.03 1425 64.40 72.92 1428 33.95 44.88 1471 31.04 36.75 1531 Compression of top 5 do cuments ; BM25 25.23 36.47 37 55.36 63.90 39 24.18 35.73 71 25.42 30.29 68 Qwen2.5-1.5B 31.94 43.03 36 57.99 66.70 30 27.36 37.47 33 25.93 31.18 32 DeepSeek-V3 (671B) 37.73 50.39 54 64.13 73.20 50 33.59 44.83 48 27.99 32.67 92 RECOMP-Abs (1.5B) 34.18 46.26 58 60.31 68.50 53 28.96 39.95 56 30.25 36.73 52 RECOMP-Ext (1.5B) 33.84 46.05 56 60.18 68.39 48 29.93 41.09 45 30.78 37.07 51 NoiseFilter-IB (1.5B) 35.15 45.94 48 59.51 68.15 35 27.97 38.62 38 27.85 34.69 40 LongLLMLingua (1.5B) 33.65 43.15 152 58.96 66.82 148 28.03 38.49 149 29.37 33.62 153 QGC (1.5B) 36.23 45.88 49 61.02 68.45 47 29.16 40.05 45 31.14 36.83 51 CORE (1.5B) 41.02 50.40 46 65.63 72.55 32 33.67 45.06 36 36.72 42.05 49 Compression of top 10 a locumen ts (with t the comp ressor ti rained or ı top 5 d locs) BM25 25.91 36.88 38 55.28 63.16 37 23.49 35.01 68 25.61 30.54 65 Qwen2.5-1.5B 32.94 44.84 40 58.45 67.31 33 28.17 38.48 36 26.22 31.57 34 DeepSeek-V3 (671B) 37.79 51.07 56 65.29 74.45 53 34.62 45.69 50 29.00 34.64 40 RECOMP-Abs (1.5B) 34.40 46.93 59 61.42 69.88 52 31.54 42.92 52 31.98 38.16 49 RECOMP-Ext (1.5B) 33.96 46.34 60 61.03 69.51 50 31.92 43.18 55 32.52 38.87 44 NoiseFilter-IB (1.5B) 35.36 46.24 50 59.92 68.32 38 28.21 38.83 38 28.63 35.16 42 LongLLMLingua (1.5B) 33.78 43.37 154 59.17 66.97 150 28.33 38.95 148 29.62 34.11 151 QGC (1.5B) 36.03 45.62 50 61.23 68.74 49 29.12 39.63 46 31.71 37.52 50 CORE (1.5B) 41.88 51.26 52 66.76 73.64 33 34.68 46.35 37 37.99 43.28 48", "source": "marker_v2", "marker_block_id": "/page/5/TableGroup/211"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0037", "section": "3.1 EXPERIMENTAL SETTINGS", "page_start": 6, "page_end": 6, "type": "Text", "text": "Retrieval Corpus and Retrievers. Following previous studies (Xu et al., 2024), we use the Wikipedia corpus from December 20, 2018, as the retrieval source for all four datasets. The articles are segmented into non-overlapping 100-word documents. To ensure that our method is not dependent on a specific retriever, we experiment with several mainstream retrievers. Specifically, we use DPR (Karpukhin et al., 2020) for NQ, a hybrid of DPR and BM25 (Robertson et al., 1995) for TriviaQA, and the Contriever model (Izacard et al., 2021) trained on the MS MARCO dataset (Nguyen et al., 2016) for HotpotQA and 2WikiMultihopQA.", "source": "marker_v2", "marker_block_id": "/page/5/Text/3"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0038", "section": "3.1 EXPERIMENTAL SETTINGS", "page_start": 6, "page_end": 6, "type": "Text", "text": "Baselines. To evaluate the effectiveness of our method, we compared it against various baselines. First, we evaluated the uncompressed approach—retaining the original in-context RALM setup—by prepending the top 1, 3, 5, and 10 retrieved documents to the prompt. We also tested alternative compression methods, including the traditional BM25 algorithm (which ranks sentences by their similarity to the input), off-the-shelf Qwen2.5-1.5B-Instruct model (with comparable parameter size to our approach), and DeepSeek-V3 model (671B parameters, far exceeding our compressor's capacity). Furthermore, we included state-of-the-art context compression methods for RAG, RECOMP (Xu et al., 2024), NoiseFilter-IB (Zhu et al., 2024), LongLLMLingua (Jiang et al., 2024) and QGC (Cao et al., 2024). For RECOMP, we evaluated both its abstractive and extractive variants. To ensure a fair comparison, all trainable approaches were all trained using the same model.", "source": "marker_v2", "marker_block_id": "/page/5/Text/4"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0039", "section": "3.1 EXPERIMENTAL SETTINGS", "page_start": 6, "page_end": 6, "type": "Text", "text": "Implementation Details. For the distillation warm-up phase, we perform full-parameter supervised fine-tuning on the off-the-shell language model for two epochs using LLaMA-Factory ^1 . This warmed-up model then serves as the initializer for the subsequent reinforcement learning phase. We adopt the Verl framework ^2 for RL training. The initialized compression model is trained for two epochs on each dataset. Training is conducted on eight NVIDIA H20 GPUs using full parameter GRPO optimization, with a learning rate of 1e-5, a batch size of 256, five rollouts per sample, and a KL loss coefficient of 0.001. The downstream LLM (M) used for reward generation is served using the vLLM inference engine during RL training.", "source": "marker_v2", "marker_block_id": "/page/5/Text/5"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0040", "section": "3.1 EXPERIMENTAL SETTINGS", "page_start": 6, "page_end": 6, "type": "Footnote", "text": "& lt;sup>1</sup>", "source": "marker_v2", "marker_block_id": "/page/5/Footnote/6"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0041", "section": "3.1 EXPERIMENTAL SETTINGS", "page_start": 6, "page_end": 6, "type": "Footnote", "text": "& lt;sup>2</sup>", "source": "marker_v2", "marker_block_id": "/page/5/Footnote/7"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0042", "section": "3.2 OVERALL PERFORMANCE", "page_start": 7, "page_end": 7, "type": "Text", "text": "The detailed comparison results are presented in Table 1. We evaluate the following approaches: traditional RAG without compression using the top 1, 3, 5, and 10 documents prepended to the original input; BM25-based compression; off-the-shelf LLMs (Qwen2.5-1.5B-Instruct and DeepSeek-V3); and state-of-the-art trainable RAG compressors (RECOMP-Abstractive, RECOMP-Extractive, and NoiseFilter-IB). For fair comparison, all trainable methods were trained using the same backbone model, Qwen2.5-1.5B-Instruct. These compressors were trained on five-document inputs, and we report their performance both on in-domain five-document compression and out-of-domain generalization to ten-document compression.", "source": "marker_v2", "marker_block_id": "/page/6/Text/2"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0043", "section": "3.2 OVERALL PERFORMANCE", "page_start": 7, "page_end": 7, "type": "Text", "text": "RAG versus No RAG. As shown in Table 1, prepending the original input with retrieved documents yields a substantial improvement over the no-retrieval baseline, indicating that these documents provide valuable information for generating the answer. Performance improves as the number of retrieved documents increases from 1 to 10, albeit with diminishing returns—a trend attributable to the decreasing relevance of lower-ranked documents. However, this performance gain comes at the cost of a significant increase in the number of additional tokens the LLM must encode, rising from 0 to over 1,400 tokens.", "source": "marker_v2", "marker_block_id": "/page/6/Text/3"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0044", "section": "3.2 OVERALL PERFORMANCE", "page_start": 7, "page_end": 7, "type": "Text", "text": "Lossless Compression of CORE. Our compressor was trained using a context of five documents. As presented in Table 1, compared to prepending the full content of all five documents to the original input (i.e., the uncompressed baseline), our method achieves a high compression rate of approximately 6% with no loss in performance . Remarkably, on all four datasets, the approach not only maintains performance but also enhances EM by 1 to 7 points. For example, on NQ, the model achieves an EM of 41.02 with compressed input, surpassing the uncompressed score of 38.03.", "source": "marker_v2", "marker_block_id": "/page/6/Text/4"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0045", "section": "3.2 OVERALL PERFORMANCE", "page_start": 7, "page_end": 7, "type": "Text", "text": "Comparison with Compression Baselines. When compressing the top five documents, all baseline compression methods result in performance degradation to varying degrees compared to the uncompressed baseline. Specifically, BM25 leads to a substantial performance drop. Using the offthe-shelf Qwen2.5-1.5B-Instruct model for compression yields better results than BM25 but still underperforms relative to the uncompressed baseline. Surprisingly, even the large-scale DeepSeek-V3 model (with 671B parameters) achieves only near-lossless compression on TriviaQA and HotpotQA, while performance on NQ and 2Wiki remains below the uncompressed baseline. As for the trained compressors (RECOMP, NoiseFilter-IB, LongLLMLingua, and QGC), all of them exhibit performance degradation compared to no compression. The decline generally ranges from 2 to 6 EM poin ts across nearly all datasets. In contrast, our method, CORE, delivers the best performance. It not only surpasses compression methods of comparable size by 4-5 EM points but also maintains an advantage over the hundreds-of-times-larger DeepSeek model. This clearly demonstrates the benefit and importance of end-to-end optimization.", "source": "marker_v2", "marker_block_id": "/page/6/Text/5"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0046", "section": "3.2 OVERALL PERFORMANCE", "page_start": 7, "page_end": 7, "type": "Text", "text": "Generalization to 10-Doc Compression. When generalizing the trained compressors to handle the top-10 documents without retraining, the aforementioned conclusions remain valid. CORE continues to achieve lossless compression and performs best among all compression methods. On NQ, it achieves a token compression ratio of 3.6% while improving the EM by 3.2 points compared to using all ten documents. Similar trends are observed on TriviaQA, with a compression ratio of 2.3% and a gain of 2.4 EM points relative to the full-document baseline. On HotpotQA, CORE achieves a compression ratio of 2.5% and an improvement of 0.7 EM points. For 2WikiMultihopQA, it obtains a compression ratio of 3.1% along with a notable increase of 6.9 EM points.", "source": "marker_v2", "marker_block_id": "/page/6/Text/6"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0047", "section": "3.3 ROBUSTNESS AND GENERALIZATION ABILITY ANALYSIS", "page_start": 7, "page_end": 7, "type": "Text", "text": "Robustness of CORE Across Compressor Architectures. To evaluate whether the effectiveness of CORE depends on the choice of compressor model, we compared the performance of compressors trained using different model architectures—while keeping the downstream LLM (Qwen2.5-14B-Instruct) fixed. The compressors tested include LLaMA3.2-1B-Instruct, Qwen2.5-1.5B-Instruct, and LLaMA3.2-3B-Instruct, which vary in architecture and parameter count. As shown in Figure 3, the results indicate that: (1) These trained compressors consistently achieve lossless compression and outperform the uncompressed baseline (represented by the red reference line in the figure, which corresponds to prepending the full document content), confirming that our training framework is robust and not tied to a specific compressor architecture. (2) Compression performance improves as the size of the compressor model increases, consistent with scaling laws. More detailed results can be found in Tables 7 and 8.", "source": "marker_v2", "marker_block_id": "/page/6/Text/8"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0048", "section": "3.3 ROBUSTNESS AND GENERALIZATION ABILITY ANALYSIS", "page_start": 8, "page_end": 8, "type": "FigureGroup", "text": "Figure 3: The impact of different models used to train the compressor.", "source": "marker_v2", "marker_block_id": "/page/7/FigureGroup/208"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0049", "section": "3.3 ROBUSTNESS AND GENERALIZATION ABILITY ANALYSIS", "page_start": 8, "page_end": 8, "type": "TableGroup", "text": "Table 2: Zero-shot transfer of the trained compressor to Llama-3.1-8B-Instruct. NQ 7 TriviaQA HotpotQA 2WikiMultihopQA EM F1 # tok EM F1 # tok EM F1 # tok EM F1 # tok No Retrieval 24.04 34.91 0 55.64 62.57 0 19.93 27.75 0 27.64 31.18 0 RAG without compres ssion Top1 Document 33.80 44.06 142 59.17 67.50 143 27.95 37.49 147 28.41 33.43 153 Top3 Documents 36.87 47.81 427 61.13 70.06 430 30.17 40.71 442 28.67 34.23 460 Top5 Documents 37.65 48.87 712 62.26 71.04 715 31.44 42.16 737 29.43 35.18 766 Top10 Documents 38.12 49.93 1425 63.95 72.71 1428 32.19 42.62 1471 30.45 36.04 1531 Compression of top 5 docume nts Qwen2.5-1.5B 32.60 44.21 36 56.76 65.77 30 26.86 36.90 33 25.45 30.88 32 DeepSeek-V3 (671B) 37.56 50.11 54 62.52 72.34 50 33.05 44.25 48 28.64 33.87 92 RECOMP-Abs (1.5B) 33.41 45.50 58 58.50 67.37 53 28.85 39.76 56 31.63 37.81 52 RECOMP-Ext (1.5B) 33.12 45.06 60 57.98 66.84 55 29.03 40.04 52 31.85 38.02 55 CORE (1.5B) 40.72 50.00 46 64.08 71.13 32 32.17 43.71 36 35.99 41.42 49 Compression of top 1 0 docum ents Qwen2.5-1.5B 32.88 44.66 40 57.44 66.56 33 27.31 37.31 36 25.80 31.30 34 DeepSeek-V3 (671B) 37.49 51.28 56 63.79 73.80 53 34.24 45.35 50 31.45 37.09 40 RECOMP-Abs (1.5B) 34.18 46.80 59 59.69 68.89 52 30.17 41.42 55 33.61 39.78 44 RECOMP-Ext (1.5B) 34.06 46.55 60 59.33 68.71 50 30.52 41.98 55 33.52 39.42 44 CORE (1.5B) 41.77 51.27 52 65.25 72.45 33 33.25 45.09 37 37.59 42.87 48", "source": "marker_v2", "marker_block_id": "/page/7/TableGroup/209"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0050", "section": "3.3 ROBUSTNESS AND GENERALIZATION ABILITY ANALYSIS", "page_start": 8, "page_end": 8, "type": "Text", "text": "Transferability of CORE Across Downstream LLMs. We evaluate the transferability of our trained compressor and other baseline compressors to a new downstream LLM, the LLaMA-3.1-8B model. The results are presented in Table 2. Note that all trainable compressors, including our own, were trained using feedback generated by Qwen2.5-14B-Instruct. The findings reveal that existing trained compressors (e.g., RECOMP) exhibit limited generalization capability, as indicated by a larger performance gap relative to the baseline of prepending full documents. In contrast, CORE demonstrates stronger generalization, achieving lossless compression on the new downstream LLM. Specifically, it not only maintains performance but also surpasses the full-document baseline (i.e., no compression) across all four datasets while retaining a high compression rate. These results suggest that the summaries produced by our method are inherently high-quality and preserve key information necessary for accurate answering, thereby enabling effective transfer to other LLMs.", "source": "marker_v2", "marker_block_id": "/page/7/Text/5"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0051", "section": "3.4 ABLATION STUDY", "page_start": 8, "page_end": 8, "type": "Text", "text": "Table 3 presents an ablation study on the two stages of our method: distillation and GRPO. Here, \"w/o distillation\" denotes training the compressor with GRPO directly, bypassing the warm-start phase, while \"w/o GRPO\" indicates using only the distillation step without subsequent reinforcement learning. The results demonstrate that removing either component leads to performance degradation, confirming the necessity of both. In addition, the decline is more substantial when GRPO is omitted, highlighting the crucial role of reinforcement learning in the absence of explicit supervision. Distillation injects external knowledge into the model, providing a favorable starting point for RL training and thereby enabling RL to more effectively unlock the compressor's full potential.", "source": "marker_v2", "marker_block_id": "/page/7/Text/7"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0052", "section": "3.4 ABLATION STUDY", "page_start": 9, "page_end": 9, "type": "TableGroup", "text": "Table 3: Ablation study. Dataset Metric w/o distillation w/o GRPO CORE EM 36.37 34.18 41.02 NQ F1 46.91 46.26 50.40 EM 65.23 60.31 65.63 TQA F1 72.41 68.50 72.55 EM 32.01 28.96 33.67 HotpotQA F1 42.73 39.95 45.06 EM 31.40 30.25 36.72 2Wiki F1 36.89 36.73 42.05", "source": "marker_v2", "marker_block_id": "/page/8/TableGroup/495"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0053", "section": "3.4 ABLATION STUDY", "page_start": 9, "page_end": 9, "type": "FigureGroup", "text": "Figure 4: The impact of α.", "source": "marker_v2", "marker_block_id": "/page/8/FigureGroup/496"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0054", "section": "3.5 THE IMPACT OF α.", "page_start": 9, "page_end": 9, "type": "Text", "text": "Figure 4 illustrates the performance of our method under different values of α, which controls the weighting coefficient of the F1 reward term. Setting α = 0 corresponds to using only the EM reward. Results indicate that performance improves when α > 0 across all datasets, demonstrating the effectiveness of the F1 reward in mitigating the sparsity issue associated with the EM reward. However, the optimal value of α is dataset-dependent; values between 0.2 and 0.5 generally yield strong performance.", "source": "marker_v2", "marker_block_id": "/page/8/Text/6"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0055", "section": "3.6 CASE STUDY", "page_start": 9, "page_end": 9, "type": "Text", "text": "To conduct an in-depth analysis of the advantages of our compressor, we performed case studies on one single-hop QA dataset (NQ) and one multi-hop QA dataset (2Wiki), with the results presented in Table 4 and Table 5, respectively. For each case, we compared the summaries generated by off-the-shelf Qwen2.5-1.5B-Instruct, RECOMP, and our method CORE based on the same set of documents, as well as the predicted answers generated by the LLM after prepending these summaries. As shown in the tables, although the summaries produced by off-the-shelf Qwen2.5-1.5B are concise, they largely fail to capture key information relevant to answering the question. In contrast, RECOMP demonstrates better summarization capability but is prone to being overwhelmed by lengthy documents, resulting in misjudgments and even generating misleading information—such as the statement in Table 4: \" The U.S. stopped drafting for the Vietnam War after the Selective Service System was officially abolished in December 1972 \"—which leads the downstream LLM to produce the incorrect answer \"1972\". Our method, CORE, accurately extracts answer-critical information from lengthy documents, exemplified by the summary: \" The U.S. stopped drafting for the Vietnam War in 1973 after announcing the decision by Secretary of Defense Melvin Laird earlier that year \", thereby enabling the LLM to generate the correct answer \"1973\". This indicates that our compressor, trained with an end-task target-oriented reinforcement learning optimization strategy, can produce document summaries that are most helpful for answering the given question while effectively filtering out irrelevant information.", "source": "marker_v2", "marker_block_id": "/page/8/Text/8"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0056", "section": "4 RELATED WORK", "page_start": 9, "page_end": 9, "type": "Text", "text": "Context Compression in RAG. RAG enhances the performance of LLMs on knowledge-intensive tasks by retrieving the most relevant documents from extensive knowledge bases and prepending them as contextual information to the original input (Ram et al., 2023; Fan et al., 2024; Lin et al., 2023; Shi et al., 2023) . However, this approach requires the LLM to process significantly longer token sequences, resulting in increased computational costs. To mitigate this issue, researchers have begun to explore methods for compressing retrieved documents in RAG systems (Xu et al., 2024; Cao et al., 2024; Jin et al., 2024b; Zhu et al., 2024; Kim & Thorne, 2025; Rau et al., 2024; Wu et al., 2025; Louis et al., 2025; Jin et al., 2024a; Li et al., 2024a; b; Zhang et al., 2024a) . For instance, Xu et al. (2024) propose compressing retrieved documents into textual summaries before in-context augmentation, training the compressor through data selection and distillation. Similarly, Cao et al. (2024) introduce a Query-Guided Compressor (QGC) that uses queries to guide the compression process, effectively preserving essential information. Jin et al. (2024b) refine retrieved documents into Key Supporting Evidence (KSE) through a combination of knowledge synthesis, supervised fine-tuning (SFT), and preference alignment. Meanwhile, Zhu et al. (2024) present an informationtheoretic approach called NoiseFilter-IB, which filters noise by maximizing the mutual information", "source": "marker_v2", "marker_block_id": "/page/8/Text/10"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0057", "section": "4 RELATED WORK", "page_start": 10, "page_end": 10, "type": "Text", "text": "between the compressed content and the ground-truth output. Additionally, Kim & Thorne (2025) train a compressor to extract critical information using reward functions based on predefined heuristic rules. However, most of these methods are heuristic in nature, and due to the lack of ideal compression labels, the compressed content they produce cannot be guaranteed to benefit downstream LLMs. In contrast, our method, CORE, adopts an end-to-end optimization approach to address these limitations.", "source": "marker_v2", "marker_block_id": "/page/9/Text/1"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0058", "section": "4 RELATED WORK", "page_start": 10, "page_end": 10, "type": "Text", "text": "Reinforcement Learning. Reinforcement learning (RL) has recently achieved notable success, enabling LLMs to develop reasoning capabilities without explicit step-by-step supervision (Liu et al., 2024; Shao et al., 2024; Guo et al., 2025) . Building on these advances, several studies have applied RL to improve RAG (Ke et al., 2024) . For example, Kulkarni et al. (2024) use RL to autonomously decide whether to retrieve documents, while Zhang et al. (2024b) employ RL to optimize the ranking of retrieved documents. Similarly, Mao et al. (2024) propose a framework for training query rewriting models for RAG without relying on human annotations. MMOA-RAG (Chen et al., 2025b) enhances RAG through multi-agent reinforcement learning, incorporating a query rewriter, retriever, and generator. Meanwhile, RL has been applied to address other challenges in RAG, such as enhancing the quality of retrieved content. For instance, Oreo (Li & Ramakrishnan, 2025) trains a reconstructor with PPO and ROUGE-based rewards to rewrite passages for improved performance. In contrast, our work CORE tackles the distinct problem of computational efficiency. Our approach leverages the GRPO algorithm, direct task-performance rewards, and a lightweight compressor to enhance efficiency without compromising accuracy. Moreover, a line of research has utilized RL to integrate search with reasoning in a step-by-step manner (Singh et al., 2025) . For instance, Chen et al. (2025a) introduce a framework called ReSearch, which trains LLMs to reason with search using RL, without requiring supervised data for reasoning steps. Related approaches include R1- Searcher (Song et al., 2025) , WebThinker (Li et al., 2025) , and DeepResearcher (Zheng et al., 2025) . Although these methods are end-to-end, they differ fundamentally from our problem setting. These approaches typically involve directly training the LLM generator—which tends to be a large-scale model with a high parameter count. However, such training becomes infeasible when the model is a black box (e.g., GPT-4), as internal weights or gradients are inaccessible. Furthermore, these methods introduce extensive internal thinking processes that substantially increase context length and inference time. In contrast, our approach treats the generator LLM as a fixed black-box model and trains only a lightweight plug-in compressor to produce document summaries. This design significantly improves both training and inference efficiency. A parallel line of work applies RL to prompt compression. For example, PCRL (Jung & Kim, 2024) and TACO-RL (Shandilya et al., 2025) learn to compress prompts via token-level keep-or-drop decisions, using the similarity between model outputs with compressed and original prompts as the reward. Our work, CORE, introduces key distinctions in both objective and methodology. First, we target the more complex problem of compressing multiple retrieved documents in RAG, rather than single prompts. Second, we employ a generative compressor that can rephrase and synthesize content, instead of making token-level binary actions. Most importantly, we optimize compression using a direct task-performance reward with GRPO, which enables us to achieve true lossless compression at significantly higher ratios—a stark contrast to the performance degradation observed in prior prompt compression methods.", "source": "marker_v2", "marker_block_id": "/page/9/Text/2"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0059", "section": "5 CONCLUSION", "page_start": 10, "page_end": 10, "type": "Text", "text": "This paper analyzes the limitations of current context compression methods for RAG. A primary challenge is the lack of optimal reference summaries for supervised learning, which often results in performance degradation in downstream tasks. To overcome this, we frame context compression as a reinforcement learning problem, utilizing downstream task performance as a reward signal to train the compression policy, thereby enabling end-to-end optimization. Extensive experiments demonstrate that our proposed method, CORE, achieves effectively lossless compression by maintaining a high compression ratio while preserving original task performance. Surprisingly, CORE not only preserves but actually enhances performance on all downstream tasks. Further in-depth analysis provides additional insights into the efficacy of our approach.", "source": "marker_v2", "marker_block_id": "/page/9/Text/4"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0060", "section": "REFERENCES", "page_start": 11, "page_end": 11, "type": "ListGroup", "text": "Zhiwei Cao, Qian Cao, Yu Lu, Ningxin Peng, Luyang Huang, Shanbo Cheng, and Jinsong Su. Retaining key information under high compression ratios: Query-guided compressor for llms. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pp. 12685–12695, 2024. Mingyang Chen, Tianpeng Li, Haoze Sun, Yijie Zhou, Chenzheng Zhu, Haofen Wang, Jeff Z Pan, Wen Zhang, Huajun Chen, Fan Yang, et al. Learning to reason with search for llms via reinforcement learning. arXiv preprint arXiv:2503.19470 , 2025a. Yiqun Chen, Lingyong Yan, Weiwei Sun, Xinyu Ma, Yi Zhang, Shuaiqiang Wang, Dawei Yin, Yiming Yang, and Jiaxin Mao. Improving retrieval-augmented generation through multi-agent reinforcement learning. arXiv preprint arXiv:2501.15228 , 2025b. Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat-Seng Chua, and Qing Li. A survey on rag meeting llms: Towards retrieval-augmented large language models. In Proceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining , pp. 6491–6501, 2024. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948 , 2025. Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa. Constructing a multi-hop qa dataset for comprehensive evaluation of reasoning steps. arXiv preprint arXiv:2011.01060 , 2020. Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. Unsupervised dense information retrieval with contrastive learning. arXiv preprint arXiv:2112.09118 , 2021. Huiqiang Jiang, Qianhui Wu, Xufang Luo, Dongsheng Li, Chin-Yew Lin, Yuqing Yang, and Lili Qiu. Longllmlingua: Accelerating and enhancing llms in long context scenarios via prompt compression. In Proceedings of the 62nd Annual Meeting of the Association for Computational Lin guistics (Volume 1: Long Papers) , pp. 1658–1677, 2024. Bowen Jin, Jinsung Yoon, Jiawei Han, and Sercan O Arik. Long-context llms meet rag: Overcoming challenges for long inputs in rag. arXiv preprint arXiv:2410.05983 , 2024a. Jiajie Jin, Yutao Zhu, Yujia Zhou, and Zhicheng Dou. Bider: Bridging knowledge inconsistency for efficient retrieval-augmented llms via key supporting evidence. In Findings of the Association for Computational Linguistics ACL 2024 , pp. 750–761, 2024b. Mandar Joshi, Eunsol Choi, Daniel S Weld, and Luke Zettlemoyer. Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension. arXiv preprint arXiv:1705.03551 , 2017.", "source": "marker_v2", "marker_block_id": "/page/10/ListGroup/318"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0061", "section": "REFERENCES", "page_start": 12, "page_end": 12, "type": "ListGroup", "text": "Hoyoun Jung and Kyung-Joong Kim. Discrete prompt compression with reinforcement learning. IEEE Access , 12:72578–72587, 2024. Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick SH Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. Dense passage retrieval for open-domain question answering. In EMNLP (1) , pp. 6769–6781, 2020. Zixuan Ke, Weize Kong, Cheng Li, Mingyang Zhang, Qiaozhu Mei, and Michael Bendersky. Bridging the preference gap between retrievers and llms. arXiv preprint arXiv:2401.06954 , 2024. Sangryul Kim and James Thorne. Context filtering with reward modeling in question answering. In Proceedings of the 31st International Conference on Computational Linguistics , pp. 11048– 11055, 2025. Mandar Kulkarni, Praveen Tangarajan, Kyung Kim, and Anusua Trivedi. Reinforcement learning for optimizing rag for domain chatbots. arXiv preprint arXiv:2401.06800 , 2024. Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, et al. Natural questions: a benchmark for question answering research. Transactions of the Association for Computational Linguistics , 7:453–466, 2019. Sha Li and Naren Ramakrishnan. Oreo: A plug-in context reconstructor to enhance retrievalaugmented generation. In Proceedings of the 2025 International ACM SIGIR Conference on Innovative Concepts and Theories in Information Retrieval (ICTIR) , pp. 238–253, 2025. Xiaoxi Li, Jiajie Jin, Guanting Dong, Hongjin Qian, Yutao Zhu, Yongkang Wu, Ji-Rong Wen, and Zhicheng Dou. Webthinker: Empowering large reasoning models with deep research capability. arXiv preprint arXiv:2504.21776 , 2025. Xinze Li, Sen Mei, Zhenghao Liu, Yukun Yan, Shuo Wang, Shi Yu, Zheni Zeng, Hao Chen, Ge Yu, Zhiyuan Liu, et al. Rag-ddr: Optimizing retrieval-augmented generation using differentiable data rewards. arXiv preprint arXiv:2410.13509 , 2024a. Zhonghao Li, Xuming Hu, Aiwei Liu, Kening Zheng, Sirui Huang, and Hui Xiong. Refiner: Restructure retrieval content efficiently to advance question-answering capabilities. arXiv preprint arXiv:2406.11357 , 2024b. Xi Victoria Lin, Xilun Chen, Mingda Chen, Weijia Shi, Maria Lomeli, Richard James, Pedro Rodriguez, Jacob Kahn, Gergely Szilvasy, Mike Lewis, et al. Ra-dit: Retrieval-augmented dual instruction tuning. In The Twelfth International Conference on Learning Representations , 2023. Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437 , 2024. Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. Lost in the middle: How language models use long contexts. arXiv preprint arXiv:2307.03172 , 2023. Maxime Louis, Herve D´ ejean, and St ´ ephane Clinchant. Pisco: Pretty simple compression for ´ retrieval-augmented generation. arXiv preprint arXiv:2501.16075 , 2025. Shengyu Mao, Yong Jiang, Boli Chen, Xiao Li, Peng Wang, Xinyu Wang, Pengjun Xie, Fei Huang, Huajun Chen, and Ningyu Zhang. Rafe: Ranking feedback improves query rewriting for rag. arXiv preprint arXiv:2405.14431 , 2024. Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh Tiwary, Rangan Majumder, and Li Deng. Ms marco: A human-generated machine reading comprehension dataset. 2016. Ori Ram, Yoav Levine, Itay Dalmedigos, Dor Muhlgay, Amnon Shashua, Kevin Leyton-Brown, and Yoav Shoham. In-context retrieval-augmented language models. Transactions of the Association for Computational Linguistics , 11:1316–1331, 2023.", "source": "marker_v2", "marker_block_id": "/page/11/ListGroup/346"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0062", "section": "REFERENCES", "page_start": 13, "page_end": 13, "type": "ListGroup", "text": "David Rau, Shuai Wang, Herve D´ ejean, and St ´ ephane Clinchant. Context embeddings for efficient ´ answer generation in rag. arXiv preprint arXiv:2407.09252 , 2024. Stephen E Robertson, Steve Walker, Susan Jones, Micheline M Hancock-Beaulieu, Mike Gatford, et al. Okapi at TREC-3 . British Library Research and Development Department, 1995. Shivam Shandilya, Menglin Xia, Supriyo Ghosh, Huiqiang Jiang, Jue Zhang, Qianhui Wu, Victor Ruhle, and Saravan Rajmohan. Taco-rl: Task aware prompt compression optimization with rein- ¨ forcement learning. In Findings of the Association for Computational Linguistics: ACL 2025 , pp. 1582–1597, 2025. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300 , 2024. Weijia Shi, Sewon Min, Michihiro Yasunaga, Minjoon Seo, Rich James, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. Replug: Retrieval-augmented black-box language models. arXiv preprint arXiv:2301.12652 , 2023. Aditi Singh, Abul Ehtesham, Saket Kumar, and Tala Talaei Khoei. Agentic retrieval-augmented generation: A survey on agentic rag. arXiv preprint arXiv:2501.09136 , 2025. Huatong Song, Jinhao Jiang, Yingqian Min, Jie Chen, Zhipeng Chen, Wayne Xin Zhao, Lei Fang, and Ji-Rong Wen. R1-searcher: Incentivizing the search capability in llms via reinforcement learning. arXiv preprint arXiv:2503.05592 , 2025. Chenyuan Wu, Ninglu Shao, Zheng Liu, Shitao Xiao, Chaozhuo Li, Chen Zhang, Senzhang Wang, and Defu Lian. Lighter and better: Towards flexible context adaptation for retrieval augmented generation. In Proceedings of the Eighteenth ACM International Conference on Web Search and Data Mining , pp. 271–280, 2025. Fangyuan Xu, Weijia Shi, and Eunsol Choi. Recomp: Improving retrieval-augmented lms with context compression and selective augmentation. In The Twelfth International Conference on Learning Representations , 2024. Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W Cohen, Ruslan Salakhutdinov, and Christopher D Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question answering. arXiv preprint arXiv:1809.09600 , 2018. Qianchi Zhang, Hainan Zhang, Liang Pang, Hongwei Zheng, and Zhiming Zheng. Adacomp: Extractive context compression with adaptive predictor for retrieval-augmented large language models. arXiv preprint arXiv:2409.01579 , 2024a. Taolin Zhang, Dongyang Li, Qizhou Chen, Chengyu Wang, Longtao Huang, Hui Xue, Xiaofeng He, and Jun Huang. R 4: Reinforced retriever-reorder-responder for retrieval-augmented large language models. In ECAI 2024 , pp. 2314–2321. IOS Press, 2024b. Yuxiang Zheng, Dayuan Fu, Xiangkun Hu, Xiaojie Cai, Lyumanshan Ye, Pengrui Lu, and Pengfei Liu. Deepresearcher: Scaling deep research via reinforcement learning in real-world environments. arXiv preprint arXiv:2504.03160 , 2025. Kun Zhu, Xiaocheng Feng, Xiyuan Du, Yuxuan Gu, Weijiang Yu, Haotian Wang, Qianglong Chen, Zheng Chu, Jingchang Chen, and Bing Qin. An information bottleneck perspective for effective noise filtering on retrieval-augmented generation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pp. 1044–1069, 2024.", "source": "marker_v2", "marker_block_id": "/page/12/ListGroup/327"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0063", "section": "REFERENCES", "page_start": 14, "page_end": 14, "type": "Text", "text": "Compress the information in the retrieved documents into a 2-sentence summary that could be used to answer the question. If the documents do not contain relevant information, simply output \" \".", "source": "marker_v2", "marker_block_id": "/page/13/Text/1"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0064", "section": "REFERENCES", "page_start": 14, "page_end": 14, "type": "Text", "text": "Question: { Question }", "source": "marker_v2", "marker_block_id": "/page/13/Text/2"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0065", "section": "REFERENCES", "page_start": 14, "page_end": 14, "type": "Text", "text": "Retrieved documents: { Documents }", "source": "marker_v2", "marker_block_id": "/page/13/Text/3"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0066", "section": "REFERENCES", "page_start": 14, "page_end": 14, "type": "Text", "text": "Compressed documents:", "source": "marker_v2", "marker_block_id": "/page/13/Text/4"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0067", "section": "REFERENCES", "page_start": 14, "page_end": 14, "type": "Caption", "text": "Figure 5: Prompt template used to generate the summary.", "source": "marker_v2", "marker_block_id": "/page/13/Caption/5"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0068", "section": "[Instruction] Answer the question.", "page_start": 14, "page_end": 14, "type": "Text", "text": "IMPORTANT: Respond ONLY with the exact answer in the same format as the examples. Do NOT add any extra text, explanations, or punctuation. Do NOT include \"Answer:\" or any similar prefix in your response.", "source": "marker_v2", "marker_block_id": "/page/13/Text/7"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0069", "section": "[Examples]", "page_start": 14, "page_end": 14, "type": "Text", "text": "Question: Which major Russian city borders the body of water in which Saaremaa is located?", "source": "marker_v2", "marker_block_id": "/page/13/Text/9"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0070", "section": "[Examples]", "page_start": 14, "page_end": 14, "type": "Text", "text": "Answer: Saint Petersburg", "source": "marker_v2", "marker_block_id": "/page/13/Text/10"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0071", "section": "[Examples]", "page_start": 14, "page_end": 14, "type": "Text", "text": "Question: Who was thee first president of the association that wrote the code of ethics for psychology?", "source": "marker_v2", "marker_block_id": "/page/13/Text/11"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0072", "section": "[Examples]", "page_start": 14, "page_end": 14, "type": "Text", "text": "Answer: G. Stanley Hall", "source": "marker_v2", "marker_block_id": "/page/13/Text/12"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0073", "section": "[Examples]", "page_start": 14, "page_end": 14, "type": "Text", "text": "Question: Where did the Baldevins bryllup director die?", "source": "marker_v2", "marker_block_id": "/page/13/Text/13"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0074", "section": "[Examples]", "page_start": 14, "page_end": 14, "type": "Text", "text": "Answer: Copenhagen", "source": "marker_v2", "marker_block_id": "/page/13/Text/14"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0075", "section": "[Current Question]", "page_start": 14, "page_end": 14, "type": "Text", "text": "{ Summary of the retrieved documents }", "source": "marker_v2", "marker_block_id": "/page/13/Text/16"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0076", "section": "[Current Question]", "page_start": 14, "page_end": 14, "type": "Text", "text": "Question: { Question }", "source": "marker_v2", "marker_block_id": "/page/13/Text/17"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0077", "section": "[Current Question]", "page_start": 14, "page_end": 14, "type": "Text", "text": "Answer:", "source": "marker_v2", "marker_block_id": "/page/13/Text/18"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0078", "section": "[Current Question]", "page_start": 14, "page_end": 14, "type": "Caption", "text": "Figure 6: Prompt template for LLM QA.", "source": "marker_v2", "marker_block_id": "/page/13/Caption/19"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0079", "section": "A USE OF LLMS", "page_start": 14, "page_end": 14, "type": "Text", "text": "Large language models (LLMs) were employed solely as writing assistants to enhance the language, improve clarity, and check grammatical correctness. They were not used to generate research ideas, design or implement methodologies, conduct data analysis, or produce any of the results presented in this work. The authors assume full responsibility for the entire content of the paper.", "source": "marker_v2", "marker_block_id": "/page/13/Text/21"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0080", "section": "B PROMPT TEMPLATES", "page_start": 14, "page_end": 14, "type": "Text", "text": "Figure 5 displays the prompt employed to train the compressor model for generating a summary of the retrieved documents, conditioned on the given query. This prompt is notably concise. For end-task answer generation, the prompt provided to the LLM M is illustrated in Figure 6, which incorporates few-shot in-context examples, the (generated summary of) retrieved documents, and the question.", "source": "marker_v2", "marker_block_id": "/page/13/Text/23"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0081", "section": "C CASE STUDY", "page_start": 14, "page_end": 14, "type": "Text", "text": "To conduct an in-depth analysis of the advantages of our compressor, we performed case studies on one single-hop QA dataset (NQ) and one multi-hop QA dataset (2Wiki), with the results presented in Table 4 and Table 5, respectively. For each case, we compared the summaries generated by off-the-shelf Qwen2.5-1.5B-Instruct, RECOMP, and our method CORE based on the same set of documents, as well as the predicted answers generated by the LLM after prepending these summaries. As shown in the tables, although the summaries produced by off-the-shelf Qwen2.5-1.5B", "source": "marker_v2", "marker_block_id": "/page/13/Text/25"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0082", "section": "Top-5 documents", "page_start": 15, "page_end": 15, "type": "Text", "text": "last men conscripted, who were born in 1952 and who reported for duty in June 1973. On February 2, 1972, a drawing was held to determine draft priority numbers for men born in 1953, but in early 1973 it was announced by Secretary of Defense Melvin Laird that no further draft orders would be issued. In March 1973, 1974, and 1975, the Selective Service assigned draft priority numbers for all men born in 1954, 1955, and 1956, in case the draft was extended, but it never was. Command Sergeant Major Jeff Mellinger, believed to be the last drafted enlisted ranked.", "source": "marker_v2", "marker_block_id": "/page/14/Text/6"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0083", "section": "Top-5 documents", "page_start": 15, "page_end": 15, "type": "Text", "text": "The Gates Commission issued its report in February 1970, describing how adequate military strength could be maintained without having conscription. The existing draft law was expiring at the end of June 1971, but the Department of Defense and Nixon administration decided the draft needed to continue for at least some time. In February 1971, the administration requested of Congress a two-year extension of the draft, to June 1973. Senatorial opponents of the war wanted to reduce this to a one-year extension, or eliminate the draft altogether, or tie the draft renewal to a timetable for troop withdrawal from Vietnam; Senator.", "source": "marker_v2", "marker_block_id": "/page/14/Text/7"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0084", "section": "Top-5 documents", "page_start": 15, "page_end": 15, "type": "Text", "text": "draft operated from 1940 until 1946 when further inductions were suspended, and its legislative authorization expired without further extension by Congress in 1947. During this time, more than 10 million men had been inducted into military service. However, the Selective Service System remained intact. Scattered opposition was encountered especially in the northern cities where some African-Americans protested the system. The tiny Nation of Islam was at the forefront, with many Black Muslims jailed for refusing the draft, and their leader Elijah Muhammed was sentenced to federal prison for 5 years for inciting draft resistance. Organized draft resistance also developed in.", "source": "marker_v2", "marker_block_id": "/page/14/Text/8"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0085", "section": "Top-5 documents", "page_start": 15, "page_end": 15, "type": "Text", "text": "was the first time any form of student deferment was used. During the Korean War a student carrying at least twelve semester hours was spared until the end of his current semester. The United States breathed easier with the Korean War Armistice on July 27, 1953; however, technology brought new promises and threats. U.S. air and nuclear power fueled the Eisenhower doctrine of \"massive retaliation\". This strategy demanded more machines and fewer foot soldiers, so the draft slipped to the back burner. However, the head of the SSS, Maj. Gen. Hershey, urged caution fearing the conflict looming in Vietnam. In.", "source": "marker_v2", "marker_block_id": "/page/14/Text/9"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0086", "section": "Top-5 documents", "page_start": 15, "page_end": 15, "type": "Text", "text": "policy of using the draft as force to compel \"voluntary\" enlistment was unique in U.S. history. Previous drafts had not aimed at encouraging individuals to sign up in order to gain preferential placement or less dangerous postings. However, the incremental buildup of Vietnam without a clear threat to the country bolstered this. Some estimates suggest conscription encompassed almost one-third of all eligible men during the period of 1965–69. This group represented those without exemption or resources to avoid military service. During the active combat phase, the possibility of avoiding combat by selecting their service and military specialty led as many.", "source": "marker_v2", "marker_block_id": "/page/14/Text/10"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0087", "section": "Top-5 documents", "page_start": 15, "page_end": 15, "type": "Table", "text": "Compressor Generated summary Predicted answer Off-the-shell Qwen2.5- 1.5B-Instruct The US stopped drafting for the Vietnam War in late 1972 after the draft expired at the end of June 1971. late 1972 RECOMP The U.S. stopped drafting for the Vietnam War after the Selective Service System was officially abolished in December 1972 due to growing public opposition and the expiration of the draft's legal authority in 1971. While the draft was formally ended, some individuals were still consripted into the military December 1972 Ours The U.S. stopped drafting for the Vietnam War in 1973 after announcing the decision by Secretary of Defense Melvin Laird earlier that year. Although the Selective Service System was later updated to include draft priority numbers, this decision marked the end of the draft's use for national service. 1973", "source": "marker_v2", "marker_block_id": "/page/14/Table/11"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0088", "section": "Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? ´ Gold answer: [Charles Armand Rene de La Tr ´ emoille] ´", "page_start": 16, "page_end": 16, "type": "Text", "text": "Top-5 documents:", "source": "marker_v2", "marker_block_id": "/page/15/Text/5"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0089", "section": "Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? ´ Gold answer: [Charles Armand Rene de La Tr ´ emoille] ´", "page_start": 16, "page_end": 16, "type": "Text", "text": "as at Versailles: he was brigadier of cavalry (January 1709), first gentleman of the King's chamber (June 1709), governor of Thouars (July 1709), and Marechal de camp (February 1719). His sister ´ Marie Armande Victoire de La Tremoille married Emmanuel Th ´ eodose de La Tour d'Auvergne. ´ On 13 April 1706 he married Marie-Madeleine Motier de La Fayette (1691–1717), the daughter of Rene-Armand, marquis de La Fayette and Marie-Madeleine de Marillac, and granddaughter of the author Marie-Madeleine Pioche de la Vergne, comtesse de la Fayette. They had one child, Charles Armand Rene de La Tr ´ emoille, born in 1708. Charles Louis Bretagne de La ´", "source": "marker_v2", "marker_block_id": "/page/15/Text/6"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0090", "section": "Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? ´ Gold answer: [Charles Armand Rene de La Tr ´ emoille] ´", "page_start": 16, "page_end": 16, "type": "Text", "text": "Charles Bretagne Marie de La Tremoille Charles Bretagne Marie de La Tr ´ emoille (24 March ´ 1764 – 10 November 1839), 8th duc de Thouars, was a French soldier and the son of Jean Bretagne Charles de La Tremoille and his wife, Marie-Maximilienne, princess of Salm-Kyrburg. La ´ Tremoille married Louise-Emmanuelle de Ch ´ atillon in 1781. She was a grand daughter of Louis ˆ Cesar de La Baume Le Blanc, the famous writer. The couple had one daughter: At the outbreak ´ of the French Revolution, he sided with King Louis XVI, as did the rest of his family. In 1789, La Tremoille and his parents ´", "source": "marker_v2", "marker_block_id": "/page/15/Text/7"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0091", "section": "Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? ´ Gold answer: [Charles Armand Rene de La Tr ´ emoille] ´", "page_start": 16, "page_end": 16, "type": "Text", "text": "Jean Bretagne Charles de La Tremoille Jean-Bretagne-Charles de La Tr ´ emoille (5 February 1737 ´ – 19 May 1792), 7th duc de Thouars, was a French soldier and the son of Charles Armand Rene´ de La Tremoille and his wife, Marie Hortense de La Tour d'Auvergne. La Tr ´ emoille began his ´ military career in the Seven Years' War as the head of the regiment of Aquitaine-Cavalry. He was seriously wounded and afterwards promoted to the rank of sergeant, and later of brigadier, of the armies of the king. During the French Revolution, La Tremoille remained faithful to King Louis ´ XVI. He emigrated in", "source": "marker_v2", "marker_block_id": "/page/15/Text/8"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0092", "section": "Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? ´ Gold answer: [Charles Armand Rene de La Tr ´ emoille] ´", "page_start": 16, "page_end": 16, "type": "Text", "text": "Charles Louis Bretagne de La Tremoille Charles Louis Bretagne de La Tr ´ emoille (15 March ´ 1683 – 9 October 1719), 6th duke of Thouars, was the son of Charles Belgique Hollande de La Tremoille and Madeleine de Cr ´ equy, daughter and heiress of Charles III de Cr ´ equy. He became ´ duke of Thouars upon his father's death in 1709. He was also count of Laval and Montfort and inherited the ancient French claim to the Kingdom of Naples. The Neapolitan claim gave the family the rank of \"princes etrangers\" at the French court. The Duke made a career in the ´ military as well", "source": "marker_v2", "marker_block_id": "/page/15/Text/9"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0093", "section": "Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? ´ Gold answer: [Charles Armand Rene de La Tr ´ emoille] ´", "page_start": 16, "page_end": 16, "type": "Text", "text": "de Saint-Didier. They had two daughters: Following her death, La Tremoille married for a third ´ time in 1830 to Valentine Eugenie Jos ´ ephine Walsh de Serrant. They had two children: La ´ Tremoille died shortly after his son's birth, in 1839. Charles Bretagne Marie de La Tr ´ emoille ´ Charles Bretagne Marie de La Tremoille (24 March 1764 – 10 November 1839), 8th duc de ´ Thouars, was a French soldier and the son of Jean Bretagne Charles de La Tremoille and his wife, ´ Marie-Maximilienne, princess of Salm-Kyrburg. La Tremoille married Louise-Emmanuelle de ´ Chatillon in 1781. She was a grand daughter of Louis C ˆ esar de ´", "source": "marker_v2", "marker_block_id": "/page/15/Text/10"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0094", "section": "Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? ´ Gold answer: [Charles Armand Rene de La Tr ´ emoille] ´", "page_start": 16, "page_end": 16, "type": "Table", "text": "Compressor Generated summary Predicted answer Off-the-shell Qwen2.5- 1.5B-Instruct Charles Bretagne Marie de La Tremoille's paternal grandfather ´ was Jean Bretagne Charles de La Tremoille. ´ Jean Bretagne Charles de La Tremoille ´ RECOMP Charles Bretagne Marie de La Tremoille's paternal grandfather ´ was Charles Louis Bretagne de La Tremoille, who was the son ´ of Charles Belgique Hollande de La Tremoille and Madeleine ´ de Crequy. This makes him the direct paternal grandfather of ´ Charles Bretagne Marie de La Tremoille. ´ Charles Louis Bretagne de La Tremoille ´ Ours ´ Charles Bretagne Marie de La Tremoille's paternal grandfather is Charles Armand Rene de La Tr ´ emoille, her father's father, the ´ 7th duc de Thouars. Charles Armand Rene´ de La Tremoille ´", "source": "marker_v2", "marker_block_id": "/page/15/Table/11"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0095", "section": "Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? ´ Gold answer: [Charles Armand Rene de La Tr ´ emoille] ´", "page_start": 17, "page_end": 17, "type": "TableGroup", "text": "Table 6: Failure Cases on NQ dataset. Question Generated Summary Predicted Answer Gold Answer who is the president of usa right now The current president of the USA is **Donald Trump**, who was elected on January 20, 2017, and is the 45th president. He is a businessman and reality TV personality, and his inauguration marked a shift in the country's political landscape. Joe Biden Donald Trump who has been ranked no. 1 in the latest football rankings announced by fifa The latest FIFA football rankings, announced in 2018, have ranked France as the top team, after winning the 2018 World Cup. This change was implemented using the Elo rating system, and the rankings were introduced to better reflect football team strengths. France Germany", "source": "marker_v2", "marker_block_id": "/page/16/TableGroup/346"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0096", "section": "Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? ´ Gold answer: [Charles Armand Rene de La Tr ´ emoille] ´", "page_start": 17, "page_end": 17, "type": "Text", "text": "are concise, they largely fail to capture key information relevant to answering the question. In contrast, RECOMP demonstrates better summarization capability but is prone to being overwhelmed by lengthy documents, resulting in misjudgments and even generating misleading information—such as the statement in Table 4: \" The U.S. stopped drafting for the Vietnam War after the Selective Service System was officially abolished in December 1972 \"—which leads the downstream LLM to produce the incorrect answer \"1972\". Our method, CORE, accurately extracts answer-critical information from lengthy documents, exemplified by the summary: \" The U.S. stopped drafting for the Vietnam War in 1973 after announcing the decision by Secretary of Defense Melvin Laird earlier that year \", thereby enabling the LLM to generate the correct answer \"1973\". This indicates that our compressor, trained with an end-task target-oriented reinforcement learning optimization strategy, can produce document summaries that are most helpful for answering the given question while effectively filtering out irrelevant information.", "source": "marker_v2", "marker_block_id": "/page/16/Text/3"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0097", "section": "Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? ´ Gold answer: [Charles Armand Rene de La Tr ´ emoille] ´", "page_start": 17, "page_end": 17, "type": "Text", "text": "To further understand the limitations of our approach, we present two failure cases from the NQ dataset where the model provided incorrect answers based on our generated summaries. As shown in Table 6, the first case reveals that although the summary contained the key information required for the correct answer, the downstream LLM still produced an error, potentially due to its over-reliance on parametric knowledge. In the second case, the summary itself omitted critical information needed to answer the question, which likely led to the incorrect response.", "source": "marker_v2", "marker_block_id": "/page/16/Text/4"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0098", "section": "D IMPACT OF DIFFERENT COMPRESSORS ON PERFORMANCE", "page_start": 17, "page_end": 17, "type": "Text", "text": "In our previous experiments, we employed Qwen2.5-1.5B as the initial model to train our compressor. In this section, we utilize two additional models—Llama3.2-1B and Llama3.2-3B—as starting points to train our compressor and the baseline compressor, respectively. The experimental results are presented in Table 7 and Table 8. As shown in the results, our method CORE continues to achieve lossless compression with both models, maintaining a high token compression ratio while exhibiting no performance degradation in terms of Exact Match (EM) and F1 score compared to uncompressed RAG. Furthermore, under both new model configurations, our approach consistently outperforms the baseline methods, indicating that its superiority is not dependent on a specific model architecture and thus demonstrates strong robustness.", "source": "marker_v2", "marker_block_id": "/page/16/Text/6"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0099", "section": "D IMPACT OF DIFFERENT COMPRESSORS ON PERFORMANCE", "page_start": 17, "page_end": 17, "type": "Text", "text": "We also observe that our method adheres to a form of scaling law: the compressor trained using the 3B model outperforms the one trained with the 1B model. Specifically, the 1B compressor improves performance by 1–4 EM points over the uncompressed baseline, while the 3B compressor yields gains of 3–9 EM points.", "source": "marker_v2", "marker_block_id": "/page/16/Text/7"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0100", "section": "D IMPACT OF DIFFERENT COMPRESSORS ON PERFORMANCE", "page_start": 18, "page_end": 18, "type": "TableGroup", "text": "Table 7: Open-domain QA results using Qwen2.5-14B-Instruct as the downstream LLM (M). The reported token counts represent the length of in-context documents, excluding few-shot examples. RECOMP and our method CORE are both trained using llama3.2-1B-Instruct . NQ TriviaQA HotpotQA 2WikiMultihopQA EM F1 # tok EM F1 # tok EM F1 # tok EM F1 # tok No Retrieval 0.2136 0.3097 0 0.5323 0.5998 0 0.2105 0.2948 0 0.2611 0.2951 0 RAG without compre ession Top1 Document 0.3446 0.4441 142 0.6082 0.6870 143 0.2920 0.3893 147 0.2679 0.3187 153 Top3 Documents 0.3778 0.4845 427 0.6260 0.7102 430 0.3164 0.4187 442 0.2789 0.3358 460 Top5 Documents 0.3803 0.4916 712 0.6410 0.7248 715 0.3299 0.4369 737 0.2964 0.3521 766 Top10 Documents 0.3867 0.5003 1425 0.6440 0.7292 1428 0.3395 0.4488 1471 0.3104 0.3675 1531 Compression of top 3 5 docs llama3.2-1B 0.3147 0.4227 64 0.5552 0.6415 60 0.2648 0.3639 58 0.2498 0.3003 61 Deepseek-V3 (671B) 0.3773 0.5039 54 0.6528 0.7433 51 0.3359 0.4483 48 0.2507 0.3031 45 RECOMP (1B) 0.3410 0.4655 57 0.6071 0.6880 48 0.2987 0.4121 49 0.3045 0.3653 33 CORE (1B) 0.3947 0.4923 47 0.6483 0.7287 43 0.3344 0.4454 45 0.3378 0.3969 34 Compression of top 10 docs (w vith the co mpresso r trained on top 5 d locs) llama3.2-1B 0.3141 0.4228 62 0.5651 0.6512 5 8 0.2663 0.3661 56 0.2493 0.3006 61 Deepseek-V3 (671B) 0.3779 0.5107 56 0.6529 0.7445 53 0.3462 0.4569 50 0.2900 0.3464 40 RECOMP (1B) 0.3421 0.4661 59 0.6095 0.6917 52 0.2982 0.4105 55 0.3072 0.3681 44 CORE (1B) 0.4033 0.5033 47 0.6521 0.7296 45 0.3412 0.4500 48 0.3586 0.4162 42", "source": "marker_v2", "marker_block_id": "/page/17/TableGroup/149"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0101", "section": "D IMPACT OF DIFFERENT COMPRESSORS ON PERFORMANCE", "page_start": 18, "page_end": 18, "type": "TableGroup", "text": "Table 8: Open-domain QA results using Qwen2.5-14B-Instruct as the downstream LLM (M). The reported token counts represent the length of in-context documents, excluding few-shot examples. RECOMP and our method CORE are both trained using llama3.2-3B-Instruct . NQ TriviaQA HotpotQA 2WikiMultihopQA EM F1 # tok EM F1 # tok EM F1 # tok EM F1 # tok No Retrieval 0.2136 0.3097 0 0.5323 0.5998 0 0.2105 0.2948 0 0.2611 0.2951 0 RAG without compre ession Top1 Document 0.3446 0.4441 142 0.6082 0.6870 143 0.2920 0.3893 147 0.2679 0.3187 153 Top3 Documents 0.3778 0.4845 427 0.6260 0.7102 430 0.3164 0.4187 442 0.2789 0.3358 460 Top5 Documents 0.3803 0.4916 712 0.6410 0.7248 715 0.3299 0.4369 737 0.2964 0.3521 766 Top10 Documents 0.3867 0.5003 1425 0.6440 0.7292 1428 0.3395 0.4488 1471 0.3104 0.3675 1531 Compression of top : 5 docs llama3.2-3B 0.3252 0.4334 60 0.5650 0.6521 59 0.2772 0.3809 58 0.2485 0.2995 60 Deepseek-V3 (671B) 0.3773 0.5039 54 0.6528 0.7433 51 0.3359 0.4483 48 0.2507 0.3031 45 RECOMP (3B) 0.3657 0.4912 55 0.6183 0.6920 47 0.3025 0.4238 52 0.3274 0.3806 42 CORE (3B) 0.4310 0.5234 32 0.6650 0.7306 38 0.3507 0.4736 51 0.3905 0.4474 40 Compression of top 10 docs (w vith the co mpresso or trained on top 5 a locs) llama3.2-3B 0.3318 0.4359 61 0.5720 0.6588 57 0.2791 0.3854 60 0.2491 0.3011 59 Deepseek-V3 (671B) 0.3779 0.5107 56 0.6529 0.7445 53 0.3462 0.4569 50 0.2900 0.3464 40 RECOMP (3B) 0.3682 0.4963 52 0.6205 0.6973 44 0.3077 0.4261 54 0.3312 0.3869 50 CORE (3B) 0.4526 0.5467 33 0.6736 0.7404 37 0.3693 0.4926 51 0.4071 0.4633 48", "source": "marker_v2", "marker_block_id": "/page/17/TableGroup/150"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0102", "section": "D IMPACT OF DIFFERENT COMPRESSORS ON PERFORMANCE", "page_start": 19, "page_end": 19, "type": "TableGroup", "text": "Table 9: Zero-Shot Evaluation on HotpotQA of Models Trained on Natural Questions. EM F1 #tok No Retrieval 21.05 29.48 0 Full Documents 32.99 43.69 737 BM25 24.18 35.73 71 NoiseFilter-IB 27.97 38.62 38 RECOMP 28.96 39.95 56 CORE 33.67 45.06 36 RECOMP-Transfer 26.68 37.29 58 CORE-Transfer 31.25 42.84 35 Table 10: Evaluation on Noisy Natural Questions.", "source": "marker_v2", "marker_block_id": "/page/18/TableGroup/304"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0103", "section": "D IMPACT OF DIFFERENT COMPRESSORS ON PERFORMANCE", "page_start": 19, "page_end": 19, "type": "Table", "text": "EM F1 #tok full documents 35.21 45.38 1427 RECOMP 33.29 43.90 59 CORE 38.19 48.85 48", "source": "marker_v2", "marker_block_id": "/page/18/Table/4"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0104", "section": "E CROSS-DATASET GENERALIZATION PERFORMANCE", "page_start": 19, "page_end": 19, "type": "Text", "text": "To verify the universal compression capability of our method, we directly transfer the model trained on the single-hop question answering dataset NQ to the multi-hop dataset HotpotQA for evaluation. The results, shown in Table 9, indicate that our transferred model achieves nearly lossless performance compared to using full documents without compression, while substantially outperforming the transfer results of the RECOMP baseline. Moreover, although both our method and the baseline underperform relative to models trained directly on the target HotpotQA dataset, our approach exhibits a smaller performance drop and demonstrates greater robustness compared to the baseline.", "source": "marker_v2", "marker_block_id": "/page/18/Text/6"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0105", "section": "F ROBUSTNESS AGAINST NOISY CONTEXTS", "page_start": 19, "page_end": 19, "type": "Text", "text": "To evaluate the robustness of our approach against adversarial retrievals and noisy contexts, we constructed a noisy version of the NQ dataset. For each question, we constructed the input context by combining the top-3 passages retrieved by the DPR retriever with 7 randomly selected passages from the Wikipedia corpus to serve as irrelevant/noisy information. This resulted in a context of 10 passages, which were then shuffled to randomize the order. We then compared the performance of our method against the full-document baseline. Experimental results are presented in the table 10. In the \"full documents\" setting, the downstream LLM directly uses all these 10 passages to answer the question, whereas in our method, the compressor first summarizes the context, and the LLM then generates an answer based on the compressed content. The model we used was trained on the standard NQ dataset without any such noise augmentation. Our method not only matches but slightly surpasses the performance of using all documents, demonstrating its strong noise resistance and ability to extract key information from cluttered contexts. In addition, we compared our approach with the RECOMP baseline, and our method consistently outperforms it, reaffirming the superior compression capability and robustness of our model. Furthermore, our method achieves a high compression rate, condensing the source content from 1,427 tokens to just 48.", "source": "marker_v2", "marker_block_id": "/page/18/Text/8"}

iclr26/6FyYLNnDL6/dataset_meta.json

@@ -0,0 +1,63 @@
+{
+  "paper_id": "6FyYLNnDL6",
+  "pipeline": "Paper2Markdown-V3",
+  "ok": true,
+  "page_count": 19,
+  "chunk_count": 106,
+  "main_body_chunk_count": 60,
+  "appendix_chunk_count": 38,
+  "reference_chunk_count": 8,
+  "model_text_chars": 43162,
+  "raw_markdown_chars": 88191,
+  "sanitized_chars": 79686,
+  "page_provenance": {
+    "min_page": 1,
+    "max_page": 19,
+    "invalid_count": 0
+  },
+  "marker_block_type_counts": {
+    "Caption": 2,
+    "Equation": 4,
+    "Figure": 1,
+    "FigureGroup": 4,
+    "Footnote": 2,
+    "ListGroup": 3,
+    "PageFooter": 19,
+    "PageHeader": 19,
+    "SectionHeader": 37,
+    "Table": 3,
+    "TableGroup": 7,
+    "Text": 330
+  },
+  "asset_count_raw": 5,
+  "asset_count_model_kept": 5,
+  "asset_count_rejected": 0,
+  "asset_reject_reasons": {
+    "kept": 5
+  },
+  "artifact_leak_audit": {
+    "ok": true,
+    "hits": {
+      "Anonymous Authors": [],
+      "ACKNOWLEDGMENT": [],
+      "OpenReview": [],
+      "\"accept_label\"": [],
+      "\"decision\"": [],
+      "\"decision_tier\"": [],
+      "\"source_status\"": [],
+      "Meta-review": [],
+      "Official Review": [],
+      "official_reviews": [],
+      "meta_reviews": [],
+      "suggested_verdict_score": []
+    },
+    "artifact_count": 2
+  },
+  "default_model_input": "model_text_v3.txt",
+  "appendix_input": "appendix_text_v3.txt",
+  "reference_input": "reference_text_v3.txt",
+  "source": "iclr2026_train",
+  "run_name": "iclr26_20260429_1952_all",
+  "partial_snapshot": true,
+  "snapshot_created_at": "2026-04-30T02:34:47.606558+00:00"
+}

iclr26/6FyYLNnDL6/main_body_chunks.jsonl

@@ -0,0 +1,60 @@
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0000", "section": "ABSTRACT", "page_start": 1, "page_end": 1, "type": "Text", "text": "Retrieval-Augmented Generation (RAG) has emerged as a promising approach to enhance the timeliness of knowledge updates and the factual accuracy of responses in large language models. However, incorporating a large number of retrieved documents significantly increases input length, leading to higher computational costs. Existing approaches to document compression tailored for RAG often degrade task performance, as they typically rely on predefined heuristics in the absence of clear compression guidelines. These heuristics fail to ensure that the compressed content effectively supports downstream tasks. To address these limitations, we propose CORE, a novel method for lossless context compression in RAG. CORE is optimized end-to-end and does not depend on predefined compression labels, which are often impractical to obtain. Instead, it leverages downstream task performance as a feedback signal, iteratively refining the compression policy to enhance task effectiveness. Extensive experiments across four datasets demonstrate the effectiveness of CORE. With a high compression ratio of 3%, CORE not only prevents performance degradation compared to including full documents (i.e., without compression) but also improves the average Exact Match (EM) score by 3.3 points. The code for CORE is available at .", "source": "marker_v2", "marker_block_id": "/page/0/Text/4"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0001", "section": "1 INTRODUCTION", "page_start": 1, "page_end": 1, "type": "Text", "text": "Large language models (LLMs) have undergone rapid development in recent years, significantly enhancing performance across various language tasks due to their emergent capabilities in semantic understanding and reasoning. Nevertheless, LLMs still face challenges in updating knowledge and providing factual responses (Fan et al., 2024) . To address these issues, Retrieval-Augmented Generation (RAG) has emerged as a promising approach. By retrieving the most relevant documents from external knowledge bases and prepending them as contextual information to the original input, RAG substantially improves LLM performance on knowledge-intensive tasks (Ram et al., 2023) .", "source": "marker_v2", "marker_block_id": "/page/0/Text/6"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0002", "section": "1 INTRODUCTION", "page_start": 1, "page_end": 1, "type": "Text", "text": "While RAG enhances performance, its effectiveness is closely tied to the number of retrieved documents used, since a broader context increases the probability of encompassing critical evidence. As illustrated in Figure 1, performance was weakest without any retrieved documents (i.e., without RAG). Accuracy improved consistently as more documents were added to the context, ultimately exceeding the no-RAG baseline by over 10 Exact Match (EM) points. However, this performance gain came with two significant limitations: (1) a substantial increase in computational cost from processing a larger number of context tokens (Xu et al., 2024) , and (2) the model's difficulty in effectively leveraging all provided documents, often resulting in the omission of key information located in the middle of the context (Liu et al., 2023) .", "source": "marker_v2", "marker_block_id": "/page/0/Text/7"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0003", "section": "1 INTRODUCTION", "page_start": 1, "page_end": 1, "type": "Text", "text": "These limitations have motivated recent research efforts aimed at compressing the retrieved context (Jin et al., 2024b; Wu et al., 2025; Jin et al., 2024a; Zhang et al., 2024a) . Prominent approaches include document summarization (Xu et al., 2024) , key information extraction (Cao et al., 2024; Xu et al., 2024) , the construction of key supporting evidence (Jin et al., 2024b) , and noise filtering based on information theory (Zhu et al., 2024) . Despite recent progress, these methods have several notable shortcomings. First, compression often results in a performance trade-off. For instance, RECOMP", "source": "marker_v2", "marker_block_id": "/page/0/Text/8"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0004", "section": "1 INTRODUCTION", "page_start": 2, "page_end": 2, "type": "FigureGroup", "text": "Figure 1: Performance evolution with an increasing number of retrieved documents on two datasets. Traditional RAG requires more documents for better performance, while our method achieves comparable or superior results with significant token compression.", "source": "marker_v2", "marker_block_id": "/page/1/FigureGroup/293"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0005", "section": "1 INTRODUCTION", "page_start": 2, "page_end": 2, "type": "Text", "text": "(Xu et al., 2024) suffers a 3–5 point drop in EM score (Table 1) , making it unsuitable for accuracysensitive applications. Second, most compression methods are heuristic in nature. The models are typically trained to generate summaries that are generally good but not necessarily useful for the downstream answer-generation LLM. This limitation arises from the lack of an ideal supervisory signal that defines what an optimal summary should be for the downstream task. This fundamental gap hinders the end-to-end optimization of existing methods. Finally, some compression models (Zhu et al., 2024) have parameter counts comparable to the LLM that performs the end task, resulting in substantial computational costs that undermine the efficiency gains of compression.", "source": "marker_v2", "marker_block_id": "/page/1/Text/4"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0006", "section": "1 INTRODUCTION", "page_start": 2, "page_end": 2, "type": "Text", "text": "Addressing these critical shortcomings requires a method that aligns compression with downstream task requirements, thereby minimizing performance trade-offs. To bridge this gap, we propose CORE, a novel method designed to achieve lossless context compression for RAG. Unlike previous compression methods, CORE is optimized in an end-to-end manner. Since obtaining predefined summary labels for supervision is impractical, we instead use downstream task performance as a feedback signal to evaluate the compression model's output. This feedback enables iterative refinement of the compression policy, guiding the compression model toward improved downstream performance. To this end, we employ Group Relative Policy Optimization (GRPO), a technique particularly well-suited for this purpose (Liu et al., 2024; Shao et al., 2024; Chen et al., 2025a) . In our framework, the accuracy of the downstream QA task is formalized as a reward, with the compression policy optimized through group-wise relative comparisons. Furthermore, our compression model is substantially smaller than the downstream LLM, which significantly reduces the computational overhead associated with encoding retrieved documents.", "source": "marker_v2", "marker_block_id": "/page/1/Text/5"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0007", "section": "1 INTRODUCTION", "page_start": 2, "page_end": 2, "type": "Text", "text": "We evaluate CORE on four benchmark datasets: two single-hop QA datasets ( Natural Questions and TriviaQA ) and two multi-hop datasets ( HotpotQA and 2WikiMultihopQA ). As shown in Table 1, CORE achieves state-of-the-art performance across all baselines. With a compression ratio of 3%, our approach not only avoids performance degradation compared to prepending full documents but also improves the average EM score by 3.3 points. We further demonstrate two key advantages of our approach. First, the effectiveness of CORE is not tied to a specific model architecture, as demonstrated by the fact that lossless compression can be achieved when various models are trained as compressors (Figure 3) . Second, the compressor exhibits strong transferability: a compression model trained using feedback from one LLM generalizes effectively to other LLMs (Table 2) . Finally, an in-depth case study (Tables 4 and 5) provides a qualitative analysis of the benefits of CORE.", "source": "marker_v2", "marker_block_id": "/page/1/Text/6"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0008", "section": "2 CORE-RAG", "page_start": 2, "page_end": 2, "type": "Text", "text": "This section introduces our proposed method, COmpression via REinforcement learning (CORE), which is shown in Figure 2. First, we provide an overview of the entire workflow. Then, we detail the end-to-end training strategy for the compression model, which is designed to drastically reduce the number of document tokens while preserving task performance.", "source": "marker_v2", "marker_block_id": "/page/1/Text/8"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0009", "section": "2 CORE-RAG", "page_start": 3, "page_end": 3, "type": "FigureGroup", "text": "Figure 2: Overview of our method CORE. The upper section illustrates the inference pipeline. The lower section depicts the end-to-end training method for the compression model.", "source": "marker_v2", "marker_block_id": "/page/2/FigureGroup/349"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0010", "section": "2.1 Problem Formulation", "page_start": 3, "page_end": 3, "type": "Text", "text": "We adopt the same problem formulation as prior work (Xu et al., 2024). Given an input question q, a target output y, and a set of k retrieved documents D, our objective is to compress D with respect to q into a summary s that preserves the most useful information while using significantly fewer tokens than D. This summary s is then prepended to the original input q and fed into an LLM to generate the final answer for the downstream task. This overall pipeline is illustrated in the upper half of Figure 2. The process involves two key components: a compression model \\pi_{\\theta} \\colon (q, D) \\mapsto s and a large language model M \\colon (s, q) \\mapsto \\hat{y} , which generates the predicted answer \\hat{y} . We treat M as a black-box system and focus exclusively on training the compressor \\pi_{\\theta} . The compressor itself is also a language model, but it is intentionally designed to be significantly smaller than M to reduce the computational cost of encoding the retrieved documents.", "source": "marker_v2", "marker_block_id": "/page/2/Text/4"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0011", "section": "2.2 Training the Compressor", "page_start": 3, "page_end": 3, "type": "Text", "text": "Our compressor is designed to generate document summaries that are highly useful to the LLM (M) for downstream tasks. This objective is challenging because the criteria for an effective summary are task-dependent, and direct supervision is unavailable. We therefore formulate this as an end-to-end training problem and employ reinforcement learning to optimize the compressor without relying on pre-defined compression labels. The overall architecture of our training framework is illustrated in the lower portion of Figure 2. The following sections describe the key components of our approach: distillation warm-up, policy optimization, and reward calculation.", "source": "marker_v2", "marker_block_id": "/page/2/Text/6"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0012", "section": "2.2.1 DISTILLATION FOR WARM-START", "page_start": 3, "page_end": 3, "type": "Text", "text": "Due to the limited parameter size of our compression model, its capability for question-focused document summarization is constrained. To provide a strong initial policy for RL and ensure training stability, we employ knowledge distillation from a teacher model to initialize our compressor. Specifically, we first utilize a large-scale language model (DeepSeek-V3) as the teacher to generate summaries of retrieved documents related to the given question. We then evaluate the performance of the downstream LLM (M) on the QA task under two conditions: (1) with the teacher-generated summary \\hat{s} prepended to the input question q, and (2) with the original question alone. The corresponding performance scores are denoted as p_{\\text{summary}} and p_{\\text{original}} , respectively. By comparing these results, we retain instances where p_{\\text{summary}} > p_{\\text{original}} , indicating that the summary enhances RAG performance. We also retain cases where p_{\\text{original}} = 1 (i.e., the model produces a fully correct answer without the summary) and p_{\\text{summary}} < p_{\\text{original}} ; for these, we set the target summary \\hat{s} to an empty string. All other instances are discarded. The resulting filtered and modified dataset is denoted as", "source": "marker_v2", "marker_block_id": "/page/2/Text/8"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0013", "section": "2.2.1 DISTILLATION FOR WARM-START", "page_start": 4, "page_end": 4, "type": "Text", "text": "\\mathcal{X}_f , which is used for supervised fine-tuning of the compression model. The fine-tuning objective is defined as:", "source": "marker_v2", "marker_block_id": "/page/3/Text/1"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0014", "section": "2.2.1 DISTILLATION FOR WARM-START", "page_start": 4, "page_end": 4, "type": "Text", "text": "\\mathcal{L}_{\\text{distill}} = \\frac{1}{|\\mathcal{X}_f|} \\sum_{(q,D,\\hat{s}) \\in \\mathcal{X}_f} \\mathcal{L}_{\\text{CE}}(\\pi_{\\theta}(q,D),\\hat{s}), \\tag{1}", "source": "marker_v2", "marker_block_id": "/page/3/Text/2"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0015", "section": "2.2.1 DISTILLATION FOR WARM-START", "page_start": 4, "page_end": 4, "type": "Text", "text": "where \\pi_{\\theta}(q, D) denotes the output of the compression model and \\mathcal{L}_{CE} is the cross-entropy loss. This distillation process yields a robust initialization for RL and promotes stability in subsequent training.", "source": "marker_v2", "marker_block_id": "/page/3/Text/3"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0016", "section": "2.2.2 END-TO-END TRAINING WITH RL", "page_start": 4, "page_end": 4, "type": "Text", "text": "Following the distillation phase, the compressor possesses a preliminary compression capability. However, as summaries from even the largest teacher models are not guaranteed to be optimal for the downstream task, further end-to-end optimization is necessary. We therefore formulate this optimization as a reinforcement learning problem. In this framework, the compressor functions as a policy that generates a summary from an input question and its corresponding documents. A reward function, which directly reflects performance on the downstream task (e.g., question-answering accuracy), then evaluates the summary's quality. The objective is to optimize the compressor's parameters to maximize the expected cumulative reward, thereby directly aligning its outputs with the downstream task's objectives.", "source": "marker_v2", "marker_block_id": "/page/3/Text/5"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0017", "section": "2.2.2 END-TO-END TRAINING WITH RL", "page_start": 4, "page_end": 4, "type": "Text", "text": "Specifically, we employ Group Relative Policy Optimization (GRPO) (Shao et al., 2024) algorithm. Unlike Proximal Policy Optimization (PPO), which trains a separate critic model, GRPO estimates the baseline directly from a group of rollouts. Given an existing policy, \\pi_{\\theta_{\\text{old}}} , and a reference policy, \\pi_{\\theta_{\\text{ref}}} , the GRPO objective maximizes the compressor policy \\pi_{\\theta} using G rollouts \\tau = \\{s_i\\}_{i=1}^G \\sim \\pi_{\\theta_{\\text{old}}}(\\cdot|x) , for each input x \\sim \\mathcal{D} :", "source": "marker_v2", "marker_block_id": "/page/3/Text/6"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0018", "section": "2.2.2 END-TO-END TRAINING WITH RL", "page_start": 4, "page_end": 4, "type": "Equation", "text": "\\mathcal{J}(\\theta) = \\mathbb{E}_{x \\sim \\mathcal{D}, \\ \\{s_i\\}_{i=1}^G \\sim \\pi_{\\theta_{\\text{old}}}(\\cdot \\mid x)} \\frac{1}{G} \\sum_{i=1}^G \\left[ \\min \\left( \\frac{\\pi_{\\theta}(s_i \\mid x)}{\\pi_{\\theta_{\\text{old}}}(s_i \\mid x)} A_i, \\ \\text{clip} \\left( \\frac{\\pi_{\\theta}(s_i \\mid x)}{\\pi_{\\theta_{\\text{old}}}(s_i \\mid x)}, 1 - \\epsilon, 1 + \\epsilon \\right) A_i \\right) - \\beta \\mathbb{D}_{\\text{KL}} \\left( \\pi_{\\theta} \\parallel \\pi_{\\theta_{\\text{ref}}} \\right) \\right], (2)", "source": "marker_v2", "marker_block_id": "/page/3/Equation/7"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0019", "section": "2.2.2 END-TO-END TRAINING WITH RL", "page_start": 4, "page_end": 4, "type": "Text", "text": "where A_i = (r_i - \\text{mean}(\\{r_j\\}_{j=1}^G))/\\text{std}(\\{r_j\\}_{j=1}^G) represents the normalized advantage of the i-th rollout within the group, \\epsilon is the clipping ratio, and \\beta is the coefficient for the KL divergence penalty. The inclusion of the KL divergence term ensures that the updated policy does not deviate significantly from the reference policy.", "source": "marker_v2", "marker_block_id": "/page/3/Text/8"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0020", "section": "2.2.3 REWARD CALCULATION", "page_start": 4, "page_end": 4, "type": "Text", "text": "Generating End-Task Output. It is important to note that the reward is not computed directly from the compressor's output summary. Instead, the summary s generated by the compressor is prepended to the original input question q, and this combined input is fed into the LLM M:(s,q)\\mapsto \\hat{y} to produce a predicted answer \\hat{y} . The reward is then calculated by comparing \\hat{y} to the gold answer y. Throughout the training process, the parameters of M remain fixed and are not updated.", "source": "marker_v2", "marker_block_id": "/page/3/Text/10"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0021", "section": "2.2.3 REWARD CALCULATION", "page_start": 4, "page_end": 4, "type": "Text", "text": "Computing Rewards. We design simple rule-based rewards based on end-task performance to guide the compressor's improvement, which consists of two components:", "source": "marker_v2", "marker_block_id": "/page/3/Text/11"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0022", "section": "2.2.3 REWARD CALCULATION", "page_start": 4, "page_end": 4, "type": "Text", "text": "• EM Reward ( r_{\\rm EM} ). We employ EM as the main reward function, which is a widely adopted metric for evaluating the accuracy of QA tasks. The EM reward is binary: it yields a value of 1 if the generated answer perfectly matches the ground truth, and 0 otherwise.", "source": "marker_v2", "marker_block_id": "/page/3/Text/12"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0023", "section": "2.2.3 REWARD CALCULATION", "page_start": 4, "page_end": 4, "type": "Equation", "text": "r_{\\rm EM} = \\begin{cases} 1 & \\text{if } y = \\hat{y}, \\\\ 0 & \\text{otherwise.} \\end{cases} (3)", "source": "marker_v2", "marker_block_id": "/page/3/Equation/13"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0024", "section": "2.2.3 REWARD CALCULATION", "page_start": 4, "page_end": 4, "type": "Text", "text": "• F1 Reward (r_{\\rm F1}) . Since exact matches occur infrequently in practice, relying solely on EM rewards leads to sparse reward signals. Furthermore, the EM metric fails to distinguish between partially correct answers, as all non-exact matches receive zero reward. Therefore, we introduce F1 reward, which provides a finer-grained evaluation by measuring the degree of partial match.", "source": "marker_v2", "marker_block_id": "/page/3/Text/14"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0025", "section": "2.2.3 REWARD CALCULATION", "page_start": 4, "page_end": 4, "type": "Equation", "text": "r_{\\rm FI} = \\frac{2 \\times I_N}{P_N + R_N},\\tag{4}", "source": "marker_v2", "marker_block_id": "/page/3/Equation/15"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0026", "section": "2.2.3 REWARD CALCULATION", "page_start": 4, "page_end": 4, "type": "Text", "text": "where P_N denotes the number of tokens in the predicted answer, R_N denotes the number of tokens in the gold answer, and I_N is the number of intersecting tokens between the two answers.", "source": "marker_v2", "marker_block_id": "/page/3/Text/16"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0027", "section": "2.2.3 REWARD CALCULATION", "page_start": 5, "page_end": 5, "type": "Text", "text": "The final reward function combines these reward signals through weighted summation:", "source": "marker_v2", "marker_block_id": "/page/4/Text/1"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0028", "section": "2.2.3 REWARD CALCULATION", "page_start": 5, "page_end": 5, "type": "Equation", "text": "r = r_{\\rm EM} + \\alpha \\cdot r_{\\rm Fl},\\tag{5}", "source": "marker_v2", "marker_block_id": "/page/4/Equation/2"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0029", "section": "2.2.3 REWARD CALCULATION", "page_start": 5, "page_end": 5, "type": "Text", "text": "where α ∈ (0, 1] is a hyperparameter that controls the relative contribution of the F1 reward.", "source": "marker_v2", "marker_block_id": "/page/4/Text/3"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0030", "section": "2.2.4 TRAINING TEMPLATE", "page_start": 5, "page_end": 5, "type": "Text", "text": "Figure 5 displays the prompt employed to train the compressor model for generating a summary of the retrieved documents, conditioned on the given question. This prompt is notably concise. For end-task answer generation, the prompt provided to the LLM M is illustrated in Figure 6, which incorporates few-shot in-context examples, the (generated summary of) retrieved documents, and the question.", "source": "marker_v2", "marker_block_id": "/page/4/Text/5"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0031", "section": "2.2.5 EFFICIENCY ANALYSIS", "page_start": 5, "page_end": 5, "type": "Text", "text": "Training Efficiency. Since our method employs reinforcement learning for training, it incurs greater time and computational costs compared to approaches that do not utilize reinforcement learning (Xu et al., 2024; Cao et al., 2024) . However, our training process only optimizes a lightweight compressor model with relatively few parameters, while the larger generator LLM responsible for producing task answers remains fixed and is not updated during training. This design ensures high training efficiency—for instance, training one epoch takes approximately 2 hours using eight H20 GPUs, and convergence is typically achieved within just two epochs. In contrast, other reinforcement learning-based methods, such as ReSearch (Chen et al., 2025a) and R1-Searcher (Song et al., 2025) , require direct fine-tuning of the large generator LLM, leading to considerably higher training time and resource consumption. Furthermore, it is important to emphasize that our method exhibits strong generalization capability. As shown in Section 3.3, a model trained only once demonstrates broad applicability, thereby reducing the need for frequent retraining and further lowering the overall training cost.", "source": "marker_v2", "marker_block_id": "/page/4/Text/7"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0032", "section": "2.2.5 EFFICIENCY ANALYSIS", "page_start": 5, "page_end": 5, "type": "Text", "text": "Inference Efficiency. Our method significantly enhances inference efficiency. In contrast to RAG approaches that do not employ a compressor—and thus require the generator LLM to directly encode lengthy documents, often spanning thousands of tokens—our approach introduces a lightweight compressor that processes long documents and summarizes them into compact representations of only a few dozen tokens before feeding them to the generator LLM. Since the compressor is an order of magnitude smaller in parameter size than the generator LLM, it substantially reduces the encoding time that would otherwise be incurred by the generator, leading to notable gains in inference efficiency. It is also important to note that the use of reinforcement learning does not adversely affect inference efficiency, as it is only involved during the training phase.", "source": "marker_v2", "marker_block_id": "/page/4/Text/8"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0033", "section": "3.1 EXPERIMENTAL SETTINGS", "page_start": 5, "page_end": 5, "type": "Text", "text": "Datasets and Evaluation Metrics. We evaluate our method on four benchmark datasets: two single-hop question-answering datasets, Natural Questions (NQ) (Kwiatkowski et al., 2019) and TriviaQA (Joshi et al., 2017) , as well as two multi-hop question-answering datasets, HotpotQA (Yang et al., 2018) and 2WikiMultihopQA (Ho et al., 2020) . Results are reported on the test sets of Natural Questions and TriviaQA, as well as the development sets of HotpotQA and 2WikiMultihopQA. Following RECOMP (Xu et al., 2024) , the performance is measured using Exact Match and token-level F1 scores, while efficiency is assessed by the number of tokens provided in the context.", "source": "marker_v2", "marker_block_id": "/page/4/Text/11"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0034", "section": "3.1 EXPERIMENTAL SETTINGS", "page_start": 5, "page_end": 5, "type": "Text", "text": "Compression Model (πθ). We trained our compression model using Qwen2.5-1.5B-Instruct to generate summaries of the retrieved documents. To evaluate the effect of using different models as compressors, we also trained compressors using Llama3.2-1B-Instruct and Llama3.2-3B-Instruct (Section 3.3) .", "source": "marker_v2", "marker_block_id": "/page/4/Text/12"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0035", "section": "3.1 EXPERIMENTAL SETTINGS", "page_start": 5, "page_end": 5, "type": "Text", "text": "Large Language Model (M). We use Qwen2.5-14B-Instruct as the primary LLM model M to generate predicted answers which are used to guide the training of the compressor. To evaluate the generalization ability of our method, we also transfer to another LLM model, LLama3.1-8B-Instruct (Table 2) .", "source": "marker_v2", "marker_block_id": "/page/4/Text/13"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0036", "section": "3.1 EXPERIMENTAL SETTINGS", "page_start": 6, "page_end": 6, "type": "TableGroup", "text": "Table 1: Open-domain QA results using Qwen2.5-14B-Instruct as the downstream LLM (M). The reported token counts represent the length of in-context documents, excluding few-shot examples. RECOMP, NoiseFilter-IB and our method CORE are all trained using Qwen2.5-1.5B-Instruct. NQ 7 TriviaQA L F IotpotQ. A 2Wiki Multiho pQA EM F1 # tok EM F1 # tok EM F1 # tok EM F1 # tok No Retrieval 21.36 30.97 0 53.23 59.98 0 21.05 29.48 0 26.11 29.51 0 RAG without compression on Top1 Document 34.46 44.41 142 60.82 68.70 143 29.20 38.93 147 26.79 31.87 153 Top3 Documents 37.78 48.45 427 62.60 71.02 430 31.64 41.87 442 27.89 33.58 460 Top5 Documents 38.03 49.16 712 64.10 72.48 715 32.99 43.69 737 29.64 35.21 766 Top10 Documents 38.67 50.03 1425 64.40 72.92 1428 33.95 44.88 1471 31.04 36.75 1531 Compression of top 5 do cuments ; BM25 25.23 36.47 37 55.36 63.90 39 24.18 35.73 71 25.42 30.29 68 Qwen2.5-1.5B 31.94 43.03 36 57.99 66.70 30 27.36 37.47 33 25.93 31.18 32 DeepSeek-V3 (671B) 37.73 50.39 54 64.13 73.20 50 33.59 44.83 48 27.99 32.67 92 RECOMP-Abs (1.5B) 34.18 46.26 58 60.31 68.50 53 28.96 39.95 56 30.25 36.73 52 RECOMP-Ext (1.5B) 33.84 46.05 56 60.18 68.39 48 29.93 41.09 45 30.78 37.07 51 NoiseFilter-IB (1.5B) 35.15 45.94 48 59.51 68.15 35 27.97 38.62 38 27.85 34.69 40 LongLLMLingua (1.5B) 33.65 43.15 152 58.96 66.82 148 28.03 38.49 149 29.37 33.62 153 QGC (1.5B) 36.23 45.88 49 61.02 68.45 47 29.16 40.05 45 31.14 36.83 51 CORE (1.5B) 41.02 50.40 46 65.63 72.55 32 33.67 45.06 36 36.72 42.05 49 Compression of top 10 a locumen ts (with t the comp ressor ti rained or ı top 5 d locs) BM25 25.91 36.88 38 55.28 63.16 37 23.49 35.01 68 25.61 30.54 65 Qwen2.5-1.5B 32.94 44.84 40 58.45 67.31 33 28.17 38.48 36 26.22 31.57 34 DeepSeek-V3 (671B) 37.79 51.07 56 65.29 74.45 53 34.62 45.69 50 29.00 34.64 40 RECOMP-Abs (1.5B) 34.40 46.93 59 61.42 69.88 52 31.54 42.92 52 31.98 38.16 49 RECOMP-Ext (1.5B) 33.96 46.34 60 61.03 69.51 50 31.92 43.18 55 32.52 38.87 44 NoiseFilter-IB (1.5B) 35.36 46.24 50 59.92 68.32 38 28.21 38.83 38 28.63 35.16 42 LongLLMLingua (1.5B) 33.78 43.37 154 59.17 66.97 150 28.33 38.95 148 29.62 34.11 151 QGC (1.5B) 36.03 45.62 50 61.23 68.74 49 29.12 39.63 46 31.71 37.52 50 CORE (1.5B) 41.88 51.26 52 66.76 73.64 33 34.68 46.35 37 37.99 43.28 48", "source": "marker_v2", "marker_block_id": "/page/5/TableGroup/211"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0037", "section": "3.1 EXPERIMENTAL SETTINGS", "page_start": 6, "page_end": 6, "type": "Text", "text": "Retrieval Corpus and Retrievers. Following previous studies (Xu et al., 2024), we use the Wikipedia corpus from December 20, 2018, as the retrieval source for all four datasets. The articles are segmented into non-overlapping 100-word documents. To ensure that our method is not dependent on a specific retriever, we experiment with several mainstream retrievers. Specifically, we use DPR (Karpukhin et al., 2020) for NQ, a hybrid of DPR and BM25 (Robertson et al., 1995) for TriviaQA, and the Contriever model (Izacard et al., 2021) trained on the MS MARCO dataset (Nguyen et al., 2016) for HotpotQA and 2WikiMultihopQA.", "source": "marker_v2", "marker_block_id": "/page/5/Text/3"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0038", "section": "3.1 EXPERIMENTAL SETTINGS", "page_start": 6, "page_end": 6, "type": "Text", "text": "Baselines. To evaluate the effectiveness of our method, we compared it against various baselines. First, we evaluated the uncompressed approach—retaining the original in-context RALM setup—by prepending the top 1, 3, 5, and 10 retrieved documents to the prompt. We also tested alternative compression methods, including the traditional BM25 algorithm (which ranks sentences by their similarity to the input), off-the-shelf Qwen2.5-1.5B-Instruct model (with comparable parameter size to our approach), and DeepSeek-V3 model (671B parameters, far exceeding our compressor's capacity). Furthermore, we included state-of-the-art context compression methods for RAG, RECOMP (Xu et al., 2024), NoiseFilter-IB (Zhu et al., 2024), LongLLMLingua (Jiang et al., 2024) and QGC (Cao et al., 2024). For RECOMP, we evaluated both its abstractive and extractive variants. To ensure a fair comparison, all trainable approaches were all trained using the same model.", "source": "marker_v2", "marker_block_id": "/page/5/Text/4"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0039", "section": "3.1 EXPERIMENTAL SETTINGS", "page_start": 6, "page_end": 6, "type": "Text", "text": "Implementation Details. For the distillation warm-up phase, we perform full-parameter supervised fine-tuning on the off-the-shell language model for two epochs using LLaMA-Factory ^1 . This warmed-up model then serves as the initializer for the subsequent reinforcement learning phase. We adopt the Verl framework ^2 for RL training. The initialized compression model is trained for two epochs on each dataset. Training is conducted on eight NVIDIA H20 GPUs using full parameter GRPO optimization, with a learning rate of 1e-5, a batch size of 256, five rollouts per sample, and a KL loss coefficient of 0.001. The downstream LLM (M) used for reward generation is served using the vLLM inference engine during RL training.", "source": "marker_v2", "marker_block_id": "/page/5/Text/5"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0040", "section": "3.1 EXPERIMENTAL SETTINGS", "page_start": 6, "page_end": 6, "type": "Footnote", "text": "& lt;sup>1</sup>", "source": "marker_v2", "marker_block_id": "/page/5/Footnote/6"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0041", "section": "3.1 EXPERIMENTAL SETTINGS", "page_start": 6, "page_end": 6, "type": "Footnote", "text": "& lt;sup>2</sup>", "source": "marker_v2", "marker_block_id": "/page/5/Footnote/7"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0042", "section": "3.2 OVERALL PERFORMANCE", "page_start": 7, "page_end": 7, "type": "Text", "text": "The detailed comparison results are presented in Table 1. We evaluate the following approaches: traditional RAG without compression using the top 1, 3, 5, and 10 documents prepended to the original input; BM25-based compression; off-the-shelf LLMs (Qwen2.5-1.5B-Instruct and DeepSeek-V3); and state-of-the-art trainable RAG compressors (RECOMP-Abstractive, RECOMP-Extractive, and NoiseFilter-IB). For fair comparison, all trainable methods were trained using the same backbone model, Qwen2.5-1.5B-Instruct. These compressors were trained on five-document inputs, and we report their performance both on in-domain five-document compression and out-of-domain generalization to ten-document compression.", "source": "marker_v2", "marker_block_id": "/page/6/Text/2"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0043", "section": "3.2 OVERALL PERFORMANCE", "page_start": 7, "page_end": 7, "type": "Text", "text": "RAG versus No RAG. As shown in Table 1, prepending the original input with retrieved documents yields a substantial improvement over the no-retrieval baseline, indicating that these documents provide valuable information for generating the answer. Performance improves as the number of retrieved documents increases from 1 to 10, albeit with diminishing returns—a trend attributable to the decreasing relevance of lower-ranked documents. However, this performance gain comes at the cost of a significant increase in the number of additional tokens the LLM must encode, rising from 0 to over 1,400 tokens.", "source": "marker_v2", "marker_block_id": "/page/6/Text/3"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0044", "section": "3.2 OVERALL PERFORMANCE", "page_start": 7, "page_end": 7, "type": "Text", "text": "Lossless Compression of CORE. Our compressor was trained using a context of five documents. As presented in Table 1, compared to prepending the full content of all five documents to the original input (i.e., the uncompressed baseline), our method achieves a high compression rate of approximately 6% with no loss in performance . Remarkably, on all four datasets, the approach not only maintains performance but also enhances EM by 1 to 7 points. For example, on NQ, the model achieves an EM of 41.02 with compressed input, surpassing the uncompressed score of 38.03.", "source": "marker_v2", "marker_block_id": "/page/6/Text/4"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0045", "section": "3.2 OVERALL PERFORMANCE", "page_start": 7, "page_end": 7, "type": "Text", "text": "Comparison with Compression Baselines. When compressing the top five documents, all baseline compression methods result in performance degradation to varying degrees compared to the uncompressed baseline. Specifically, BM25 leads to a substantial performance drop. Using the offthe-shelf Qwen2.5-1.5B-Instruct model for compression yields better results than BM25 but still underperforms relative to the uncompressed baseline. Surprisingly, even the large-scale DeepSeek-V3 model (with 671B parameters) achieves only near-lossless compression on TriviaQA and HotpotQA, while performance on NQ and 2Wiki remains below the uncompressed baseline. As for the trained compressors (RECOMP, NoiseFilter-IB, LongLLMLingua, and QGC), all of them exhibit performance degradation compared to no compression. The decline generally ranges from 2 to 6 EM poin ts across nearly all datasets. In contrast, our method, CORE, delivers the best performance. It not only surpasses compression methods of comparable size by 4-5 EM points but also maintains an advantage over the hundreds-of-times-larger DeepSeek model. This clearly demonstrates the benefit and importance of end-to-end optimization.", "source": "marker_v2", "marker_block_id": "/page/6/Text/5"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0046", "section": "3.2 OVERALL PERFORMANCE", "page_start": 7, "page_end": 7, "type": "Text", "text": "Generalization to 10-Doc Compression. When generalizing the trained compressors to handle the top-10 documents without retraining, the aforementioned conclusions remain valid. CORE continues to achieve lossless compression and performs best among all compression methods. On NQ, it achieves a token compression ratio of 3.6% while improving the EM by 3.2 points compared to using all ten documents. Similar trends are observed on TriviaQA, with a compression ratio of 2.3% and a gain of 2.4 EM points relative to the full-document baseline. On HotpotQA, CORE achieves a compression ratio of 2.5% and an improvement of 0.7 EM points. For 2WikiMultihopQA, it obtains a compression ratio of 3.1% along with a notable increase of 6.9 EM points.", "source": "marker_v2", "marker_block_id": "/page/6/Text/6"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0047", "section": "3.3 ROBUSTNESS AND GENERALIZATION ABILITY ANALYSIS", "page_start": 7, "page_end": 7, "type": "Text", "text": "Robustness of CORE Across Compressor Architectures. To evaluate whether the effectiveness of CORE depends on the choice of compressor model, we compared the performance of compressors trained using different model architectures—while keeping the downstream LLM (Qwen2.5-14B-Instruct) fixed. The compressors tested include LLaMA3.2-1B-Instruct, Qwen2.5-1.5B-Instruct, and LLaMA3.2-3B-Instruct, which vary in architecture and parameter count. As shown in Figure 3, the results indicate that: (1) These trained compressors consistently achieve lossless compression and outperform the uncompressed baseline (represented by the red reference line in the figure, which corresponds to prepending the full document content), confirming that our training framework is robust and not tied to a specific compressor architecture. (2) Compression performance improves as the size of the compressor model increases, consistent with scaling laws. More detailed results can be found in Tables 7 and 8.", "source": "marker_v2", "marker_block_id": "/page/6/Text/8"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0048", "section": "3.3 ROBUSTNESS AND GENERALIZATION ABILITY ANALYSIS", "page_start": 8, "page_end": 8, "type": "FigureGroup", "text": "Figure 3: The impact of different models used to train the compressor.", "source": "marker_v2", "marker_block_id": "/page/7/FigureGroup/208"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0049", "section": "3.3 ROBUSTNESS AND GENERALIZATION ABILITY ANALYSIS", "page_start": 8, "page_end": 8, "type": "TableGroup", "text": "Table 2: Zero-shot transfer of the trained compressor to Llama-3.1-8B-Instruct. NQ 7 TriviaQA HotpotQA 2WikiMultihopQA EM F1 # tok EM F1 # tok EM F1 # tok EM F1 # tok No Retrieval 24.04 34.91 0 55.64 62.57 0 19.93 27.75 0 27.64 31.18 0 RAG without compres ssion Top1 Document 33.80 44.06 142 59.17 67.50 143 27.95 37.49 147 28.41 33.43 153 Top3 Documents 36.87 47.81 427 61.13 70.06 430 30.17 40.71 442 28.67 34.23 460 Top5 Documents 37.65 48.87 712 62.26 71.04 715 31.44 42.16 737 29.43 35.18 766 Top10 Documents 38.12 49.93 1425 63.95 72.71 1428 32.19 42.62 1471 30.45 36.04 1531 Compression of top 5 docume nts Qwen2.5-1.5B 32.60 44.21 36 56.76 65.77 30 26.86 36.90 33 25.45 30.88 32 DeepSeek-V3 (671B) 37.56 50.11 54 62.52 72.34 50 33.05 44.25 48 28.64 33.87 92 RECOMP-Abs (1.5B) 33.41 45.50 58 58.50 67.37 53 28.85 39.76 56 31.63 37.81 52 RECOMP-Ext (1.5B) 33.12 45.06 60 57.98 66.84 55 29.03 40.04 52 31.85 38.02 55 CORE (1.5B) 40.72 50.00 46 64.08 71.13 32 32.17 43.71 36 35.99 41.42 49 Compression of top 1 0 docum ents Qwen2.5-1.5B 32.88 44.66 40 57.44 66.56 33 27.31 37.31 36 25.80 31.30 34 DeepSeek-V3 (671B) 37.49 51.28 56 63.79 73.80 53 34.24 45.35 50 31.45 37.09 40 RECOMP-Abs (1.5B) 34.18 46.80 59 59.69 68.89 52 30.17 41.42 55 33.61 39.78 44 RECOMP-Ext (1.5B) 34.06 46.55 60 59.33 68.71 50 30.52 41.98 55 33.52 39.42 44 CORE (1.5B) 41.77 51.27 52 65.25 72.45 33 33.25 45.09 37 37.59 42.87 48", "source": "marker_v2", "marker_block_id": "/page/7/TableGroup/209"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0050", "section": "3.3 ROBUSTNESS AND GENERALIZATION ABILITY ANALYSIS", "page_start": 8, "page_end": 8, "type": "Text", "text": "Transferability of CORE Across Downstream LLMs. We evaluate the transferability of our trained compressor and other baseline compressors to a new downstream LLM, the LLaMA-3.1-8B model. The results are presented in Table 2. Note that all trainable compressors, including our own, were trained using feedback generated by Qwen2.5-14B-Instruct. The findings reveal that existing trained compressors (e.g., RECOMP) exhibit limited generalization capability, as indicated by a larger performance gap relative to the baseline of prepending full documents. In contrast, CORE demonstrates stronger generalization, achieving lossless compression on the new downstream LLM. Specifically, it not only maintains performance but also surpasses the full-document baseline (i.e., no compression) across all four datasets while retaining a high compression rate. These results suggest that the summaries produced by our method are inherently high-quality and preserve key information necessary for accurate answering, thereby enabling effective transfer to other LLMs.", "source": "marker_v2", "marker_block_id": "/page/7/Text/5"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0051", "section": "3.4 ABLATION STUDY", "page_start": 8, "page_end": 8, "type": "Text", "text": "Table 3 presents an ablation study on the two stages of our method: distillation and GRPO. Here, \"w/o distillation\" denotes training the compressor with GRPO directly, bypassing the warm-start phase, while \"w/o GRPO\" indicates using only the distillation step without subsequent reinforcement learning. The results demonstrate that removing either component leads to performance degradation, confirming the necessity of both. In addition, the decline is more substantial when GRPO is omitted, highlighting the crucial role of reinforcement learning in the absence of explicit supervision. Distillation injects external knowledge into the model, providing a favorable starting point for RL training and thereby enabling RL to more effectively unlock the compressor's full potential.", "source": "marker_v2", "marker_block_id": "/page/7/Text/7"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0052", "section": "3.4 ABLATION STUDY", "page_start": 9, "page_end": 9, "type": "TableGroup", "text": "Table 3: Ablation study. Dataset Metric w/o distillation w/o GRPO CORE EM 36.37 34.18 41.02 NQ F1 46.91 46.26 50.40 EM 65.23 60.31 65.63 TQA F1 72.41 68.50 72.55 EM 32.01 28.96 33.67 HotpotQA F1 42.73 39.95 45.06 EM 31.40 30.25 36.72 2Wiki F1 36.89 36.73 42.05", "source": "marker_v2", "marker_block_id": "/page/8/TableGroup/495"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0053", "section": "3.4 ABLATION STUDY", "page_start": 9, "page_end": 9, "type": "FigureGroup", "text": "Figure 4: The impact of α.", "source": "marker_v2", "marker_block_id": "/page/8/FigureGroup/496"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0054", "section": "3.5 THE IMPACT OF α.", "page_start": 9, "page_end": 9, "type": "Text", "text": "Figure 4 illustrates the performance of our method under different values of α, which controls the weighting coefficient of the F1 reward term. Setting α = 0 corresponds to using only the EM reward. Results indicate that performance improves when α > 0 across all datasets, demonstrating the effectiveness of the F1 reward in mitigating the sparsity issue associated with the EM reward. However, the optimal value of α is dataset-dependent; values between 0.2 and 0.5 generally yield strong performance.", "source": "marker_v2", "marker_block_id": "/page/8/Text/6"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0055", "section": "3.6 CASE STUDY", "page_start": 9, "page_end": 9, "type": "Text", "text": "To conduct an in-depth analysis of the advantages of our compressor, we performed case studies on one single-hop QA dataset (NQ) and one multi-hop QA dataset (2Wiki), with the results presented in Table 4 and Table 5, respectively. For each case, we compared the summaries generated by off-the-shelf Qwen2.5-1.5B-Instruct, RECOMP, and our method CORE based on the same set of documents, as well as the predicted answers generated by the LLM after prepending these summaries. As shown in the tables, although the summaries produced by off-the-shelf Qwen2.5-1.5B are concise, they largely fail to capture key information relevant to answering the question. In contrast, RECOMP demonstrates better summarization capability but is prone to being overwhelmed by lengthy documents, resulting in misjudgments and even generating misleading information—such as the statement in Table 4: \" The U.S. stopped drafting for the Vietnam War after the Selective Service System was officially abolished in December 1972 \"—which leads the downstream LLM to produce the incorrect answer \"1972\". Our method, CORE, accurately extracts answer-critical information from lengthy documents, exemplified by the summary: \" The U.S. stopped drafting for the Vietnam War in 1973 after announcing the decision by Secretary of Defense Melvin Laird earlier that year \", thereby enabling the LLM to generate the correct answer \"1973\". This indicates that our compressor, trained with an end-task target-oriented reinforcement learning optimization strategy, can produce document summaries that are most helpful for answering the given question while effectively filtering out irrelevant information.", "source": "marker_v2", "marker_block_id": "/page/8/Text/8"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0056", "section": "4 RELATED WORK", "page_start": 9, "page_end": 9, "type": "Text", "text": "Context Compression in RAG. RAG enhances the performance of LLMs on knowledge-intensive tasks by retrieving the most relevant documents from extensive knowledge bases and prepending them as contextual information to the original input (Ram et al., 2023; Fan et al., 2024; Lin et al., 2023; Shi et al., 2023) . However, this approach requires the LLM to process significantly longer token sequences, resulting in increased computational costs. To mitigate this issue, researchers have begun to explore methods for compressing retrieved documents in RAG systems (Xu et al., 2024; Cao et al., 2024; Jin et al., 2024b; Zhu et al., 2024; Kim & Thorne, 2025; Rau et al., 2024; Wu et al., 2025; Louis et al., 2025; Jin et al., 2024a; Li et al., 2024a; b; Zhang et al., 2024a) . For instance, Xu et al. (2024) propose compressing retrieved documents into textual summaries before in-context augmentation, training the compressor through data selection and distillation. Similarly, Cao et al. (2024) introduce a Query-Guided Compressor (QGC) that uses queries to guide the compression process, effectively preserving essential information. Jin et al. (2024b) refine retrieved documents into Key Supporting Evidence (KSE) through a combination of knowledge synthesis, supervised fine-tuning (SFT), and preference alignment. Meanwhile, Zhu et al. (2024) present an informationtheoretic approach called NoiseFilter-IB, which filters noise by maximizing the mutual information", "source": "marker_v2", "marker_block_id": "/page/8/Text/10"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0057", "section": "4 RELATED WORK", "page_start": 10, "page_end": 10, "type": "Text", "text": "between the compressed content and the ground-truth output. Additionally, Kim & Thorne (2025) train a compressor to extract critical information using reward functions based on predefined heuristic rules. However, most of these methods are heuristic in nature, and due to the lack of ideal compression labels, the compressed content they produce cannot be guaranteed to benefit downstream LLMs. In contrast, our method, CORE, adopts an end-to-end optimization approach to address these limitations.", "source": "marker_v2", "marker_block_id": "/page/9/Text/1"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0058", "section": "4 RELATED WORK", "page_start": 10, "page_end": 10, "type": "Text", "text": "Reinforcement Learning. Reinforcement learning (RL) has recently achieved notable success, enabling LLMs to develop reasoning capabilities without explicit step-by-step supervision (Liu et al., 2024; Shao et al., 2024; Guo et al., 2025) . Building on these advances, several studies have applied RL to improve RAG (Ke et al., 2024) . For example, Kulkarni et al. (2024) use RL to autonomously decide whether to retrieve documents, while Zhang et al. (2024b) employ RL to optimize the ranking of retrieved documents. Similarly, Mao et al. (2024) propose a framework for training query rewriting models for RAG without relying on human annotations. MMOA-RAG (Chen et al., 2025b) enhances RAG through multi-agent reinforcement learning, incorporating a query rewriter, retriever, and generator. Meanwhile, RL has been applied to address other challenges in RAG, such as enhancing the quality of retrieved content. For instance, Oreo (Li & Ramakrishnan, 2025) trains a reconstructor with PPO and ROUGE-based rewards to rewrite passages for improved performance. In contrast, our work CORE tackles the distinct problem of computational efficiency. Our approach leverages the GRPO algorithm, direct task-performance rewards, and a lightweight compressor to enhance efficiency without compromising accuracy. Moreover, a line of research has utilized RL to integrate search with reasoning in a step-by-step manner (Singh et al., 2025) . For instance, Chen et al. (2025a) introduce a framework called ReSearch, which trains LLMs to reason with search using RL, without requiring supervised data for reasoning steps. Related approaches include R1- Searcher (Song et al., 2025) , WebThinker (Li et al., 2025) , and DeepResearcher (Zheng et al., 2025) . Although these methods are end-to-end, they differ fundamentally from our problem setting. These approaches typically involve directly training the LLM generator—which tends to be a large-scale model with a high parameter count. However, such training becomes infeasible when the model is a black box (e.g., GPT-4), as internal weights or gradients are inaccessible. Furthermore, these methods introduce extensive internal thinking processes that substantially increase context length and inference time. In contrast, our approach treats the generator LLM as a fixed black-box model and trains only a lightweight plug-in compressor to produce document summaries. This design significantly improves both training and inference efficiency. A parallel line of work applies RL to prompt compression. For example, PCRL (Jung & Kim, 2024) and TACO-RL (Shandilya et al., 2025) learn to compress prompts via token-level keep-or-drop decisions, using the similarity between model outputs with compressed and original prompts as the reward. Our work, CORE, introduces key distinctions in both objective and methodology. First, we target the more complex problem of compressing multiple retrieved documents in RAG, rather than single prompts. Second, we employ a generative compressor that can rephrase and synthesize content, instead of making token-level binary actions. Most importantly, we optimize compression using a direct task-performance reward with GRPO, which enables us to achieve true lossless compression at significantly higher ratios—a stark contrast to the performance degradation observed in prior prompt compression methods.", "source": "marker_v2", "marker_block_id": "/page/9/Text/2"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0059", "section": "5 CONCLUSION", "page_start": 10, "page_end": 10, "type": "Text", "text": "This paper analyzes the limitations of current context compression methods for RAG. A primary challenge is the lack of optimal reference summaries for supervised learning, which often results in performance degradation in downstream tasks. To overcome this, we frame context compression as a reinforcement learning problem, utilizing downstream task performance as a reward signal to train the compression policy, thereby enabling end-to-end optimization. Extensive experiments demonstrate that our proposed method, CORE, achieves effectively lossless compression by maintaining a high compression ratio while preserving original task performance. Surprisingly, CORE not only preserves but actually enhances performance on all downstream tasks. Further in-depth analysis provides additional insights into the efficacy of our approach.", "source": "marker_v2", "marker_block_id": "/page/9/Text/4"}

iclr26/6FyYLNnDL6/marker_meta.json

@@ -0,0 +1,1816 @@
+{
+  "table_of_contents": [
+    {
+      "title": "CORE: PERFORMANCE-LOSSLESS CONTEXT COM-\nPRESSION FOR RETRIEVAL-AUGMENTED GENERA-\nTION",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          106.3828125,
+          79.6640625
+        ],
+        [
+          503.5697326660156,
+          79.6640625
+        ],
+        [
+          503.5697326660156,
+          136.72955322265625
+        ],
+        [
+          106.3828125,
+          136.72955322265625
+        ]
+      ]
+    },
+    {
+      "title": "ABSTRACT",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          277.3125,
+          207.63934326171875
+        ],
+        [
+          333.72119140625,
+          207.63934326171875
+        ],
+        [
+          333.72119140625,
+          219.59454345703125
+        ],
+        [
+          277.3125,
+          219.59454345703125
+        ]
+      ]
+    },
+    {
+      "title": "1 INTRODUCTION",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          107.578125,
+          453.62109375
+        ],
+        [
+          205.98886108398438,
+          453.62109375
+        ],
+        [
+          205.98886108398438,
+          466.3693542480469
+        ],
+        [
+          107.578125,
+          466.3693542480469
+        ]
+      ]
+    },
+    {
+      "title": "2 CORE-RAG",
+      "heading_level": null,
+      "page_id": 1,
+      "polygon": [
+        [
+          107.578125,
+          667.0570220947266
+        ],
+        [
+          191.63870239257812,
+          667.0570220947266
+        ],
+        [
+          191.63870239257812,
+          679.0122222900391
+        ],
+        [
+          107.578125,
+          679.0122222900391
+        ]
+      ]
+    },
+    {
+      "title": "2.1 Problem Formulation",
+      "heading_level": null,
+      "page_id": 2,
+      "polygon": [
+        [
+          107.25,
+          337.9921875
+        ],
+        [
+          240.0,
+          337.9921875
+        ],
+        [
+          240.0,
+          346.5
+        ],
+        [
+          107.25,
+          346.5
+        ]
+      ]
+    },
+    {
+      "title": "2.2 Training the Compressor",
+      "heading_level": null,
+      "page_id": 2,
+      "polygon": [
+        [
+          107.25,
+          472.18359375
+        ],
+        [
+          257.25,
+          472.18359375
+        ],
+        [
+          257.25,
+          481.5
+        ],
+        [
+          107.25,
+          481.5
+        ]
+      ]
+    },
+    {
+      "title": "2.2.1 DISTILLATION FOR WARM-START",
+      "heading_level": null,
+      "page_id": 2,
+      "polygon": [
+        [
+          106.5,
+          580.46484375
+        ],
+        [
+          284.25,
+          580.46484375
+        ],
+        [
+          284.25,
+          591.0
+        ],
+        [
+          106.5,
+          591.0
+        ]
+      ]
+    },
+    {
+      "title": "2.2.2 END-TO-END TRAINING WITH RL",
+      "heading_level": null,
+      "page_id": 3,
+      "polygon": [
+        [
+          105.1875,
+          169.3828125
+        ],
+        [
+          285.0,
+          169.3828125
+        ],
+        [
+          285.0,
+          178.5
+        ],
+        [
+          105.1875,
+          178.5
+        ]
+      ]
+    },
+    {
+      "title": "2.2.3 REWARD CALCULATION",
+      "heading_level": null,
+      "page_id": 3,
+      "polygon": [
+        [
+          107.578125,
+          462.75
+        ],
+        [
+          243.84375,
+          462.75
+        ],
+        [
+          243.84375,
+          471.41015625
+        ],
+        [
+          107.578125,
+          471.41015625
+        ]
+      ]
+    },
+    {
+      "title": "2.2.4 TRAINING TEMPLATE",
+      "heading_level": null,
+      "page_id": 4,
+      "polygon": [
+        [
+          107.279296875,
+          146.1796875
+        ],
+        [
+          232.19337463378906,
+          146.1796875
+        ],
+        [
+          232.19337463378906,
+          157.8551025390625
+        ],
+        [
+          107.279296875,
+          157.8551025390625
+        ]
+      ]
+    },
+    {
+      "title": "2.2.5 EFFICIENCY ANALYSIS",
+      "heading_level": null,
+      "page_id": 4,
+      "polygon": [
+        [
+          106.98046875,
+          248.58648681640625
+        ],
+        [
+          237.5297393798828,
+          248.58648681640625
+        ],
+        [
+          237.5297393798828,
+          258.549072265625
+        ],
+        [
+          106.98046875,
+          258.549072265625
+        ]
+      ]
+    },
+    {
+      "title": "3 EXPERIMENTS",
+      "heading_level": null,
+      "page_id": 4,
+      "polygon": [
+        [
+          106.98046875,
+          522.8830871582031
+        ],
+        [
+          200.08351135253906,
+          522.8830871582031
+        ],
+        [
+          200.08351135253906,
+          534.8382873535156
+        ],
+        [
+          106.98046875,
+          534.8382873535156
+        ]
+      ]
+    },
+    {
+      "title": "3.1 EXPERIMENTAL SETTINGS",
+      "heading_level": null,
+      "page_id": 4,
+      "polygon": [
+        [
+          107.578125,
+          543.33984375
+        ],
+        [
+          244.3253936767578,
+          543.33984375
+        ],
+        [
+          244.3253936767578,
+          553.9218444824219
+        ],
+        [
+          107.578125,
+          553.9218444824219
+        ]
+      ]
+    },
+    {
+      "title": "3.2 OVERALL PERFORMANCE",
+      "heading_level": null,
+      "page_id": 6,
+      "polygon": [
+        [
+          106.083984375,
+          82.7578125
+        ],
+        [
+          241.21005249023438,
+          82.7578125
+        ],
+        [
+          241.21005249023438,
+          94.2310791015625
+        ],
+        [
+          106.083984375,
+          94.2310791015625
+        ]
+      ]
+    },
+    {
+      "title": "3.3 ROBUSTNESS AND GENERALIZATION ABILITY ANALYSIS",
+      "heading_level": null,
+      "page_id": 6,
+      "polygon": [
+        [
+          106.3828125,
+          595.93359375
+        ],
+        [
+          374.51373291015625,
+          595.93359375
+        ],
+        [
+          374.51373291015625,
+          606.733642578125
+        ],
+        [
+          106.3828125,
+          606.733642578125
+        ]
+      ]
+    },
+    {
+      "title": "3.4 ABLATION STUDY",
+      "heading_level": null,
+      "page_id": 7,
+      "polygon": [
+        [
+          105.486328125,
+          626.87109375
+        ],
+        [
+          209.25,
+          626.87109375
+        ],
+        [
+          209.25,
+          636.75
+        ],
+        [
+          105.486328125,
+          636.75
+        ]
+      ]
+    },
+    {
+      "title": "3.5 THE IMPACT OF \u03b1.",
+      "heading_level": null,
+      "page_id": 8,
+      "polygon": [
+        [
+          106.681640625,
+          234.58087158203125
+        ],
+        [
+          210.57858276367188,
+          234.58087158203125
+        ],
+        [
+          210.57858276367188,
+          244.77398681640625
+        ],
+        [
+          106.681640625,
+          244.77398681640625
+        ]
+      ]
+    },
+    {
+      "title": "3.6 CASE STUDY",
+      "heading_level": null,
+      "page_id": 8,
+      "polygon": [
+        [
+          107.279296875,
+          324.0703125
+        ],
+        [
+          187.52699279785156,
+          324.0703125
+        ],
+        [
+          187.52699279785156,
+          334.8929443359375
+        ],
+        [
+          107.279296875,
+          334.8929443359375
+        ]
+      ]
+    },
+    {
+      "title": "4 RELATED WORK",
+      "heading_level": null,
+      "page_id": 8,
+      "polygon": [
+        [
+          106.083984375,
+          548.9309844970703
+        ],
+        [
+          211.1956787109375,
+          548.9309844970703
+        ],
+        [
+          211.1956787109375,
+          560.8861846923828
+        ],
+        [
+          106.083984375,
+          560.8861846923828
+        ]
+      ]
+    },
+    {
+      "title": "5 CONCLUSION",
+      "heading_level": null,
+      "page_id": 9,
+      "polygon": [
+        [
+          105.78515625,
+          549.52734375
+        ],
+        [
+          195.37747192382812,
+          549.52734375
+        ],
+        [
+          195.37747192382812,
+          563.3050231933594
+        ],
+        [
+          105.78515625,
+          563.3050231933594
+        ]
+      ]
+    },
+    {
+      "title": "ETHICS STATEMENT",
+      "heading_level": null,
+      "page_id": 10,
+      "polygon": [
+        [
+          104.291015625,
+          81.2109375
+        ],
+        [
+          209.38401794433594,
+          81.2109375
+        ],
+        [
+          209.38401794433594,
+          94.7125244140625
+        ],
+        [
+          104.291015625,
+          94.7125244140625
+        ]
+      ]
+    },
+    {
+      "title": "REPRODUCIBILITY STATEMENT",
+      "heading_level": null,
+      "page_id": 10,
+      "polygon": [
+        [
+          108.17578125,
+          167.98126220703125
+        ],
+        [
+          267.8599853515625,
+          167.98126220703125
+        ],
+        [
+          267.8599853515625,
+          179.93646240234375
+        ],
+        [
+          108.17578125,
+          179.93646240234375
+        ]
+      ]
+    },
+    {
+      "title": "REFERENCES",
+      "heading_level": null,
+      "page_id": 10,
+      "polygon": [
+        [
+          106.3828125,
+          264.1632080078125
+        ],
+        [
+          175.2598114013672,
+          264.1632080078125
+        ],
+        [
+          175.2598114013672,
+          276.118408203125
+        ],
+        [
+          106.3828125,
+          276.118408203125
+        ]
+      ]
+    },
+    {
+      "title": "[Instruction] Answer the question.",
+      "heading_level": null,
+      "page_id": 13,
+      "polygon": [
+        [
+          117.439453125,
+          213.85546875
+        ],
+        [
+          259.5873718261719,
+          213.85546875
+        ],
+        [
+          259.5873718261719,
+          224.05865478515625
+        ],
+        [
+          117.439453125,
+          224.05865478515625
+        ]
+      ]
+    },
+    {
+      "title": "[Examples]",
+      "heading_level": null,
+      "page_id": 13,
+      "polygon": [
+        [
+          116.244140625,
+          274.81878662109375
+        ],
+        [
+          165.271240234375,
+          274.81878662109375
+        ],
+        [
+          165.271240234375,
+          285.0151062011719
+        ],
+        [
+          116.244140625,
+          285.0151062011719
+        ]
+      ]
+    },
+    {
+      "title": "[Current Question]",
+      "heading_level": null,
+      "page_id": 13,
+      "polygon": [
+        [
+          116.841796875,
+          365.0342102050781
+        ],
+        [
+          195.42184448242188,
+          365.0342102050781
+        ],
+        [
+          195.42184448242188,
+          375.2305603027344
+        ],
+        [
+          116.841796875,
+          375.2305603027344
+        ]
+      ]
+    },
+    {
+      "title": "A USE OF LLMS",
+      "heading_level": null,
+      "page_id": 13,
+      "polygon": [
+        [
+          106.3828125,
+          459.80859375
+        ],
+        [
+          202.77764892578125,
+          459.80859375
+        ],
+        [
+          202.77764892578125,
+          472.3905334472656
+        ],
+        [
+          106.3828125,
+          472.3905334472656
+        ]
+      ]
+    },
+    {
+      "title": "B PROMPT TEMPLATES",
+      "heading_level": null,
+      "page_id": 13,
+      "polygon": [
+        [
+          108.17578125,
+          545.9012756347656
+        ],
+        [
+          236.07421875,
+          545.9012756347656
+        ],
+        [
+          236.07421875,
+          557.8564758300781
+        ],
+        [
+          108.17578125,
+          557.8564758300781
+        ]
+      ]
+    },
+    {
+      "title": "C CASE STUDY",
+      "heading_level": null,
+      "page_id": 13,
+      "polygon": [
+        [
+          105.78515625,
+          642.3252868652344
+        ],
+        [
+          195.26687622070312,
+          642.3252868652344
+        ],
+        [
+          195.26687622070312,
+          654.2804870605469
+        ],
+        [
+          105.78515625,
+          654.2804870605469
+        ]
+      ]
+    },
+    {
+      "title": "Table 4: Case study on NQ dataset.",
+      "heading_level": null,
+      "page_id": 14,
+      "polygon": [
+        [
+          234.28125,
+          88.55859375
+        ],
+        [
+          376.321044921875,
+          88.55859375
+        ],
+        [
+          376.321044921875,
+          99.5140380859375
+        ],
+        [
+          234.28125,
+          99.5140380859375
+        ]
+      ]
+    },
+    {
+      "title": "Question: when did the us stop drafting for the vietnam war? Gold answer: [1973]",
+      "heading_level": null,
+      "page_id": 14,
+      "polygon": [
+        [
+          113.9780044555664,
+          113.967529296875
+        ],
+        [
+          464.48211669921875,
+          113.967529296875
+        ],
+        [
+          464.48211669921875,
+          130.7109375
+        ],
+        [
+          113.9780044555664,
+          130.7109375
+        ]
+      ]
+    },
+    {
+      "title": "Top-5 documents:",
+      "heading_level": null,
+      "page_id": 14,
+      "polygon": [
+        [
+          111.1640625,
+          128.77734375
+        ],
+        [
+          189.7226104736328,
+          128.77734375
+        ],
+        [
+          189.7226104736328,
+          139.98199462890625
+        ],
+        [
+          111.1640625,
+          139.98199462890625
+        ]
+      ]
+    },
+    {
+      "title": "Table 5: Case study on 2Wiki dataset.",
+      "heading_level": null,
+      "page_id": 15,
+      "polygon": [
+        [
+          230.09765625,
+          93.97265625
+        ],
+        [
+          381.90234375,
+          93.97265625
+        ],
+        [
+          381.90234375,
+          104.9940185546875
+        ],
+        [
+          230.09765625,
+          104.9940185546875
+        ]
+      ]
+    },
+    {
+      "title": "Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? \u00b4\nGold answer: [Charles Armand Rene de La Tr \u00b4 emoille] \u00b4",
+      "heading_level": null,
+      "page_id": 15,
+      "polygon": [
+        [
+          112.359375,
+          119.32745361328125
+        ],
+        [
+          468.9930114746094,
+          119.32745361328125
+        ],
+        [
+          468.9930114746094,
+          140.3680419921875
+        ],
+        [
+          112.359375,
+          140.3680419921875
+        ]
+      ]
+    },
+    {
+      "title": "D IMPACT OF DIFFERENT COMPRESSORS ON PERFORMANCE",
+      "heading_level": null,
+      "page_id": 16,
+      "polygon": [
+        [
+          108.29900360107422,
+          548.7171173095703
+        ],
+        [
+          424.6056823730469,
+          548.7171173095703
+        ],
+        [
+          424.6056823730469,
+          560.6723175048828
+        ],
+        [
+          108.29900360107422,
+          560.6723175048828
+        ]
+      ]
+    },
+    {
+      "title": "E CROSS-DATASET GENERALIZATION PERFORMANCE",
+      "heading_level": null,
+      "page_id": 18,
+      "polygon": [
+        [
+          106.98046875,
+          325.23046875
+        ],
+        [
+          390.4736633300781,
+          325.23046875
+        ],
+        [
+          390.4736633300781,
+          338.1045227050781
+        ],
+        [
+          106.98046875,
+          338.1045227050781
+        ]
+      ]
+    },
+    {
+      "title": "F ROBUSTNESS AGAINST NOISY CONTEXTS",
+      "heading_level": null,
+      "page_id": 18,
+      "polygon": [
+        [
+          107.578125,
+          444.33984375
+        ],
+        [
+          339.8441467285156,
+          444.33984375
+        ],
+        [
+          339.8441467285156,
+          456.41046142578125
+        ],
+        [
+          107.578125,
+          456.41046142578125
+        ]
+      ]
+    }
+  ],
+  "page_stats": [
+    {
+      "page_id": 0,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          207
+        ],
+        [
+          "Line",
+          87
+        ],
+        [
+          "Text",
+          14
+        ],
+        [
+          "SectionHeader",
+          3
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 1,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          193
+        ],
+        [
+          "Line",
+          81
+        ],
+        [
+          "Text",
+          13
+        ],
+        [
+          "Figure",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "SectionHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "FigureGroup",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 2,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "Line",
+          186
+        ],
+        [
+          "Span",
+          110
+        ],
+        [
+          "Text",
+          46
+        ],
+        [
+          "SectionHeader",
+          3
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Figure",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "FigureGroup",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 3,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "Line",
+          145
+        ],
+        [
+          "Span",
+          107
+        ],
+        [
+          "Text",
+          46
+        ],
+        [
+          "Equation",
+          3
+        ],
+        [
+          "SectionHeader",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 4,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          277
+        ],
+        [
+          "Line",
+          97
+        ],
+        [
+          "Text",
+          15
+        ],
+        [
+          "SectionHeader",
+          4
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Equation",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 5,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "TableCell",
+          364
+        ],
+        [
+          "Line",
+          118
+        ],
+        [
+          "Span",
+          55
+        ],
+        [
+          "Text",
+          32
+        ],
+        [
+          "Footnote",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "TableGroup",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 6,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          222
+        ],
+        [
+          "Line",
+          94
+        ],
+        [
+          "Text",
+          15
+        ],
+        [
+          "SectionHeader",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 7,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "TableCell",
+          253
+        ],
+        [
+          "Line",
+          128
+        ],
+        [
+          "Span",
+          47
+        ],
+        [
+          "Text",
+          26
+        ],
+        [
+          "Caption",
+          2
+        ],
+        [
+          "Reference",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Figure",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "SectionHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "FigureGroup",
+          1
+        ],
+        [
+          "TableGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 8,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          358
+        ],
+        [
+          "Line",
+          124
+        ],
+        [
+          "TableCell",
+          45
+        ],
+        [
+          "Text",
+          4
+        ],
+        [
+          "SectionHeader",
+          3
+        ],
+        [
+          "Caption",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "Figure",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "TableGroup",
+          1
+        ],
+        [
+          "FigureGroup",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 9,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          224
+        ],
+        [
+          "Line",
+          93
+        ],
+        [
+          "Text",
+          11
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "SectionHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 10,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          205
+        ],
+        [
+          "Line",
+          87
+        ],
+        [
+          "ListItem",
+          11
+        ],
+        [
+          "Reference",
+          11
+        ],
+        [
+          "Text",
+          10
+        ],
+        [
+          "SectionHeader",
+          3
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "ListGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 11,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          227
+        ],
+        [
+          "Line",
+          91
+        ],
+        [
+          "ListItem",
+          17
+        ],
+        [
+          "Reference",
+          17
+        ],
+        [
+          "Text",
+          9
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "ListGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 12,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          218
+        ],
+        [
+          "Line",
+          86
+        ],
+        [
+          "ListItem",
+          14
+        ],
+        [
+          "Reference",
+          14
+        ],
+        [
+          "Text",
+          7
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "ListGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 13,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          211
+        ],
+        [
+          "Line",
+          88
+        ],
+        [
+          "Text",
+          22
+        ],
+        [
+          "SectionHeader",
+          6
+        ],
+        [
+          "Caption",
+          2
+        ],
+        [
+          "Reference",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 14,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          221
+        ],
+        [
+          "Line",
+          107
+        ],
+        [
+          "Text",
+          15
+        ],
+        [
+          "TableCell",
+          12
+        ],
+        [
+          "SectionHeader",
+          3
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 15,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          347
+        ],
+        [
+          "Line",
+          109
+        ],
+        [
+          "Text",
+          17
+        ],
+        [
+          "TableCell",
+          12
+        ],
+        [
+          "SectionHeader",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 16,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          227
+        ],
+        [
+          "Line",
+          102
+        ],
+        [
+          "TableCell",
+          16
+        ],
+        [
+          "Text",
+          12
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "SectionHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "TableGroup",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 17,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "TableCell",
+          456
+        ],
+        [
+          "Line",
+          73
+        ],
+        [
+          "Span",
+          62
+        ],
+        [
+          "Text",
+          8
+        ],
+        [
+          "Caption",
+          2
+        ],
+        [
+          "Table",
+          2
+        ],
+        [
+          "TableGroup",
+          2
+        ],
+        [
+          "Reference",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 18,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          204
+        ],
+        [
+          "Line",
+          84
+        ],
+        [
+          "TableCell",
+          60
+        ],
+        [
+          "Text",
+          8
+        ],
+        [
+          "Caption",
+          2
+        ],
+        [
+          "Table",
+          2
+        ],
+        [
+          "SectionHeader",
+          2
+        ],
+        [
+          "Reference",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "TableGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    }
+  ],
+  "debug_data_path": "debug_data/6FyYLNnDL6"
+}

iclr26/6FyYLNnDL6/model_text_v3.txt

@@ -0,0 +1,179 @@
+[p. 1 | section: ABSTRACT | type: Text]
+Retrieval-Augmented Generation (RAG) has emerged as a promising approach to enhance the timeliness of knowledge updates and the factual accuracy of responses in large language models. However, incorporating a large number of retrieved documents significantly increases input length, leading to higher computational costs. Existing approaches to document compression tailored for RAG often degrade task performance, as they typically rely on predefined heuristics in the absence of clear compression guidelines. These heuristics fail to ensure that the compressed content effectively supports downstream tasks. To address these limitations, we propose CORE, a novel method for lossless context compression in RAG. CORE is optimized end-to-end and does not depend on predefined compression labels, which are often impractical to obtain. Instead, it leverages downstream task performance as a feedback signal, iteratively refining the compression policy to enhance task effectiveness. Extensive experiments across four datasets demonstrate the effectiveness of CORE. With a high compression ratio of 3%, CORE not only prevents performance degradation compared to including full documents (i.e., without compression) but also improves the average Exact Match (EM) score by 3.3 points. The code for CORE is available at .
+
+[p. 1 | section: 1 INTRODUCTION | type: Text]
+Large language models (LLMs) have undergone rapid development in recent years, significantly enhancing performance across various language tasks due to their emergent capabilities in semantic understanding and reasoning. Nevertheless, LLMs still face challenges in updating knowledge and providing factual responses (Fan et al., 2024) . To address these issues, Retrieval-Augmented Generation (RAG) has emerged as a promising approach. By retrieving the most relevant documents from external knowledge bases and prepending them as contextual information to the original input, RAG substantially improves LLM performance on knowledge-intensive tasks (Ram et al., 2023) .
+
+[p. 1 | section: 1 INTRODUCTION | type: Text]
+While RAG enhances performance, its effectiveness is closely tied to the number of retrieved documents used, since a broader context increases the probability of encompassing critical evidence. As illustrated in Figure 1, performance was weakest without any retrieved documents (i.e., without RAG). Accuracy improved consistently as more documents were added to the context, ultimately exceeding the no-RAG baseline by over 10 Exact Match (EM) points. However, this performance gain came with two significant limitations: (1) a substantial increase in computational cost from processing a larger number of context tokens (Xu et al., 2024) , and (2) the model's difficulty in effectively leveraging all provided documents, often resulting in the omission of key information located in the middle of the context (Liu et al., 2023) .
+
+[p. 1 | section: 1 INTRODUCTION | type: Text]
+These limitations have motivated recent research efforts aimed at compressing the retrieved context (Jin et al., 2024b; Wu et al., 2025; Jin et al., 2024a; Zhang et al., 2024a) . Prominent approaches include document summarization (Xu et al., 2024) , key information extraction (Cao et al., 2024; Xu et al., 2024) , the construction of key supporting evidence (Jin et al., 2024b) , and noise filtering based on information theory (Zhu et al., 2024) . Despite recent progress, these methods have several notable shortcomings. First, compression often results in a performance trade-off. For instance, RECOMP
+
+[p. 2 | section: 1 INTRODUCTION | type: FigureGroup]
+Figure 1: Performance evolution with an increasing number of retrieved documents on two datasets. Traditional RAG requires more documents for better performance, while our method achieves comparable or superior results with significant token compression.
+
+[p. 2 | section: 1 INTRODUCTION | type: Text]
+(Xu et al., 2024) suffers a 3–5 point drop in EM score (Table 1) , making it unsuitable for accuracysensitive applications. Second, most compression methods are heuristic in nature. The models are typically trained to generate summaries that are generally good but not necessarily useful for the downstream answer-generation LLM. This limitation arises from the lack of an ideal supervisory signal that defines what an optimal summary should be for the downstream task. This fundamental gap hinders the end-to-end optimization of existing methods. Finally, some compression models (Zhu et al., 2024) have parameter counts comparable to the LLM that performs the end task, resulting in substantial computational costs that undermine the efficiency gains of compression.
+
+[p. 2 | section: 1 INTRODUCTION | type: Text]
+Addressing these critical shortcomings requires a method that aligns compression with downstream task requirements, thereby minimizing performance trade-offs. To bridge this gap, we propose CORE, a novel method designed to achieve lossless context compression for RAG. Unlike previous compression methods, CORE is optimized in an end-to-end manner. Since obtaining predefined summary labels for supervision is impractical, we instead use downstream task performance as a feedback signal to evaluate the compression model's output. This feedback enables iterative refinement of the compression policy, guiding the compression model toward improved downstream performance. To this end, we employ Group Relative Policy Optimization (GRPO), a technique particularly well-suited for this purpose (Liu et al., 2024; Shao et al., 2024; Chen et al., 2025a) . In our framework, the accuracy of the downstream QA task is formalized as a reward, with the compression policy optimized through group-wise relative comparisons. Furthermore, our compression model is substantially smaller than the downstream LLM, which significantly reduces the computational overhead associated with encoding retrieved documents.
+
+[p. 2 | section: 1 INTRODUCTION | type: Text]
+We evaluate CORE on four benchmark datasets: two single-hop QA datasets ( Natural Questions and TriviaQA ) and two multi-hop datasets ( HotpotQA and 2WikiMultihopQA ). As shown in Table 1, CORE achieves state-of-the-art performance across all baselines. With a compression ratio of 3%, our approach not only avoids performance degradation compared to prepending full documents but also improves the average EM score by 3.3 points. We further demonstrate two key advantages of our approach. First, the effectiveness of CORE is not tied to a specific model architecture, as demonstrated by the fact that lossless compression can be achieved when various models are trained as compressors (Figure 3) . Second, the compressor exhibits strong transferability: a compression model trained using feedback from one LLM generalizes effectively to other LLMs (Table 2) . Finally, an in-depth case study (Tables 4 and 5) provides a qualitative analysis of the benefits of CORE.
+
+[p. 2 | section: 2 CORE-RAG | type: Text]
+This section introduces our proposed method, COmpression via REinforcement learning (CORE), which is shown in Figure 2. First, we provide an overview of the entire workflow. Then, we detail the end-to-end training strategy for the compression model, which is designed to drastically reduce the number of document tokens while preserving task performance.
+
+[p. 3 | section: 2 CORE-RAG | type: FigureGroup]
+Figure 2: Overview of our method CORE. The upper section illustrates the inference pipeline. The lower section depicts the end-to-end training method for the compression model.
+
+[p. 3 | section: 2.1 Problem Formulation | type: Text]
+We adopt the same problem formulation as prior work (Xu et al., 2024). Given an input question q, a target output y, and a set of k retrieved documents D, our objective is to compress D with respect to q into a summary s that preserves the most useful information while using significantly fewer tokens than D. This summary s is then prepended to the original input q and fed into an LLM to generate the final answer for the downstream task. This overall pipeline is illustrated in the upper half of Figure 2. The process involves two key components: a compression model \pi_{\theta} \colon (q, D) \mapsto s and a large language model M \colon (s, q) \mapsto \hat{y} , which generates the predicted answer \hat{y} . We treat M as a black-box system and focus exclusively on training the compressor \pi_{\theta} . The compressor itself is also a language model, but it is intentionally designed to be significantly smaller than M to reduce the computational cost of encoding the retrieved documents.
+
+[p. 3 | section: 2.2 Training the Compressor | type: Text]
+Our compressor is designed to generate document summaries that are highly useful to the LLM (M) for downstream tasks. This objective is challenging because the criteria for an effective summary are task-dependent, and direct supervision is unavailable. We therefore formulate this as an end-to-end training problem and employ reinforcement learning to optimize the compressor without relying on pre-defined compression labels. The overall architecture of our training framework is illustrated in the lower portion of Figure 2. The following sections describe the key components of our approach: distillation warm-up, policy optimization, and reward calculation.
+
+[p. 3 | section: 2.2.1 DISTILLATION FOR WARM-START | type: Text]
+Due to the limited parameter size of our compression model, its capability for question-focused document summarization is constrained. To provide a strong initial policy for RL and ensure training stability, we employ knowledge distillation from a teacher model to initialize our compressor. Specifically, we first utilize a large-scale language model (DeepSeek-V3) as the teacher to generate summaries of retrieved documents related to the given question. We then evaluate the performance of the downstream LLM (M) on the QA task under two conditions: (1) with the teacher-generated summary \hat{s} prepended to the input question q, and (2) with the original question alone. The corresponding performance scores are denoted as p_{\text{summary}} and p_{\text{original}} , respectively. By comparing these results, we retain instances where p_{\text{summary}} > p_{\text{original}} , indicating that the summary enhances RAG performance. We also retain cases where p_{\text{original}} = 1 (i.e., the model produces a fully correct answer without the summary) and p_{\text{summary}} < p_{\text{original}} ; for these, we set the target summary \hat{s} to an empty string. All other instances are discarded. The resulting filtered and modified dataset is denoted as
+
+[p. 4 | section: 2.2.1 DISTILLATION FOR WARM-START | type: Text]
+\mathcal{X}_f , which is used for supervised fine-tuning of the compression model. The fine-tuning objective is defined as:
+
+[p. 4 | section: 2.2.1 DISTILLATION FOR WARM-START | type: Text]
+\mathcal{L}_{\text{distill}} = \frac{1}{|\mathcal{X}_f|} \sum_{(q,D,\hat{s}) \in \mathcal{X}_f} \mathcal{L}_{\text{CE}}(\pi_{\theta}(q,D),\hat{s}), \tag{1}
+
+[p. 4 | section: 2.2.1 DISTILLATION FOR WARM-START | type: Text]
+where \pi_{\theta}(q, D) denotes the output of the compression model and \mathcal{L}_{CE} is the cross-entropy loss. This distillation process yields a robust initialization for RL and promotes stability in subsequent training.
+
+[p. 4 | section: 2.2.2 END-TO-END TRAINING WITH RL | type: Text]
+Following the distillation phase, the compressor possesses a preliminary compression capability. However, as summaries from even the largest teacher models are not guaranteed to be optimal for the downstream task, further end-to-end optimization is necessary. We therefore formulate this optimization as a reinforcement learning problem. In this framework, the compressor functions as a policy that generates a summary from an input question and its corresponding documents. A reward function, which directly reflects performance on the downstream task (e.g., question-answering accuracy), then evaluates the summary's quality. The objective is to optimize the compressor's parameters to maximize the expected cumulative reward, thereby directly aligning its outputs with the downstream task's objectives.
+
+[p. 4 | section: 2.2.2 END-TO-END TRAINING WITH RL | type: Text]
+Specifically, we employ Group Relative Policy Optimization (GRPO) (Shao et al., 2024) algorithm. Unlike Proximal Policy Optimization (PPO), which trains a separate critic model, GRPO estimates the baseline directly from a group of rollouts. Given an existing policy, \pi_{\theta_{\text{old}}} , and a reference policy, \pi_{\theta_{\text{ref}}} , the GRPO objective maximizes the compressor policy \pi_{\theta} using G rollouts \tau = \{s_i\}_{i=1}^G \sim \pi_{\theta_{\text{old}}}(\cdot|x) , for each input x \sim \mathcal{D} :
+
+[p. 4 | section: 2.2.2 END-TO-END TRAINING WITH RL | type: Equation]
+\mathcal{J}(\theta) = \mathbb{E}_{x \sim \mathcal{D}, \ \{s_i\}_{i=1}^G \sim \pi_{\theta_{\text{old}}}(\cdot \mid x)} \frac{1}{G} \sum_{i=1}^G \left[ \min \left( \frac{\pi_{\theta}(s_i \mid x)}{\pi_{\theta_{\text{old}}}(s_i \mid x)} A_i, \ \text{clip} \left( \frac{\pi_{\theta}(s_i \mid x)}{\pi_{\theta_{\text{old}}}(s_i \mid x)}, 1 - \epsilon, 1 + \epsilon \right) A_i \right) - \beta \mathbb{D}_{\text{KL}} \left( \pi_{\theta} \parallel \pi_{\theta_{\text{ref}}} \right) \right], (2)
+
+[p. 4 | section: 2.2.2 END-TO-END TRAINING WITH RL | type: Text]
+where A_i = (r_i - \text{mean}(\{r_j\}_{j=1}^G))/\text{std}(\{r_j\}_{j=1}^G) represents the normalized advantage of the i-th rollout within the group, \epsilon is the clipping ratio, and \beta is the coefficient for the KL divergence penalty. The inclusion of the KL divergence term ensures that the updated policy does not deviate significantly from the reference policy.
+
+[p. 4 | section: 2.2.3 REWARD CALCULATION | type: Text]
+Generating End-Task Output. It is important to note that the reward is not computed directly from the compressor's output summary. Instead, the summary s generated by the compressor is prepended to the original input question q, and this combined input is fed into the LLM M:(s,q)\mapsto \hat{y} to produce a predicted answer \hat{y} . The reward is then calculated by comparing \hat{y} to the gold answer y. Throughout the training process, the parameters of M remain fixed and are not updated.
+
+[p. 4 | section: 2.2.3 REWARD CALCULATION | type: Text]
+Computing Rewards. We design simple rule-based rewards based on end-task performance to guide the compressor's improvement, which consists of two components:
+
+[p. 4 | section: 2.2.3 REWARD CALCULATION | type: Text]
+• EM Reward ( r_{\rm EM} ). We employ EM as the main reward function, which is a widely adopted metric for evaluating the accuracy of QA tasks. The EM reward is binary: it yields a value of 1 if the generated answer perfectly matches the ground truth, and 0 otherwise.
+
+[p. 4 | section: 2.2.3 REWARD CALCULATION | type: Equation]
+r_{\rm EM} = \begin{cases} 1 & \text{if } y = \hat{y}, \\ 0 & \text{otherwise.} \end{cases} (3)
+
+[p. 4 | section: 2.2.3 REWARD CALCULATION | type: Text]
+• F1 Reward (r_{\rm F1}) . Since exact matches occur infrequently in practice, relying solely on EM rewards leads to sparse reward signals. Furthermore, the EM metric fails to distinguish between partially correct answers, as all non-exact matches receive zero reward. Therefore, we introduce F1 reward, which provides a finer-grained evaluation by measuring the degree of partial match.
+
+[p. 4 | section: 2.2.3 REWARD CALCULATION | type: Equation]
+r_{\rm FI} = \frac{2 \times I_N}{P_N + R_N},\tag{4}
+
+[p. 4 | section: 2.2.3 REWARD CALCULATION | type: Text]
+where P_N denotes the number of tokens in the predicted answer, R_N denotes the number of tokens in the gold answer, and I_N is the number of intersecting tokens between the two answers.
+
+[p. 5 | section: 2.2.3 REWARD CALCULATION | type: Text]
+The final reward function combines these reward signals through weighted summation:
+
+[p. 5 | section: 2.2.3 REWARD CALCULATION | type: Equation]
+r = r_{\rm EM} + \alpha \cdot r_{\rm Fl},\tag{5}
+
+[p. 5 | section: 2.2.3 REWARD CALCULATION | type: Text]
+where α ∈ (0, 1] is a hyperparameter that controls the relative contribution of the F1 reward.
+
+[p. 5 | section: 2.2.4 TRAINING TEMPLATE | type: Text]
+Figure 5 displays the prompt employed to train the compressor model for generating a summary of the retrieved documents, conditioned on the given question. This prompt is notably concise. For end-task answer generation, the prompt provided to the LLM M is illustrated in Figure 6, which incorporates few-shot in-context examples, the (generated summary of) retrieved documents, and the question.
+
+[p. 5 | section: 2.2.5 EFFICIENCY ANALYSIS | type: Text]
+Training Efficiency. Since our method employs reinforcement learning for training, it incurs greater time and computational costs compared to approaches that do not utilize reinforcement learning (Xu et al., 2024; Cao et al., 2024) . However, our training process only optimizes a lightweight compressor model with relatively few parameters, while the larger generator LLM responsible for producing task answers remains fixed and is not updated during training. This design ensures high training efficiency—for instance, training one epoch takes approximately 2 hours using eight H20 GPUs, and convergence is typically achieved within just two epochs. In contrast, other reinforcement learning-based methods, such as ReSearch (Chen et al., 2025a) and R1-Searcher (Song et al., 2025) , require direct fine-tuning of the large generator LLM, leading to considerably higher training time and resource consumption. Furthermore, it is important to emphasize that our method exhibits strong generalization capability. As shown in Section 3.3, a model trained only once demonstrates broad applicability, thereby reducing the need for frequent retraining and further lowering the overall training cost.
+
+[p. 5 | section: 2.2.5 EFFICIENCY ANALYSIS | type: Text]
+Inference Efficiency. Our method significantly enhances inference efficiency. In contrast to RAG approaches that do not employ a compressor—and thus require the generator LLM to directly encode lengthy documents, often spanning thousands of tokens—our approach introduces a lightweight compressor that processes long documents and summarizes them into compact representations of only a few dozen tokens before feeding them to the generator LLM. Since the compressor is an order of magnitude smaller in parameter size than the generator LLM, it substantially reduces the encoding time that would otherwise be incurred by the generator, leading to notable gains in inference efficiency. It is also important to note that the use of reinforcement learning does not adversely affect inference efficiency, as it is only involved during the training phase.
+
+[p. 5 | section: 3.1 EXPERIMENTAL SETTINGS | type: Text]
+Datasets and Evaluation Metrics. We evaluate our method on four benchmark datasets: two single-hop question-answering datasets, Natural Questions (NQ) (Kwiatkowski et al., 2019) and TriviaQA (Joshi et al., 2017) , as well as two multi-hop question-answering datasets, HotpotQA (Yang et al., 2018) and 2WikiMultihopQA (Ho et al., 2020) . Results are reported on the test sets of Natural Questions and TriviaQA, as well as the development sets of HotpotQA and 2WikiMultihopQA. Following RECOMP (Xu et al., 2024) , the performance is measured using Exact Match and token-level F1 scores, while efficiency is assessed by the number of tokens provided in the context.
+
+[p. 5 | section: 3.1 EXPERIMENTAL SETTINGS | type: Text]
+Compression Model (πθ). We trained our compression model using Qwen2.5-1.5B-Instruct to generate summaries of the retrieved documents. To evaluate the effect of using different models as compressors, we also trained compressors using Llama3.2-1B-Instruct and Llama3.2-3B-Instruct (Section 3.3) .
+
+[p. 5 | section: 3.1 EXPERIMENTAL SETTINGS | type: Text]
+Large Language Model (M). We use Qwen2.5-14B-Instruct as the primary LLM model M to generate predicted answers which are used to guide the training of the compressor. To evaluate the generalization ability of our method, we also transfer to another LLM model, LLama3.1-8B-Instruct (Table 2) .
+
+[p. 6 | section: 3.1 EXPERIMENTAL SETTINGS | type: TableGroup]
+Table 1: Open-domain QA results using Qwen2.5-14B-Instruct as the downstream LLM (M). The reported token counts represent the length of in-context documents, excluding few-shot examples. RECOMP, NoiseFilter-IB and our method CORE are all trained using Qwen2.5-1.5B-Instruct. NQ 7 TriviaQA L F IotpotQ. A 2Wiki Multiho pQA EM F1 # tok EM F1 # tok EM F1 # tok EM F1 # tok No Retrieval 21.36 30.97 0 53.23 59.98 0 21.05 29.48 0 26.11 29.51 0 RAG without compression on Top1 Document 34.46 44.41 142 60.82 68.70 143 29.20 38.93 147 26.79 31.87 153 Top3 Documents 37.78 48.45 427 62.60 71.02 430 31.64 41.87 442 27.89 33.58 460 Top5 Documents 38.03 49.16 712 64.10 72.48 715 32.99 43.69 737 29.64 35.21 766 Top10 Documents 38.67 50.03 1425 64.40 72.92 1428 33.95 44.88 1471 31.04 36.75 1531 Compression of top 5 do cuments ; BM25 25.23 36.47 37 55.36 63.90 39 24.18 35.73 71 25.42 30.29 68 Qwen2.5-1.5B 31.94 43.03 36 57.99 66.70 30 27.36 37.47 33 25.93 31.18 32 DeepSeek-V3 (671B) 37.73 50.39 54 64.13 73.20 50 33.59 44.83 48 27.99 32.67 92 RECOMP-Abs (1.5B) 34.18 46.26 58 60.31 68.50 53 28.96 39.95 56 30.25 36.73 52 RECOMP-Ext (1.5B) 33.84 46.05 56 60.18 68.39 48 29.93 41.09 45 30.78 37.07 51 NoiseFilter-IB (1.5B) 35.15 45.94 48 59.51 68.15 35 27.97 38.62 38 27.85 34.69 40 LongLLMLingua (1.5B) 33.65 43.15 152 58.96 66.82 148 28.03 38.49 149 29.37 33.62 153 QGC (1.5B) 36.23 45.88 49 61.02 68.45 47 29.16 40.05 45 31.14 36.83 51 CORE (1.5B) 41.02 50.40 46 65.63 72.55 32 33.67 45.06 36 36.72 42.05 49 Compression of top 10 a locumen ts (with t the comp ressor ti rained or ı top 5 d locs) BM25 25.91 36.88 38 55.28 63.16 37 23.49 35.01 68 25.61 30.54 65 Qwen2.5-1.5B 32.94 44.84 40 58.45 67.31 33 28.17 38.48 36 26.22 31.57 34 DeepSeek-V3 (671B) 37.79 51.07 56 65.29 74.45 53 34.62 45.69 50 29.00 34.64 40 RECOMP-Abs (1.5B) 34.40 46.93 59 61.42 69.88 52 31.54 42.92 52 31.98 38.16 49 RECOMP-Ext (1.5B) 33.96 46.34 60 61.03 69.51 50 31.92 43.18 55 32.52 38.87 44 NoiseFilter-IB (1.5B) 35.36 46.24 50 59.92 68.32 38 28.21 38.83 38 28.63 35.16 42 LongLLMLingua (1.5B) 33.78 43.37 154 59.17 66.97 150 28.33 38.95 148 29.62 34.11 151 QGC (1.5B) 36.03 45.62 50 61.23 68.74 49 29.12 39.63 46 31.71 37.52 50 CORE (1.5B) 41.88 51.26 52 66.76 73.64 33 34.68 46.35 37 37.99 43.28 48
+
+[p. 6 | section: 3.1 EXPERIMENTAL SETTINGS | type: Text]
+Retrieval Corpus and Retrievers. Following previous studies (Xu et al., 2024), we use the Wikipedia corpus from December 20, 2018, as the retrieval source for all four datasets. The articles are segmented into non-overlapping 100-word documents. To ensure that our method is not dependent on a specific retriever, we experiment with several mainstream retrievers. Specifically, we use DPR (Karpukhin et al., 2020) for NQ, a hybrid of DPR and BM25 (Robertson et al., 1995) for TriviaQA, and the Contriever model (Izacard et al., 2021) trained on the MS MARCO dataset (Nguyen et al., 2016) for HotpotQA and 2WikiMultihopQA.
+
+[p. 6 | section: 3.1 EXPERIMENTAL SETTINGS | type: Text]
+Baselines. To evaluate the effectiveness of our method, we compared it against various baselines. First, we evaluated the uncompressed approach—retaining the original in-context RALM setup—by prepending the top 1, 3, 5, and 10 retrieved documents to the prompt. We also tested alternative compression methods, including the traditional BM25 algorithm (which ranks sentences by their similarity to the input), off-the-shelf Qwen2.5-1.5B-Instruct model (with comparable parameter size to our approach), and DeepSeek-V3 model (671B parameters, far exceeding our compressor's capacity). Furthermore, we included state-of-the-art context compression methods for RAG, RECOMP (Xu et al., 2024), NoiseFilter-IB (Zhu et al., 2024), LongLLMLingua (Jiang et al., 2024) and QGC (Cao et al., 2024). For RECOMP, we evaluated both its abstractive and extractive variants. To ensure a fair comparison, all trainable approaches were all trained using the same model.
+
+[p. 6 | section: 3.1 EXPERIMENTAL SETTINGS | type: Text]
+Implementation Details. For the distillation warm-up phase, we perform full-parameter supervised fine-tuning on the off-the-shell language model for two epochs using LLaMA-Factory ^1 . This warmed-up model then serves as the initializer for the subsequent reinforcement learning phase. We adopt the Verl framework ^2 for RL training. The initialized compression model is trained for two epochs on each dataset. Training is conducted on eight NVIDIA H20 GPUs using full parameter GRPO optimization, with a learning rate of 1e-5, a batch size of 256, five rollouts per sample, and a KL loss coefficient of 0.001. The downstream LLM (M) used for reward generation is served using the vLLM inference engine during RL training.
+
+[p. 6 | section: 3.1 EXPERIMENTAL SETTINGS | type: Footnote]
+& lt;sup>1</sup>
+
+[p. 6 | section: 3.1 EXPERIMENTAL SETTINGS | type: Footnote]
+& lt;sup>2</sup>
+
+[p. 7 | section: 3.2 OVERALL PERFORMANCE | type: Text]
+The detailed comparison results are presented in Table 1. We evaluate the following approaches: traditional RAG without compression using the top 1, 3, 5, and 10 documents prepended to the original input; BM25-based compression; off-the-shelf LLMs (Qwen2.5-1.5B-Instruct and DeepSeek-V3); and state-of-the-art trainable RAG compressors (RECOMP-Abstractive, RECOMP-Extractive, and NoiseFilter-IB). For fair comparison, all trainable methods were trained using the same backbone model, Qwen2.5-1.5B-Instruct. These compressors were trained on five-document inputs, and we report their performance both on in-domain five-document compression and out-of-domain generalization to ten-document compression.
+
+[p. 7 | section: 3.2 OVERALL PERFORMANCE | type: Text]
+RAG versus No RAG. As shown in Table 1, prepending the original input with retrieved documents yields a substantial improvement over the no-retrieval baseline, indicating that these documents provide valuable information for generating the answer. Performance improves as the number of retrieved documents increases from 1 to 10, albeit with diminishing returns—a trend attributable to the decreasing relevance of lower-ranked documents. However, this performance gain comes at the cost of a significant increase in the number of additional tokens the LLM must encode, rising from 0 to over 1,400 tokens.
+
+[p. 7 | section: 3.2 OVERALL PERFORMANCE | type: Text]
+Lossless Compression of CORE. Our compressor was trained using a context of five documents. As presented in Table 1, compared to prepending the full content of all five documents to the original input (i.e., the uncompressed baseline), our method achieves a high compression rate of approximately 6% with no loss in performance . Remarkably, on all four datasets, the approach not only maintains performance but also enhances EM by 1 to 7 points. For example, on NQ, the model achieves an EM of 41.02 with compressed input, surpassing the uncompressed score of 38.03.
+
+[p. 7 | section: 3.2 OVERALL PERFORMANCE | type: Text]
+Comparison with Compression Baselines. When compressing the top five documents, all baseline compression methods result in performance degradation to varying degrees compared to the uncompressed baseline. Specifically, BM25 leads to a substantial performance drop. Using the offthe-shelf Qwen2.5-1.5B-Instruct model for compression yields better results than BM25 but still underperforms relative to the uncompressed baseline. Surprisingly, even the large-scale DeepSeek-V3 model (with 671B parameters) achieves only near-lossless compression on TriviaQA and HotpotQA, while performance on NQ and 2Wiki remains below the uncompressed baseline. As for the trained compressors (RECOMP, NoiseFilter-IB, LongLLMLingua, and QGC), all of them exhibit performance degradation compared to no compression. The decline generally ranges from 2 to 6 EM poin ts across nearly all datasets. In contrast, our method, CORE, delivers the best performance. It not only surpasses compression methods of comparable size by 4-5 EM points but also maintains an advantage over the hundreds-of-times-larger DeepSeek model. This clearly demonstrates the benefit and importance of end-to-end optimization.
+
+[p. 7 | section: 3.2 OVERALL PERFORMANCE | type: Text]
+Generalization to 10-Doc Compression. When generalizing the trained compressors to handle the top-10 documents without retraining, the aforementioned conclusions remain valid. CORE continues to achieve lossless compression and performs best among all compression methods. On NQ, it achieves a token compression ratio of 3.6% while improving the EM by 3.2 points compared to using all ten documents. Similar trends are observed on TriviaQA, with a compression ratio of 2.3% and a gain of 2.4 EM points relative to the full-document baseline. On HotpotQA, CORE achieves a compression ratio of 2.5% and an improvement of 0.7 EM points. For 2WikiMultihopQA, it obtains a compression ratio of 3.1% along with a notable increase of 6.9 EM points.
+
+[p. 7 | section: 3.3 ROBUSTNESS AND GENERALIZATION ABILITY ANALYSIS | type: Text]
+Robustness of CORE Across Compressor Architectures. To evaluate whether the effectiveness of CORE depends on the choice of compressor model, we compared the performance of compressors trained using different model architectures—while keeping the downstream LLM (Qwen2.5-14B-Instruct) fixed. The compressors tested include LLaMA3.2-1B-Instruct, Qwen2.5-1.5B-Instruct, and LLaMA3.2-3B-Instruct, which vary in architecture and parameter count. As shown in Figure 3, the results indicate that: (1) These trained compressors consistently achieve lossless compression and outperform the uncompressed baseline (represented by the red reference line in the figure, which corresponds to prepending the full document content), confirming that our training framework is robust and not tied to a specific compressor architecture. (2) Compression performance improves as the size of the compressor model increases, consistent with scaling laws. More detailed results can be found in Tables 7 and 8.
+
+[p. 8 | section: 3.3 ROBUSTNESS AND GENERALIZATION ABILITY ANALYSIS | type: FigureGroup]
+Figure 3: The impact of different models used to train the compressor.
+
+[p. 8 | section: 3.3 ROBUSTNESS AND GENERALIZATION ABILITY ANALYSIS | type: TableGroup]
+Table 2: Zero-shot transfer of the trained compressor to Llama-3.1-8B-Instruct. NQ 7 TriviaQA HotpotQA 2WikiMultihopQA EM F1 # tok EM F1 # tok EM F1 # tok EM F1 # tok No Retrieval 24.04 34.91 0 55.64 62.57 0 19.93 27.75 0 27.64 31.18 0 RAG without compres ssion Top1 Document 33.80 44.06 142 59.17 67.50 143 27.95 37.49 147 28.41 33.43 153 Top3 Documents 36.87 47.81 427 61.13 70.06 430 30.17 40.71 442 28.67 34.23 460 Top5 Documents 37.65 48.87 712 62.26 71.04 715 31.44 42.16 737 29.43 35.18 766 Top10 Documents 38.12 49.93 1425 63.95 72.71 1428 32.19 42.62 1471 30.45 36.04 1531 Compression of top 5 docume nts Qwen2.5-1.5B 32.60 44.21 36 56.76 65.77 30 26.86 36.90 33 25.45 30.88 32 DeepSeek-V3 (671B) 37.56 50.11 54 62.52 72.34 50 33.05 44.25 48 28.64 33.87 92 RECOMP-Abs (1.5B) 33.41 45.50 58 58.50 67.37 53 28.85 39.76 56 31.63 37.81 52 RECOMP-Ext (1.5B) 33.12 45.06 60 57.98 66.84 55 29.03 40.04 52 31.85 38.02 55 CORE (1.5B) 40.72 50.00 46 64.08 71.13 32 32.17 43.71 36 35.99 41.42 49 Compression of top 1 0 docum ents Qwen2.5-1.5B 32.88 44.66 40 57.44 66.56 33 27.31 37.31 36 25.80 31.30 34 DeepSeek-V3 (671B) 37.49 51.28 56 63.79 73.80 53 34.24 45.35 50 31.45 37.09 40 RECOMP-Abs (1.5B) 34.18 46.80 59 59.69 68.89 52 30.17 41.42 55 33.61 39.78 44 RECOMP-Ext (1.5B) 34.06 46.55 60 59.33 68.71 50 30.52 41.98 55 33.52 39.42 44 CORE (1.5B) 41.77 51.27 52 65.25 72.45 33 33.25 45.09 37 37.59 42.87 48
+
+[p. 8 | section: 3.3 ROBUSTNESS AND GENERALIZATION ABILITY ANALYSIS | type: Text]
+Transferability of CORE Across Downstream LLMs. We evaluate the transferability of our trained compressor and other baseline compressors to a new downstream LLM, the LLaMA-3.1-8B model. The results are presented in Table 2. Note that all trainable compressors, including our own, were trained using feedback generated by Qwen2.5-14B-Instruct. The findings reveal that existing trained compressors (e.g., RECOMP) exhibit limited generalization capability, as indicated by a larger performance gap relative to the baseline of prepending full documents. In contrast, CORE demonstrates stronger generalization, achieving lossless compression on the new downstream LLM. Specifically, it not only maintains performance but also surpasses the full-document baseline (i.e., no compression) across all four datasets while retaining a high compression rate. These results suggest that the summaries produced by our method are inherently high-quality and preserve key information necessary for accurate answering, thereby enabling effective transfer to other LLMs.
+
+[p. 8 | section: 3.4 ABLATION STUDY | type: Text]
+Table 3 presents an ablation study on the two stages of our method: distillation and GRPO. Here, "w/o distillation" denotes training the compressor with GRPO directly, bypassing the warm-start phase, while "w/o GRPO" indicates using only the distillation step without subsequent reinforcement learning. The results demonstrate that removing either component leads to performance degradation, confirming the necessity of both. In addition, the decline is more substantial when GRPO is omitted, highlighting the crucial role of reinforcement learning in the absence of explicit supervision. Distillation injects external knowledge into the model, providing a favorable starting point for RL training and thereby enabling RL to more effectively unlock the compressor's full potential.
+
+[p. 9 | section: 3.4 ABLATION STUDY | type: TableGroup]
+Table 3: Ablation study. Dataset Metric w/o distillation w/o GRPO CORE EM 36.37 34.18 41.02 NQ F1 46.91 46.26 50.40 EM 65.23 60.31 65.63 TQA F1 72.41 68.50 72.55 EM 32.01 28.96 33.67 HotpotQA F1 42.73 39.95 45.06 EM 31.40 30.25 36.72 2Wiki F1 36.89 36.73 42.05
+
+[p. 9 | section: 3.4 ABLATION STUDY | type: FigureGroup]
+Figure 4: The impact of α.
+
+[p. 9 | section: 3.5 THE IMPACT OF α. | type: Text]
+Figure 4 illustrates the performance of our method under different values of α, which controls the weighting coefficient of the F1 reward term. Setting α = 0 corresponds to using only the EM reward. Results indicate that performance improves when α > 0 across all datasets, demonstrating the effectiveness of the F1 reward in mitigating the sparsity issue associated with the EM reward. However, the optimal value of α is dataset-dependent; values between 0.2 and 0.5 generally yield strong performance.
+
+[p. 9 | section: 3.6 CASE STUDY | type: Text]
+To conduct an in-depth analysis of the advantages of our compressor, we performed case studies on one single-hop QA dataset (NQ) and one multi-hop QA dataset (2Wiki), with the results presented in Table 4 and Table 5, respectively. For each case, we compared the summaries generated by off-the-shelf Qwen2.5-1.5B-Instruct, RECOMP, and our method CORE based on the same set of documents, as well as the predicted answers generated by the LLM after prepending these summaries. As shown in the tables, although the summaries produced by off-the-shelf Qwen2.5-1.5B are concise, they largely fail to capture key information relevant to answering the question. In contrast, RECOMP demonstrates better summarization capability but is prone to being overwhelmed by lengthy documents, resulting in misjudgments and even generating misleading information—such as the statement in Table 4: " The U.S. stopped drafting for the Vietnam War after the Selective Service System was officially abolished in December 1972 "—which leads the downstream LLM to produce the incorrect answer "1972". Our method, CORE, accurately extracts answer-critical information from lengthy documents, exemplified by the summary: " The U.S. stopped drafting for the Vietnam War in 1973 after announcing the decision by Secretary of Defense Melvin Laird earlier that year ", thereby enabling the LLM to generate the correct answer "1973". This indicates that our compressor, trained with an end-task target-oriented reinforcement learning optimization strategy, can produce document summaries that are most helpful for answering the given question while effectively filtering out irrelevant information.
+
+[p. 9 | section: 4 RELATED WORK | type: Text]
+Context Compression in RAG. RAG enhances the performance of LLMs on knowledge-intensive tasks by retrieving the most relevant documents from extensive knowledge bases and prepending them as contextual information to the original input (Ram et al., 2023; Fan et al., 2024; Lin et al., 2023; Shi et al., 2023) . However, this approach requires the LLM to process significantly longer token sequences, resulting in increased computational costs. To mitigate this issue, researchers have begun to explore methods for compressing retrieved documents in RAG systems (Xu et al., 2024; Cao et al., 2024; Jin et al., 2024b; Zhu et al., 2024; Kim & Thorne, 2025; Rau et al., 2024; Wu et al., 2025; Louis et al., 2025; Jin et al., 2024a; Li et al., 2024a; b; Zhang et al., 2024a) . For instance, Xu et al. (2024) propose compressing retrieved documents into textual summaries before in-context augmentation, training the compressor through data selection and distillation. Similarly, Cao et al. (2024) introduce a Query-Guided Compressor (QGC) that uses queries to guide the compression process, effectively preserving essential information. Jin et al. (2024b) refine retrieved documents into Key Supporting Evidence (KSE) through a combination of knowledge synthesis, supervised fine-tuning (SFT), and preference alignment. Meanwhile, Zhu et al. (2024) present an informationtheoretic approach called NoiseFilter-IB, which filters noise by maximizing the mutual information
+
+[p. 10 | section: 4 RELATED WORK | type: Text]
+between the compressed content and the ground-truth output. Additionally, Kim & Thorne (2025) train a compressor to extract critical information using reward functions based on predefined heuristic rules. However, most of these methods are heuristic in nature, and due to the lack of ideal compression labels, the compressed content they produce cannot be guaranteed to benefit downstream LLMs. In contrast, our method, CORE, adopts an end-to-end optimization approach to address these limitations.
+
+[p. 10 | section: 4 RELATED WORK | type: Text]
+Reinforcement Learning. Reinforcement learning (RL) has recently achieved notable success, enabling LLMs to develop reasoning capabilities without explicit step-by-step supervision (Liu et al., 2024; Shao et al., 2024; Guo et al., 2025) . Building on these advances, several studies have applied RL to improve RAG (Ke et al., 2024) . For example, Kulkarni et al. (2024) use RL to autonomously decide whether to retrieve documents, while Zhang et al. (2024b) employ RL to optimize the ranking of retrieved documents. Similarly, Mao et al. (2024) propose a framework for training query rewriting models for RAG without relying on human annotations. MMOA-RAG (Chen et al., 2025b) enhances RAG through multi-agent reinforcement learning, incorporating a query rewriter, retriever, and generator. Meanwhile, RL has been applied to address other challenges in RAG, such as enhancing the quality of retrieved content. For instance, Oreo (Li & Ramakrishnan, 2025) trains a reconstructor with PPO and ROUGE-based rewards to rewrite passages for improved performance. In contrast, our work CORE tackles the distinct problem of computational efficiency. Our approach leverages the GRPO algorithm, direct task-performance rewards, and a lightweight compressor to enhance efficiency without compromising accuracy. Moreover, a line of research has utilized RL to integrate search with reasoning in a step-by-step manner (Singh et al., 2025) . For instance, Chen et al. (2025a) introduce a framework called ReSearch, which trains LLMs to reason with search using RL, without requiring supervised data for reasoning steps. Related approaches include R1- Searcher (Song et al., 2025) , WebThinker (Li et al., 2025) , and DeepResearcher (Zheng et al., 2025) . Although these methods are end-to-end, they differ fundamentally from our problem setting. These approaches typically involve directly training the LLM generator—which tends to be a large-scale model with a high parameter count. However, such training becomes infeasible when the model is a black box (e.g., GPT-4), as internal weights or gradients are inaccessible. Furthermore, these methods introduce extensive internal thinking processes that substantially increase context length and inference time. In contrast, our approach treats the generator LLM as a fixed black-box model and trains only a lightweight plug-in compressor to produce document summaries. This design significantly improves both training and inference efficiency. A parallel line of work applies RL to prompt compression. For example, PCRL (Jung & Kim, 2024) and TACO-RL (Shandilya et al., 2025) learn to compress prompts via token-level keep-or-drop decisions, using the similarity between model outputs with compressed and original prompts as the reward. Our work, CORE, introduces key distinctions in both objective and methodology. First, we target the more complex problem of compressing multiple retrieved documents in RAG, rather than single prompts. Second, we employ a generative compressor that can rephrase and synthesize content, instead of making token-level binary actions. Most importantly, we optimize compression using a direct task-performance reward with GRPO, which enables us to achieve true lossless compression at significantly higher ratios—a stark contrast to the performance degradation observed in prior prompt compression methods.
+
+[p. 10 | section: 5 CONCLUSION | type: Text]
+This paper analyzes the limitations of current context compression methods for RAG. A primary challenge is the lack of optimal reference summaries for supervised learning, which often results in performance degradation in downstream tasks. To overcome this, we frame context compression as a reinforcement learning problem, utilizing downstream task performance as a reward signal to train the compression policy, thereby enabling end-to-end optimization. Extensive experiments demonstrate that our proposed method, CORE, achieves effectively lossless compression by maintaining a high compression ratio while preserving original task performance. Surprisingly, CORE not only preserves but actually enhances performance on all downstream tasks. Further in-depth analysis provides additional insights into the efficacy of our approach.

iclr26/6FyYLNnDL6/paper.md

@@ -0,0 +1,627 @@
+
+
+{0}------------------------------------------------
+
+# CORE: PERFORMANCE-LOSSLESS CONTEXT COM-PRESSION FOR RETRIEVAL-AUGMENTED GENERA-TION
+
+Anonymous authors Paper under double-blind review
+
+<span id="page-0-0"></span>**000**
+
+**003 004**
+
+**006**
+
+**024**
+
+**028 029 030**
+
+**052 053**
+
+## ABSTRACT
+
+Retrieval-Augmented Generation (RAG) has emerged as a promising approach to enhance the timeliness of knowledge updates and the factual accuracy of responses in large language models. However, incorporating a large number of retrieved documents significantly increases input length, leading to higher computational costs. Existing approaches to document compression tailored for RAG often degrade task performance, as they typically rely on predefined heuristics in the absence of clear compression guidelines. These heuristics fail to ensure that the compressed content effectively supports downstream tasks. To address these limitations, we propose CORE, a novel method for lossless context compression in RAG. CORE is optimized end-to-end and does not depend on predefined compression labels, which are often impractical to obtain. Instead, it leverages downstream task performance as a feedback signal, iteratively refining the compression policy to enhance task effectiveness. Extensive experiments across four datasets demonstrate the effectiveness of CORE. With a high compression ratio of 3%, CORE not only prevents performance degradation compared to including full documents (i.e., without compression) but also improves the average Exact Match (EM) score by 3.3 points. The code for CORE is available at <https://anonymous.4open.science/r/CORE-28B4>.
+
+## 1 INTRODUCTION
+
+Large language models (LLMs) have undergone rapid development in recent years, significantly enhancing performance across various language tasks due to their emergent capabilities in semantic understanding and reasoning. Nevertheless, LLMs still face challenges in updating knowledge and providing factual responses [\(Fan et al., 2024\)](#page-10-0). To address these issues, Retrieval-Augmented Generation (RAG) has emerged as a promising approach. By retrieving the most relevant documents from external knowledge bases and prepending them as contextual information to the original input, RAG substantially improves LLM performance on knowledge-intensive tasks [\(Ram et al., 2023\)](#page-11-0).
+
+While RAG enhances performance, its effectiveness is closely tied to the number of retrieved documents used, since a broader context increases the probability of encompassing critical evidence. As illustrated in Figure [1,](#page-1-0) performance was weakest without any retrieved documents (i.e., without RAG). Accuracy improved consistently as more documents were added to the context, ultimately exceeding the no-RAG baseline by over 10 Exact Match (EM) points. However, this performance gain came with two significant limitations: (1) a substantial increase in computational cost from processing a larger number of context tokens [\(Xu et al., 2024\)](#page-12-0), and (2) the model's difficulty in effectively leveraging all provided documents, often resulting in the omission of key information located in the middle of the context [\(Liu et al., 2023\)](#page-11-1).
+
+These limitations have motivated recent research efforts aimed at compressing the retrieved context [\(Jin et al., 2024b;](#page-10-1) [Wu et al., 2025;](#page-12-1) [Jin et al., 2024a;](#page-10-2) [Zhang et al., 2024a\)](#page-12-2). Prominent approaches include document summarization [\(Xu et al., 2024\)](#page-12-0), key information extraction [\(Cao et al., 2024;](#page-10-3) [Xu](#page-12-0) [et al., 2024\)](#page-12-0), the construction of key supporting evidence [\(Jin et al., 2024b\)](#page-10-1), and noise filtering based on information theory [\(Zhu et al., 2024\)](#page-12-3). Despite recent progress, these methods have several notable shortcomings. First, compression often results in a performance trade-off. For instance, RECOMP
+
+{1}------------------------------------------------
+
+**059**
+
+**061**
+
+**072 073 074**
+
+**079**
+
+**094**
+
+<span id="page-1-0"></span>![](_page_1_Figure_1.jpeg)
+
+![](_page_1_Figure_2.jpeg)
+
+Figure 1: Performance evolution with an increasing number of retrieved documents on two datasets. Traditional RAG requires more documents for better performance, while our method achieves comparable or superior results with significant token compression.
+
+[\(Xu et al., 2024\)](#page-12-0) suffers a 3–5 point drop in EM score (Table [1\)](#page-5-0), making it unsuitable for accuracysensitive applications. Second, most compression methods are heuristic in nature. The models are typically trained to generate summaries that are generally good but not necessarily useful for the downstream answer-generation LLM. This limitation arises from the lack of an ideal supervisory signal that defines what an optimal summary should be for the downstream task. This fundamental gap hinders the end-to-end optimization of existing methods. Finally, some compression models [\(Zhu et al., 2024\)](#page-12-3) have parameter counts comparable to the LLM that performs the end task, resulting in substantial computational costs that undermine the efficiency gains of compression.
+
+Addressing these critical shortcomings requires a method that aligns compression with downstream task requirements, thereby minimizing performance trade-offs. To bridge this gap, we propose CORE, a novel method designed to achieve lossless context compression for RAG. Unlike previous compression methods, CORE is optimized in an end-to-end manner. Since obtaining predefined summary labels for supervision is impractical, we instead use downstream task performance as a feedback signal to evaluate the compression model's output. This feedback enables iterative refinement of the compression policy, guiding the compression model toward improved downstream performance. To this end, we employ Group Relative Policy Optimization (GRPO), a technique particularly well-suited for this purpose [\(Liu et al., 2024;](#page-11-2) [Shao et al., 2024;](#page-12-4) [Chen et al., 2025a\)](#page-10-4). In our framework, the accuracy of the downstream QA task is formalized as a reward, with the compression policy optimized through group-wise relative comparisons. Furthermore, our compression model is substantially smaller than the downstream LLM, which significantly reduces the computational overhead associated with encoding retrieved documents.
+
+We evaluate CORE on four benchmark datasets: two single-hop QA datasets (*Natural Questions* and *TriviaQA*) and two multi-hop datasets (*HotpotQA* and *2WikiMultihopQA*). As shown in Table [1,](#page-5-0) CORE achieves state-of-the-art performance across all baselines. With a compression ratio of 3%, our approach not only avoids performance degradation compared to prepending full documents but also improves the average EM score by 3.3 points. We further demonstrate two key advantages of our approach. First, the effectiveness of CORE is not tied to a specific model architecture, as demonstrated by the fact that lossless compression can be achieved when various models are trained as compressors (Figure [3\)](#page-7-0). Second, the compressor exhibits strong transferability: a compression model trained using feedback from one LLM generalizes effectively to other LLMs (Table [2\)](#page-7-1). Finally, an in-depth case study (Tables [4](#page-14-0) and [5\)](#page-15-0) provides a qualitative analysis of the benefits of CORE.
+
+## 2 CORE-RAG
+
+This section introduces our proposed method, COmpression via REinforcement learning (CORE), which is shown in Figure [2.](#page-2-0) First, we provide an overview of the entire workflow. Then, we detail the end-to-end training strategy for the compression model, which is designed to drastically reduce the number of document tokens while preserving task performance.
+
+{2}------------------------------------------------
+
+<span id="page-2-0"></span>![](_page_2_Figure_1.jpeg)
+
+Figure 2: Overview of our method CORE. The upper section illustrates the inference pipeline. The lower section depicts the end-to-end training method for the compression model.
+
+#### 2.1 Problem Formulation
+
+We adopt the same problem formulation as prior work (Xu et al., 2024). Given an input question q, a target output y, and a set of k retrieved documents D, our objective is to compress D with respect to q into a summary s that preserves the most useful information while using significantly fewer tokens than D. This summary s is then prepended to the original input q and fed into an LLM to generate the final answer for the downstream task. This overall pipeline is illustrated in the upper half of Figure 2. The process involves two key components: a compression model  $\pi_{\theta} \colon (q, D) \mapsto s$  and a large language model  $M \colon (s, q) \mapsto \hat{y}$ , which generates the predicted answer  $\hat{y}$ . We treat M as a black-box system and focus exclusively on training the compressor  $\pi_{\theta}$ . The compressor itself is also a language model, but it is intentionally designed to be significantly smaller than M to reduce the computational cost of encoding the retrieved documents.
+
+#### 2.2 Training the Compressor
+
+Our compressor is designed to generate document summaries that are highly useful to the LLM (M) for downstream tasks. This objective is challenging because the criteria for an effective summary are task-dependent, and direct supervision is unavailable. We therefore formulate this as an end-to-end training problem and employ reinforcement learning to optimize the compressor without relying on pre-defined compression labels. The overall architecture of our training framework is illustrated in the lower portion of Figure 2. The following sections describe the key components of our approach: distillation warm-up, policy optimization, and reward calculation.
+
+### 2.2.1 DISTILLATION FOR WARM-START
+
+Due to the limited parameter size of our compression model, its capability for question-focused document summarization is constrained. To provide a strong initial policy for RL and ensure training stability, we employ knowledge distillation from a teacher model to initialize our compressor. Specifically, we first utilize a large-scale language model (DeepSeek-V3) as the teacher to generate summaries of retrieved documents related to the given question. We then evaluate the performance of the downstream LLM (M) on the QA task under two conditions: (1) with the teacher-generated summary  $\hat{s}$  prepended to the input question q, and (2) with the original question alone. The corresponding performance scores are denoted as  $p_{\text{summary}}$  and  $p_{\text{original}}$ , respectively. By comparing these results, we retain instances where  $p_{\text{summary}} > p_{\text{original}}$ , indicating that the summary enhances RAG performance. We also retain cases where  $p_{\text{original}} = 1$  (i.e., the model produces a fully correct answer without the summary) and  $p_{\text{summary}} < p_{\text{original}}$ ; for these, we set the target summary  $\hat{s}$  to an empty string. All other instances are discarded. The resulting filtered and modified dataset is denoted as
+
+{3}------------------------------------------------
+
+ $\mathcal{X}_f$ , which is used for supervised fine-tuning of the compression model. The fine-tuning objective is defined as:
+
+ $\mathcal{L}_{\text{distill}} = \frac{1}{|\mathcal{X}_f|} \sum_{(q,D,\hat{s}) \in \mathcal{X}_f} \mathcal{L}_{\text{CE}}(\pi_{\theta}(q,D),\hat{s}), \tag{1}$ 
+
+where  $\pi_{\theta}(q, D)$  denotes the output of the compression model and  $\mathcal{L}_{CE}$  is the cross-entropy loss. This distillation process yields a robust initialization for RL and promotes stability in subsequent training.
+
+#### 2.2.2 END-TO-END TRAINING WITH RL
+
+Following the distillation phase, the compressor possesses a preliminary compression capability. However, as summaries from even the largest teacher models are not guaranteed to be optimal for the downstream task, further end-to-end optimization is necessary. We therefore formulate this optimization as a reinforcement learning problem. In this framework, the compressor functions as a policy that generates a summary from an input question and its corresponding documents. A reward function, which directly reflects performance on the downstream task (e.g., question-answering accuracy), then evaluates the summary's quality. The objective is to optimize the compressor's parameters to maximize the expected cumulative reward, thereby directly aligning its outputs with the downstream task's objectives.
+
+Specifically, we employ Group Relative Policy Optimization (GRPO) (Shao et al., 2024) algorithm. Unlike Proximal Policy Optimization (PPO), which trains a separate critic model, GRPO estimates the baseline directly from a group of rollouts. Given an existing policy,  $\pi_{\theta_{\text{old}}}$ , and a reference policy,  $\pi_{\theta_{\text{ref}}}$ , the GRPO objective maximizes the compressor policy  $\pi_{\theta}$  using G rollouts  $\tau = \{s_i\}_{i=1}^G \sim \pi_{\theta_{\text{old}}}(\cdot|x)$ , for each input  $x \sim \mathcal{D}$ :
+
+$$\mathcal{J}(\theta) = \mathbb{E}_{x \sim \mathcal{D}, \ \{s_i\}_{i=1}^G \sim \pi_{\theta_{\text{old}}}(\cdot \mid x)}$$
+
+$$\frac{1}{G} \sum_{i=1}^G \left[ \min \left( \frac{\pi_{\theta}(s_i \mid x)}{\pi_{\theta_{\text{old}}}(s_i \mid x)} A_i, \ \text{clip} \left( \frac{\pi_{\theta}(s_i \mid x)}{\pi_{\theta_{\text{old}}}(s_i \mid x)}, 1 - \epsilon, 1 + \epsilon \right) A_i \right) - \beta \mathbb{D}_{\text{KL}} \left( \pi_{\theta} \parallel \pi_{\theta_{\text{ref}}} \right) \right],$$
+
+$$(2)$$
+
+where  $A_i = (r_i - \text{mean}(\{r_j\}_{j=1}^G))/\text{std}(\{r_j\}_{j=1}^G)$  represents the normalized advantage of the i-th rollout within the group,  $\epsilon$  is the clipping ratio, and  $\beta$  is the coefficient for the KL divergence penalty. The inclusion of the KL divergence term ensures that the updated policy does not deviate significantly from the reference policy.
+
+#### 2.2.3 REWARD CALCULATION
+
+**Generating End-Task Output.** It is important to note that the reward is not computed directly from the compressor's output summary. Instead, the summary s generated by the compressor is prepended to the original input question q, and this combined input is fed into the LLM  $M:(s,q)\mapsto \hat{y}$  to produce a predicted answer  $\hat{y}$ . The reward is then calculated by comparing  $\hat{y}$  to the gold answer y. Throughout the training process, the parameters of M remain fixed and are not updated.
+
+**Computing Rewards.** We design simple rule-based rewards based on end-task performance to guide the compressor's improvement, which consists of two components:
+
+• EM Reward ( $r_{\rm EM}$ ). We employ EM as the main reward function, which is a widely adopted metric for evaluating the accuracy of QA tasks. The EM reward is binary: it yields a value of 1 if the generated answer perfectly matches the ground truth, and 0 otherwise.
+
+$$r_{\rm EM} = \begin{cases} 1 & \text{if } y = \hat{y}, \\ 0 & \text{otherwise.} \end{cases}$$
+ (3)
+
+• F1 Reward  $(r_{\rm F1})$ . Since exact matches occur infrequently in practice, relying solely on EM rewards leads to sparse reward signals. Furthermore, the EM metric fails to distinguish between partially correct answers, as all non-exact matches receive zero reward. Therefore, we introduce F1 reward, which provides a finer-grained evaluation by measuring the degree of partial match.
+
+$$r_{\rm FI} = \frac{2 \times I_N}{P_N + R_N},\tag{4}$$
+
+where  $P_N$  denotes the number of tokens in the predicted answer,  $R_N$  denotes the number of tokens in the gold answer, and  $I_N$  is the number of intersecting tokens between the two answers.
+
+{4}------------------------------------------------
+
+The final reward function combines these reward signals through weighted summation:
+
+$$r = r_{\rm EM} + \alpha \cdot r_{\rm Fl},\tag{5}$$
+
+where α ∈ (0, 1] is a hyperparameter that controls the relative contribution of the F1 reward.
+
+### 2.2.4 TRAINING TEMPLATE
+
+**221**
+
+**224**
+
+**256**
+
+Figure [5](#page-13-0) displays the prompt employed to train the compressor model for generating a summary of the retrieved documents, conditioned on the given question. This prompt is notably concise. For end-task answer generation, the prompt provided to the LLM M is illustrated in Figure [6,](#page-13-1) which incorporates few-shot in-context examples, the (generated summary of) retrieved documents, and the question.
+
+### 2.2.5 EFFICIENCY ANALYSIS
+
+Training Efficiency. Since our method employs reinforcement learning for training, it incurs greater time and computational costs compared to approaches that do not utilize reinforcement learning [\(Xu et al., 2024;](#page-12-0) [Cao et al., 2024\)](#page-10-3). However, our training process only optimizes a lightweight compressor model with relatively few parameters, while the larger generator LLM responsible for producing task answers remains fixed and is not updated during training. This design ensures high training efficiency—for instance, training one epoch takes approximately 2 hours using eight H20 GPUs, and convergence is typically achieved within just two epochs. In contrast, other reinforcement learning-based methods, such as ReSearch [\(Chen et al., 2025a\)](#page-10-4) and R1-Searcher [\(Song et al.,](#page-12-5) [2025\)](#page-12-5), require direct fine-tuning of the large generator LLM, leading to considerably higher training time and resource consumption. Furthermore, it is important to emphasize that our method exhibits strong generalization capability. As shown in Section 3.3, a model trained only once demonstrates broad applicability, thereby reducing the need for frequent retraining and further lowering the overall training cost.
+
+Inference Efficiency. Our method significantly enhances inference efficiency. In contrast to RAG approaches that do not employ a compressor—and thus require the generator LLM to directly encode lengthy documents, often spanning thousands of tokens—our approach introduces a lightweight compressor that processes long documents and summarizes them into compact representations of only a few dozen tokens before feeding them to the generator LLM. Since the compressor is an order of magnitude smaller in parameter size than the generator LLM, it substantially reduces the encoding time that would otherwise be incurred by the generator, leading to notable gains in inference efficiency. It is also important to note that the use of reinforcement learning does not adversely affect inference efficiency, as it is only involved during the training phase.
+
+## 3 EXPERIMENTS
+
+### 3.1 EXPERIMENTAL SETTINGS
+
+Datasets and Evaluation Metrics. We evaluate our method on four benchmark datasets: two single-hop question-answering datasets, Natural Questions (NQ) [\(Kwiatkowski et al., 2019\)](#page-11-3) and TriviaQA [\(Joshi et al., 2017\)](#page-10-5), as well as two multi-hop question-answering datasets, HotpotQA [\(Yang et al., 2018\)](#page-12-6) and 2WikiMultihopQA [\(Ho et al., 2020\)](#page-10-6). Results are reported on the test sets of Natural Questions and TriviaQA, as well as the development sets of HotpotQA and 2WikiMultihopQA. Following RECOMP [\(Xu et al., 2024\)](#page-12-0), the performance is measured using Exact Match and token-level F1 scores, while efficiency is assessed by the number of tokens provided in the context.
+
+Compression Model (πθ). We trained our compression model using Qwen2.5-1.5B-Instruct to generate summaries of the retrieved documents. To evaluate the effect of using different models as compressors, we also trained compressors using Llama3.2-1B-Instruct and Llama3.2-3B-Instruct (Section [3.3\)](#page-7-0).
+
+Large Language Model (M). We use Qwen2.5-14B-Instruct as the primary LLM model M to generate predicted answers which are used to guide the training of the compressor. To evaluate the generalization ability of our method, we also transfer to another LLM model, LLama3.1-8B-Instruct (Table [2\)](#page-7-1).
+
+{5}------------------------------------------------
+
+<span id="page-5-0"></span>Table 1: Open-domain QA results using Qwen2.5-14B-Instruct as the downstream LLM (M). The reported token counts represent the length of in-context documents, excluding few-shot examples. RECOMP, NoiseFilter-IB and our method CORE are all trained using Qwen2.5-1.5B-Instruct.
+
+|                         |         | NQ         |          | 7         | TriviaQA  | <b>L</b>  | F     | IotpotQ. | A     | 2Wiki | Multiho | pQA   |
+|-------------------------|---------|------------|----------|-----------|-----------|-----------|-------|----------|-------|-------|---------|-------|
+|                         | EM      | F1         | # tok    | EM        | F1        | # tok     | EM    | F1       | # tok | EM    | F1      | # tok |
+| No Retrieval            | 21.36   | 30.97      | 0        | 53.23     | 59.98     | 0         | 21.05 | 29.48    | 0     | 26.11 | 29.51   | 0     |
+| RAG without compression | on      |            |          |           |           |           |       |          |       |       |         |       |
+| Top1 Document           | 34.46   | 44.41      | 142      | 60.82     | 68.70     | 143       | 29.20 | 38.93    | 147   | 26.79 | 31.87   | 153   |
+| Top3 Documents          | 37.78   | 48.45      | 427      | 62.60     | 71.02     | 430       | 31.64 | 41.87    | 442   | 27.89 | 33.58   | 460   |
+| Top5 Documents          | 38.03   | 49.16      | 712      | 64.10     | 72.48     | 715       | 32.99 | 43.69    | 737   | 29.64 | 35.21   | 766   |
+| Top10 Documents         | 38.67   | 50.03      | 1425     | 64.40     | 72.92     | 1428      | 33.95 | 44.88    | 1471  | 31.04 | 36.75   | 1531  |
+| Compression of top 5 do | cuments | ;          |          |           |           |           |       |          |       |       |         |       |
+| BM25                    | 25.23   | 36.47      | 37       | 55.36     | 63.90     | 39        | 24.18 | 35.73    | 71    | 25.42 | 30.29   | 68    |
+| Qwen2.5-1.5B            | 31.94   | 43.03      | 36       | 57.99     | 66.70     | 30        | 27.36 | 37.47    | 33    | 25.93 | 31.18   | 32    |
+| DeepSeek-V3 (671B)      | 37.73   | 50.39      | 54       | 64.13     | 73.20     | 50        | 33.59 | 44.83    | 48    | 27.99 | 32.67   | 92    |
+| RECOMP-Abs (1.5B)       | 34.18   | 46.26      | 58       | 60.31     | 68.50     | 53        | 28.96 | 39.95    | 56    | 30.25 | 36.73   | 52    |
+| RECOMP-Ext (1.5B)       | 33.84   | 46.05      | 56       | 60.18     | 68.39     | 48        | 29.93 | 41.09    | 45    | 30.78 | 37.07   | 51    |
+| NoiseFilter-IB (1.5B)   | 35.15   | 45.94      | 48       | 59.51     | 68.15     | 35        | 27.97 | 38.62    | 38    | 27.85 | 34.69   | 40    |
+| LongLLMLingua (1.5B)    | 33.65   | 43.15      | 152      | 58.96     | 66.82     | 148       | 28.03 | 38.49    | 149   | 29.37 | 33.62   | 153   |
+| QGC (1.5B)              | 36.23   | 45.88      | 49       | 61.02     | 68.45     | 47        | 29.16 | 40.05    | 45    | 31.14 | 36.83   | 51    |
+| CORE (1.5B)             | 41.02   | 50.40      | 46       | 65.63     | 72.55     | 32        | 33.67 | 45.06    | 36    | 36.72 | 42.05   | 49    |
+| Compression of top 10 a | locumen | ts (with t | the comp | ressor ti | rained or | ı top 5 d | locs) |          |       |       |         |       |
+| BM25                    | 25.91   | 36.88      | 38       | 55.28     | 63.16     | 37        | 23.49 | 35.01    | 68    | 25.61 | 30.54   | 65    |
+| Qwen2.5-1.5B            | 32.94   | 44.84      | 40       | 58.45     | 67.31     | 33        | 28.17 | 38.48    | 36    | 26.22 | 31.57   | 34    |
+| DeepSeek-V3 (671B)      | 37.79   | 51.07      | 56       | 65.29     | 74.45     | 53        | 34.62 | 45.69    | 50    | 29.00 | 34.64   | 40    |
+| RECOMP-Abs (1.5B)       | 34.40   | 46.93      | 59       | 61.42     | 69.88     | 52        | 31.54 | 42.92    | 52    | 31.98 | 38.16   | 49    |
+| RECOMP-Ext (1.5B)       | 33.96   | 46.34      | 60       | 61.03     | 69.51     | 50        | 31.92 | 43.18    | 55    | 32.52 | 38.87   | 44    |
+| NoiseFilter-IB (1.5B)   | 35.36   | 46.24      | 50       | 59.92     | 68.32     | 38        | 28.21 | 38.83    | 38    | 28.63 | 35.16   | 42    |
+| LongLLMLingua (1.5B)    | 33.78   | 43.37      | 154      | 59.17     | 66.97     | 150       | 28.33 | 38.95    | 148   | 29.62 | 34.11   | 151   |
+| QGC (1.5B)              | 36.03   | 45.62      | 50       | 61.23     | 68.74     | 49        | 29.12 | 39.63    | 46    | 31.71 | 37.52   | 50    |
+| CORE (1.5B)             | 41.88   | 51.26      | 52       | 66.76     | 73.64     | 33        | 34.68 | 46.35    | 37    | 37.99 | 43.28   | 48    |
+
+**Retrieval Corpus and Retrievers.** Following previous studies (Xu et al., 2024), we use the Wikipedia corpus from December 20, 2018, as the retrieval source for all four datasets. The articles are segmented into non-overlapping 100-word documents. To ensure that our method is not dependent on a specific retriever, we experiment with several mainstream retrievers. Specifically, we use DPR (Karpukhin et al., 2020) for NQ, a hybrid of DPR and BM25 (Robertson et al., 1995) for TriviaQA, and the Contriever model (Izacard et al., 2021) trained on the MS MARCO dataset (Nguyen et al., 2016) for HotpotQA and 2WikiMultihopQA.
+
+**Baselines.** To evaluate the effectiveness of our method, we compared it against various baselines. First, we evaluated the uncompressed approach—retaining the original in-context RALM setup—by prepending the top 1, 3, 5, and 10 retrieved documents to the prompt. We also tested alternative compression methods, including the traditional BM25 algorithm (which ranks sentences by their similarity to the input), off-the-shelf Qwen2.5-1.5B-Instruct model (with comparable parameter size to our approach), and DeepSeek-V3 model (671B parameters, far exceeding our compressor's capacity). Furthermore, we included state-of-the-art context compression methods for RAG, RECOMP (Xu et al., 2024), NoiseFilter-IB (Zhu et al., 2024), LongLLMLingua (Jiang et al., 2024) and QGC (Cao et al., 2024). For RECOMP, we evaluated both its abstractive and extractive variants. To ensure a fair comparison, all trainable approaches were all trained using the same model.
+
+**Implementation Details.** For the distillation warm-up phase, we perform full-parameter supervised fine-tuning on the off-the-shell language model for two epochs using LLaMA-Factory  $^1$ . This warmed-up model then serves as the initializer for the subsequent reinforcement learning phase. We adopt the Verl framework  $^2$  for RL training. The initialized compression model is trained for two epochs on each dataset. Training is conducted on eight NVIDIA H20 GPUs using full parameter GRPO optimization, with a learning rate of 1e-5, a batch size of 256, five rollouts per sample, and a KL loss coefficient of 0.001. The downstream LLM (M) used for reward generation is served using the vLLM inference engine during RL training.
+
+<sup>&</sup>lt;sup>1</sup>https://github.com/hiyouga/LLaMA-Factory
+
+<sup>&</sup>lt;sup>2</sup>https://github.com/volcengine/verl
+
+{6}------------------------------------------------
+
+### 3.2 OVERALL PERFORMANCE
+
+**329**
+
+**334 335 336**
+
+**364**
+
+**368 369 370**
+
+The detailed comparison results are presented in Table [1.](#page-5-0) We evaluate the following approaches: traditional RAG without compression using the top 1, 3, 5, and 10 documents prepended to the original input; BM25-based compression; off-the-shelf LLMs (Qwen2.5-1.5B-Instruct and DeepSeek-V3); and state-of-the-art trainable RAG compressors (RECOMP-Abstractive, RECOMP-Extractive, and NoiseFilter-IB). For fair comparison, all trainable methods were trained using the same backbone model, Qwen2.5-1.5B-Instruct. These compressors were trained on five-document inputs, and we report their performance both on in-domain five-document compression and out-of-domain generalization to ten-document compression.
+
+RAG versus No RAG. As shown in Table [1,](#page-5-0) prepending the original input with retrieved documents yields a substantial improvement over the no-retrieval baseline, indicating that these documents provide valuable information for generating the answer. Performance improves as the number of retrieved documents increases from 1 to 10, albeit with diminishing returns—a trend attributable to the decreasing relevance of lower-ranked documents. However, this performance gain comes at the cost of a significant increase in the number of additional tokens the LLM must encode, rising from 0 to over 1,400 tokens.
+
+Lossless Compression of CORE. Our compressor was trained using a context of five documents. As presented in Table [1,](#page-5-0) compared to prepending the full content of all five documents to the original input (i.e., the uncompressed baseline), our method achieves a high compression rate of approximately 6% *with no loss in performance*. Remarkably, on all four datasets, the approach not only maintains performance but also enhances EM by 1 to 7 points. For example, on NQ, the model achieves an EM of 41.02 with compressed input, surpassing the uncompressed score of 38.03.
+
+Comparison with Compression Baselines. When compressing the top five documents, all baseline compression methods result in performance degradation to varying degrees compared to the uncompressed baseline. Specifically, BM25 leads to a substantial performance drop. Using the offthe-shelf Qwen2.5-1.5B-Instruct model for compression yields better results than BM25 but still underperforms relative to the uncompressed baseline. Surprisingly, even the large-scale DeepSeek-V3 model (with 671B parameters) achieves only near-lossless compression on TriviaQA and HotpotQA, while performance on NQ and 2Wiki remains below the uncompressed baseline. As for the trained compressors (RECOMP, NoiseFilter-IB, LongLLMLingua, and QGC), all of them exhibit performance degradation compared to no compression. The decline generally ranges from 2 to 6 EM poin ts across nearly all datasets. In contrast, our method, CORE, delivers the best performance. It not only surpasses compression methods of comparable size by 4-5 EM points but also maintains an advantage over the hundreds-of-times-larger DeepSeek model. This clearly demonstrates the benefit and importance of end-to-end optimization.
+
+Generalization to 10-Doc Compression. When generalizing the trained compressors to handle the top-10 documents without retraining, the aforementioned conclusions remain valid. CORE continues to achieve lossless compression and performs best among all compression methods. On NQ, it achieves a token compression ratio of 3.6% while improving the EM by 3.2 points compared to using all ten documents. Similar trends are observed on TriviaQA, with a compression ratio of 2.3% and a gain of 2.4 EM points relative to the full-document baseline. On HotpotQA, CORE achieves a compression ratio of 2.5% and an improvement of 0.7 EM points. For 2WikiMultihopQA, it obtains a compression ratio of 3.1% along with a notable increase of 6.9 EM points.
+
+### 3.3 ROBUSTNESS AND GENERALIZATION ABILITY ANALYSIS
+
+Robustness of CORE Across Compressor Architectures. To evaluate whether the effectiveness of CORE depends on the choice of compressor model, we compared the performance of compressors trained using different model architectures—while keeping the downstream LLM (Qwen2.5-14B-Instruct) fixed. The compressors tested include LLaMA3.2-1B-Instruct, Qwen2.5-1.5B-Instruct, and LLaMA3.2-3B-Instruct, which vary in architecture and parameter count. As shown in Figure [3,](#page-7-0) the results indicate that: (1) These trained compressors consistently achieve lossless compression and outperform the uncompressed baseline (represented by the red reference line in the figure, which corresponds to prepending the full document content), confirming that our training framework is robust and not tied to a specific compressor architecture. (2) Compression performance improves as the size of the compressor model increases, consistent with scaling laws. More detailed results can be found in Tables [7](#page-17-0) and [8.](#page-17-1)
+
+{7}------------------------------------------------
+
+<span id="page-7-0"></span>![](_page_7_Figure_1.jpeg)
+
+Figure 3: The impact of different models used to train the compressor.
+
+Table 2: Zero-shot transfer of the trained compressor to Llama-3.1-8B-Instruct.
+
+<span id="page-7-1"></span>
+
+|                      | NQ      |       |       | 7     | TriviaQA |       |       | HotpotQA |       |       | 2WikiMultihopQA |       |  |
+|----------------------|---------|-------|-------|-------|----------|-------|-------|----------|-------|-------|-----------------|-------|--|
+|                      | EM      | F1    | # tok | EM    | F1       | # tok | EM    | F1       | # tok | EM    | F1              | # tok |  |
+| No Retrieval         | 24.04   | 34.91 | 0     | 55.64 | 62.57    | 0     | 19.93 | 27.75    | 0     | 27.64 | 31.18           | 0     |  |
+| RAG without compres  | ssion   |       |       |       |          |       |       |          |       |       |                 |       |  |
+| Top1 Document        | 33.80   | 44.06 | 142   | 59.17 | 67.50    | 143   | 27.95 | 37.49    | 147   | 28.41 | 33.43           | 153   |  |
+| Top3 Documents       | 36.87   | 47.81 | 427   | 61.13 | 70.06    | 430   | 30.17 | 40.71    | 442   | 28.67 | 34.23           | 460   |  |
+| Top5 Documents       | 37.65   | 48.87 | 712   | 62.26 | 71.04    | 715   | 31.44 | 42.16    | 737   | 29.43 | 35.18           | 766   |  |
+| Top10 Documents      | 38.12   | 49.93 | 1425  | 63.95 | 72.71    | 1428  | 32.19 | 42.62    | 1471  | 30.45 | 36.04           | 1531  |  |
+| Compression of top 5 | docume  | nts   |       |       |          |       |       |          |       |       |                 |       |  |
+| Qwen2.5-1.5B         | 32.60   | 44.21 | 36    | 56.76 | 65.77    | 30    | 26.86 | 36.90    | 33    | 25.45 | 30.88           | 32    |  |
+| DeepSeek-V3 (671B)   | 37.56   | 50.11 | 54    | 62.52 | 72.34    | 50    | 33.05 | 44.25    | 48    | 28.64 | 33.87           | 92    |  |
+| RECOMP-Abs (1.5B)    | 33.41   | 45.50 | 58    | 58.50 | 67.37    | 53    | 28.85 | 39.76    | 56    | 31.63 | 37.81           | 52    |  |
+| RECOMP-Ext (1.5B)    | 33.12   | 45.06 | 60    | 57.98 | 66.84    | 55    | 29.03 | 40.04    | 52    | 31.85 | 38.02           | 55    |  |
+| CORE (1.5B)          | 40.72   | 50.00 | 46    | 64.08 | 71.13    | 32    | 32.17 | 43.71    | 36    | 35.99 | 41.42           | 49    |  |
+| Compression of top 1 | 0 docum | ents  |       |       |          |       |       |          |       |       |                 |       |  |
+| Qwen2.5-1.5B         | 32.88   | 44.66 | 40    | 57.44 | 66.56    | 33    | 27.31 | 37.31    | 36    | 25.80 | 31.30           | 34    |  |
+| DeepSeek-V3 (671B)   | 37.49   | 51.28 | 56    | 63.79 | 73.80    | 53    | 34.24 | 45.35    | 50    | 31.45 | 37.09           | 40    |  |
+| RECOMP-Abs (1.5B)    | 34.18   | 46.80 | 59    | 59.69 | 68.89    | 52    | 30.17 | 41.42    | 55    | 33.61 | 39.78           | 44    |  |
+| RECOMP-Ext (1.5B)    | 34.06   | 46.55 | 60    | 59.33 | 68.71    | 50    | 30.52 | 41.98    | 55    | 33.52 | 39.42           | 44    |  |
+| CORE (1.5B)          | 41.77   | 51.27 | 52    | 65.25 | 72.45    | 33    | 33.25 | 45.09    | 37    | 37.59 | 42.87           | 48    |  |
+
+Transferability of CORE Across Downstream LLMs. We evaluate the transferability of our trained compressor and other baseline compressors to a new downstream LLM, the LLaMA-3.1-8B model. The results are presented in Table 2. Note that all trainable compressors, including our own, were trained using feedback generated by Qwen2.5-14B-Instruct. The findings reveal that existing trained compressors (e.g., RECOMP) exhibit limited generalization capability, as indicated by a larger performance gap relative to the baseline of prepending full documents. In contrast, CORE demonstrates stronger generalization, achieving lossless compression on the new downstream LLM. Specifically, it not only maintains performance but also surpasses the full-document baseline (i.e., no compression) across all four datasets while retaining a high compression rate. These results suggest that the summaries produced by our method are inherently high-quality and preserve key information necessary for accurate answering, thereby enabling effective transfer to other LLMs.
+
+### 3.4 ABLATION STUDY
+
+Table 3 presents an ablation study on the two stages of our method: distillation and GRPO. Here, "w/o distillation" denotes training the compressor with GRPO directly, bypassing the warm-start phase, while "w/o GRPO" indicates using only the distillation step without subsequent reinforcement learning. The results demonstrate that removing either component leads to performance degradation, confirming the necessity of both. In addition, the decline is more substantial when GRPO is omitted, highlighting the crucial role of reinforcement learning in the absence of explicit supervision. Distillation injects external knowledge into the model, providing a favorable starting point for RL training and thereby enabling RL to more effectively unlock the compressor's full potential.
+
+{8}------------------------------------------------
+
+Table 3: Ablation study.
+
+<span id="page-8-0"></span>
+
+| Dataset  | Metric | w/o distillation | w/o GRPO | CORE  |
+|----------|--------|------------------|----------|-------|
+|          | EM     | 36.37            | 34.18    | 41.02 |
+| NQ       | F1     | 46.91            | 46.26    | 50.40 |
+|          | EM     | 65.23            | 60.31    | 65.63 |
+| TQA      | F1     | 72.41            | 68.50    | 72.55 |
+|          | EM     | 32.01            | 28.96    | 33.67 |
+| HotpotQA | F1     | 42.73            | 39.95    | 45.06 |
+|          | EM     | 31.40            | 30.25    | 36.72 |
+| 2Wiki    | F1     | 36.89            | 36.73    | 42.05 |
+
+![](_page_8_Figure_3.jpeg)
+
+Figure 4: The impact of α.
+
+### 3.5 THE IMPACT OF α.
+
+Figure [4](#page-8-0) illustrates the performance of our method under different values of α, which controls the weighting coefficient of the F1 reward term. Setting α = 0 corresponds to using only the EM reward. Results indicate that performance improves when α > 0 across all datasets, demonstrating the effectiveness of the F1 reward in mitigating the sparsity issue associated with the EM reward. However, the optimal value of α is dataset-dependent; values between 0.2 and 0.5 generally yield strong performance.
+
+### 3.6 CASE STUDY
+
+To conduct an in-depth analysis of the advantages of our compressor, we performed case studies on one single-hop QA dataset (NQ) and one multi-hop QA dataset (2Wiki), with the results presented in Table [4](#page-14-0) and Table [5,](#page-15-0) respectively. For each case, we compared the summaries generated by off-the-shelf Qwen2.5-1.5B-Instruct, RECOMP, and our method CORE based on the same set of documents, as well as the predicted answers generated by the LLM after prepending these summaries. As shown in the tables, although the summaries produced by off-the-shelf Qwen2.5-1.5B are concise, they largely fail to capture key information relevant to answering the question. In contrast, RECOMP demonstrates better summarization capability but is prone to being overwhelmed by lengthy documents, resulting in misjudgments and even generating misleading information—such as the statement in Table [4:](#page-14-0) "*The U.S. stopped drafting for the Vietnam War after the Selective Service System was officially abolished in December 1972*"—which leads the downstream LLM to produce the incorrect answer "1972". Our method, CORE, accurately extracts answer-critical information from lengthy documents, exemplified by the summary: "*The U.S. stopped drafting for the Vietnam War in 1973 after announcing the decision by Secretary of Defense Melvin Laird earlier that year*", thereby enabling the LLM to generate the correct answer "1973". This indicates that our compressor, trained with an end-task target-oriented reinforcement learning optimization strategy, can produce document summaries that are most helpful for answering the given question while effectively filtering out irrelevant information.
+
+## 4 RELATED WORK
+
+Context Compression in RAG. RAG enhances the performance of LLMs on knowledge-intensive tasks by retrieving the most relevant documents from extensive knowledge bases and prepending them as contextual information to the original input [\(Ram et al., 2023;](#page-11-0) [Fan et al., 2024;](#page-10-0) [Lin et al.,](#page-11-6) [2023;](#page-11-6) [Shi et al., 2023\)](#page-12-8). However, this approach requires the LLM to process significantly longer token sequences, resulting in increased computational costs. To mitigate this issue, researchers have begun to explore methods for compressing retrieved documents in RAG systems [\(Xu et al., 2024;](#page-12-0) [Cao et al., 2024;](#page-10-3) [Jin et al., 2024b;](#page-10-1) [Zhu et al., 2024;](#page-12-3) [Kim & Thorne, 2025;](#page-11-7) [Rau et al., 2024;](#page-12-9) [Wu et al.,](#page-12-1) [2025;](#page-12-1) [Louis et al., 2025;](#page-11-8) [Jin et al., 2024a;](#page-10-2) [Li et al., 2024a;](#page-11-9)[b;](#page-11-10) [Zhang et al., 2024a\)](#page-12-2). For instance, [Xu et al.](#page-12-0) [\(2024\)](#page-12-0) propose compressing retrieved documents into textual summaries before in-context augmentation, training the compressor through data selection and distillation. Similarly, [Cao et al.](#page-10-3) [\(2024\)](#page-10-3) introduce a Query-Guided Compressor (QGC) that uses queries to guide the compression process, effectively preserving essential information. [Jin et al.](#page-10-1) [\(2024b\)](#page-10-1) refine retrieved documents into Key Supporting Evidence (KSE) through a combination of knowledge synthesis, supervised fine-tuning (SFT), and preference alignment. Meanwhile, [Zhu et al.](#page-12-3) [\(2024\)](#page-12-3) present an informationtheoretic approach called NoiseFilter-IB, which filters noise by maximizing the mutual information
+
+{9}------------------------------------------------
+
+**509**
+
+**529 530**
+
+**538 539** between the compressed content and the ground-truth output. Additionally, [Kim & Thorne](#page-11-7) [\(2025\)](#page-11-7) train a compressor to extract critical information using reward functions based on predefined heuristic rules. However, most of these methods are heuristic in nature, and due to the lack of ideal compression labels, the compressed content they produce cannot be guaranteed to benefit downstream LLMs. In contrast, our method, CORE, adopts an end-to-end optimization approach to address these limitations.
+
+Reinforcement Learning. Reinforcement learning (RL) has recently achieved notable success, enabling LLMs to develop reasoning capabilities without explicit step-by-step supervision [\(Liu et al.,](#page-11-2) [2024;](#page-11-2) [Shao et al., 2024;](#page-12-4) [Guo et al., 2025\)](#page-10-9). Building on these advances, several studies have applied RL to improve RAG [\(Ke et al., 2024\)](#page-11-11). For example, [Kulkarni et al.](#page-11-12) [\(2024\)](#page-11-12) use RL to autonomously decide whether to retrieve documents, while [Zhang et al.](#page-12-10) [\(2024b\)](#page-12-10) employ RL to optimize the ranking of retrieved documents. Similarly, [Mao et al.](#page-11-13) [\(2024\)](#page-11-13) propose a framework for training query rewriting models for RAG without relying on human annotations. MMOA-RAG [\(Chen et al., 2025b\)](#page-10-10) enhances RAG through multi-agent reinforcement learning, incorporating a query rewriter, retriever, and generator. Meanwhile, RL has been applied to address other challenges in RAG, such as enhancing the quality of retrieved content. For instance, Oreo [\(Li & Ramakrishnan, 2025\)](#page-11-14) trains a reconstructor with PPO and ROUGE-based rewards to rewrite passages for improved performance. In contrast, our work CORE tackles the distinct problem of computational efficiency. Our approach leverages the GRPO algorithm, direct task-performance rewards, and a lightweight compressor to enhance efficiency without compromising accuracy. Moreover, a line of research has utilized RL to integrate search with reasoning in a step-by-step manner [\(Singh et al., 2025\)](#page-12-11). For instance, [Chen](#page-10-4) [et al.](#page-10-4) [\(2025a\)](#page-10-4) introduce a framework called ReSearch, which trains LLMs to reason with search using RL, without requiring supervised data for reasoning steps. Related approaches include R1- Searcher [\(Song et al., 2025\)](#page-12-5), WebThinker [\(Li et al., 2025\)](#page-11-15), and DeepResearcher [\(Zheng et al., 2025\)](#page-12-12). Although these methods are end-to-end, they differ fundamentally from our problem setting. These approaches typically involve directly training the LLM generator—which tends to be a large-scale model with a high parameter count. However, such training becomes infeasible when the model is a black box (e.g., GPT-4), as internal weights or gradients are inaccessible. Furthermore, these methods introduce extensive internal thinking processes that substantially increase context length and inference time. In contrast, our approach treats the generator LLM as a fixed black-box model and trains only a lightweight plug-in compressor to produce document summaries. This design significantly improves both training and inference efficiency. A parallel line of work applies RL to prompt compression. For example, PCRL [\(Jung & Kim, 2024\)](#page-11-16) and TACO-RL [\(Shandilya et al., 2025\)](#page-12-13) learn to compress prompts via token-level keep-or-drop decisions, using the similarity between model outputs with compressed and original prompts as the reward. Our work, CORE, introduces key distinctions in both objective and methodology. First, we target the more complex problem of compressing multiple retrieved documents in RAG, rather than single prompts. Second, we employ a generative compressor that can rephrase and synthesize content, instead of making token-level binary actions. Most importantly, we optimize compression using a direct task-performance reward with GRPO, which enables us to achieve true lossless compression at significantly higher ratios—a stark contrast to the performance degradation observed in prior prompt compression methods.
+
+## 5 CONCLUSION
+
+This paper analyzes the limitations of current context compression methods for RAG. A primary challenge is the lack of optimal reference summaries for supervised learning, which often results in performance degradation in downstream tasks. To overcome this, we frame context compression as a reinforcement learning problem, utilizing downstream task performance as a reward signal to train the compression policy, thereby enabling end-to-end optimization. Extensive experiments demonstrate that our proposed method, CORE, achieves effectively lossless compression by maintaining a high compression ratio while preserving original task performance. Surprisingly, CORE not only preserves but actually enhances performance on all downstream tasks. Further in-depth analysis provides additional insights into the efficacy of our approach.
+
+{10}------------------------------------------------
+
+## ETHICS STATEMENT
+
+<span id="page-10-3"></span>**558 559 560**
+
+<span id="page-10-10"></span>**564**
+
+<span id="page-10-7"></span>**579**
+
+The authors affirm that this work adheres to the ICLR Code of Ethics. It involves no human subjects, sensitive or private data, or applications posing potential ethical risks. All resources utilized are publicly available and appropriately licensed. The research was conducted in accordance with ethical and legal standards.
+
+## REPRODUCIBILITY STATEMENT
+
+This paper includes detailed descriptions of the experimental setups, implementation details, hyperparameter selections, and evaluation procedures to facilitate full verification of the reported results. To further support reproducibility, the complete source code and experimental scripts are available at the following anonymous repository: [https://anonymous.4open.science/](https://anonymous.4open.science/r/CORE-28B4) [r/CORE-28B4](https://anonymous.4open.science/r/CORE-28B4).
+
+## REFERENCES
+
+- Zhiwei Cao, Qian Cao, Yu Lu, Ningxin Peng, Luyang Huang, Shanbo Cheng, and Jinsong Su. Retaining key information under high compression ratios: Query-guided compressor for llms. In *Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)*, pp. 12685–12695, 2024.
+- <span id="page-10-4"></span>Mingyang Chen, Tianpeng Li, Haoze Sun, Yijie Zhou, Chenzheng Zhu, Haofen Wang, Jeff Z Pan, Wen Zhang, Huajun Chen, Fan Yang, et al. Learning to reason with search for llms via reinforcement learning. *arXiv preprint arXiv:2503.19470*, 2025a.
+- Yiqun Chen, Lingyong Yan, Weiwei Sun, Xinyu Ma, Yi Zhang, Shuaiqiang Wang, Dawei Yin, Yiming Yang, and Jiaxin Mao. Improving retrieval-augmented generation through multi-agent reinforcement learning. *arXiv preprint arXiv:2501.15228*, 2025b.
+- <span id="page-10-0"></span>Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat-Seng Chua, and Qing Li. A survey on rag meeting llms: Towards retrieval-augmented large language models. In *Proceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining*, pp. 6491–6501, 2024.
+- <span id="page-10-9"></span>Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. *arXiv preprint arXiv:2501.12948*, 2025.
+- <span id="page-10-6"></span>Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa. Constructing a multi-hop qa dataset for comprehensive evaluation of reasoning steps. *arXiv preprint arXiv:2011.01060*, 2020.
+- Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. Unsupervised dense information retrieval with contrastive learning. *arXiv preprint arXiv:2112.09118*, 2021.
+- <span id="page-10-8"></span>Huiqiang Jiang, Qianhui Wu, Xufang Luo, Dongsheng Li, Chin-Yew Lin, Yuqing Yang, and Lili Qiu. Longllmlingua: Accelerating and enhancing llms in long context scenarios via prompt compression. In *Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)*, pp. 1658–1677, 2024.
+- <span id="page-10-2"></span>Bowen Jin, Jinsung Yoon, Jiawei Han, and Sercan O Arik. Long-context llms meet rag: Overcoming challenges for long inputs in rag. *arXiv preprint arXiv:2410.05983*, 2024a.
+- <span id="page-10-1"></span>Jiajie Jin, Yutao Zhu, Yujia Zhou, and Zhicheng Dou. Bider: Bridging knowledge inconsistency for efficient retrieval-augmented llms via key supporting evidence. In *Findings of the Association for Computational Linguistics ACL 2024*, pp. 750–761, 2024b.
+- <span id="page-10-5"></span>Mandar Joshi, Eunsol Choi, Daniel S Weld, and Luke Zettlemoyer. Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension. *arXiv preprint arXiv:1705.03551*, 2017.
+
+{11}------------------------------------------------
+
+<span id="page-11-12"></span><span id="page-11-11"></span><span id="page-11-7"></span><span id="page-11-3"></span>**604 605 606**
+
+<span id="page-11-15"></span><span id="page-11-14"></span>**617**
+
+<span id="page-11-10"></span><span id="page-11-9"></span>**619**
+
+<span id="page-11-6"></span><span id="page-11-2"></span><span id="page-11-1"></span>**634**
+
+<span id="page-11-13"></span><span id="page-11-8"></span><span id="page-11-5"></span><span id="page-11-0"></span>**636**
+
+- <span id="page-11-16"></span><span id="page-11-4"></span>Hoyoun Jung and Kyung-Joong Kim. Discrete prompt compression with reinforcement learning. *IEEE Access*, 12:72578–72587, 2024.
+  - Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick SH Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. Dense passage retrieval for open-domain question answering. In *EMNLP (1)*, pp. 6769–6781, 2020.
+  - Zixuan Ke, Weize Kong, Cheng Li, Mingyang Zhang, Qiaozhu Mei, and Michael Bendersky. Bridging the preference gap between retrievers and llms. *arXiv preprint arXiv:2401.06954*, 2024.
+  - Sangryul Kim and James Thorne. Context filtering with reward modeling in question answering. In *Proceedings of the 31st International Conference on Computational Linguistics*, pp. 11048– 11055, 2025.
+  - Mandar Kulkarni, Praveen Tangarajan, Kyung Kim, and Anusua Trivedi. Reinforcement learning for optimizing rag for domain chatbots. *arXiv preprint arXiv:2401.06800*, 2024.
+  - Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, et al. Natural questions: a benchmark for question answering research. *Transactions of the Association for Computational Linguistics*, 7:453–466, 2019.
+  - Sha Li and Naren Ramakrishnan. Oreo: A plug-in context reconstructor to enhance retrievalaugmented generation. In *Proceedings of the 2025 International ACM SIGIR Conference on Innovative Concepts and Theories in Information Retrieval (ICTIR)*, pp. 238–253, 2025.
+  - Xiaoxi Li, Jiajie Jin, Guanting Dong, Hongjin Qian, Yutao Zhu, Yongkang Wu, Ji-Rong Wen, and Zhicheng Dou. Webthinker: Empowering large reasoning models with deep research capability. *arXiv preprint arXiv:2504.21776*, 2025.
+  - Xinze Li, Sen Mei, Zhenghao Liu, Yukun Yan, Shuo Wang, Shi Yu, Zheni Zeng, Hao Chen, Ge Yu, Zhiyuan Liu, et al. Rag-ddr: Optimizing retrieval-augmented generation using differentiable data rewards. *arXiv preprint arXiv:2410.13509*, 2024a.
+  - Zhonghao Li, Xuming Hu, Aiwei Liu, Kening Zheng, Sirui Huang, and Hui Xiong. Refiner: Restructure retrieval content efficiently to advance question-answering capabilities. *arXiv preprint arXiv:2406.11357*, 2024b.
+  - Xi Victoria Lin, Xilun Chen, Mingda Chen, Weijia Shi, Maria Lomeli, Richard James, Pedro Rodriguez, Jacob Kahn, Gergely Szilvasy, Mike Lewis, et al. Ra-dit: Retrieval-augmented dual instruction tuning. In *The Twelfth International Conference on Learning Representations*, 2023.
+  - Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report. *arXiv preprint arXiv:2412.19437*, 2024.
+  - Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. Lost in the middle: How language models use long contexts. *arXiv preprint arXiv:2307.03172*, 2023.
+  - Maxime Louis, Herve D´ ejean, and St ´ ephane Clinchant. Pisco: Pretty simple compression for ´ retrieval-augmented generation. *arXiv preprint arXiv:2501.16075*, 2025.
+  - Shengyu Mao, Yong Jiang, Boli Chen, Xiao Li, Peng Wang, Xinyu Wang, Pengjun Xie, Fei Huang, Huajun Chen, and Ningyu Zhang. Rafe: Ranking feedback improves query rewriting for rag. *arXiv preprint arXiv:2405.14431*, 2024.
+  - Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh Tiwary, Rangan Majumder, and Li Deng. Ms marco: A human-generated machine reading comprehension dataset. 2016.
+  - Ori Ram, Yoav Levine, Itay Dalmedigos, Dor Muhlgay, Amnon Shashua, Kevin Leyton-Brown, and Yoav Shoham. In-context retrieval-augmented language models. *Transactions of the Association for Computational Linguistics*, 11:1316–1331, 2023.
+
+{12}------------------------------------------------
+
+<span id="page-12-13"></span>**654**
+
+<span id="page-12-11"></span><span id="page-12-8"></span><span id="page-12-5"></span><span id="page-12-4"></span><span id="page-12-1"></span>**656**
+
+<span id="page-12-12"></span><span id="page-12-10"></span><span id="page-12-6"></span><span id="page-12-3"></span><span id="page-12-2"></span><span id="page-12-0"></span>**684 685**
+
+- <span id="page-12-9"></span><span id="page-12-7"></span>David Rau, Shuai Wang, Herve D´ ejean, and St ´ ephane Clinchant. Context embeddings for efficient ´ answer generation in rag. *arXiv preprint arXiv:2407.09252*, 2024.
+  - Stephen E Robertson, Steve Walker, Susan Jones, Micheline M Hancock-Beaulieu, Mike Gatford, et al. *Okapi at TREC-3*. British Library Research and Development Department, 1995.
+  - Shivam Shandilya, Menglin Xia, Supriyo Ghosh, Huiqiang Jiang, Jue Zhang, Qianhui Wu, Victor Ruhle, and Saravan Rajmohan. Taco-rl: Task aware prompt compression optimization with rein- ¨ forcement learning. In *Findings of the Association for Computational Linguistics: ACL 2025*, pp. 1582–1597, 2025.
+  - Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. *arXiv preprint arXiv:2402.03300*, 2024.
+  - Weijia Shi, Sewon Min, Michihiro Yasunaga, Minjoon Seo, Rich James, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. Replug: Retrieval-augmented black-box language models. *arXiv preprint arXiv:2301.12652*, 2023.
+  - Aditi Singh, Abul Ehtesham, Saket Kumar, and Tala Talaei Khoei. Agentic retrieval-augmented generation: A survey on agentic rag. *arXiv preprint arXiv:2501.09136*, 2025.
+  - Huatong Song, Jinhao Jiang, Yingqian Min, Jie Chen, Zhipeng Chen, Wayne Xin Zhao, Lei Fang, and Ji-Rong Wen. R1-searcher: Incentivizing the search capability in llms via reinforcement learning. *arXiv preprint arXiv:2503.05592*, 2025.
+  - Chenyuan Wu, Ninglu Shao, Zheng Liu, Shitao Xiao, Chaozhuo Li, Chen Zhang, Senzhang Wang, and Defu Lian. Lighter and better: Towards flexible context adaptation for retrieval augmented generation. In *Proceedings of the Eighteenth ACM International Conference on Web Search and Data Mining*, pp. 271–280, 2025.
+  - Fangyuan Xu, Weijia Shi, and Eunsol Choi. Recomp: Improving retrieval-augmented lms with context compression and selective augmentation. In *The Twelfth International Conference on Learning Representations*, 2024.
+  - Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W Cohen, Ruslan Salakhutdinov, and Christopher D Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question answering. *arXiv preprint arXiv:1809.09600*, 2018.
+  - Qianchi Zhang, Hainan Zhang, Liang Pang, Hongwei Zheng, and Zhiming Zheng. Adacomp: Extractive context compression with adaptive predictor for retrieval-augmented large language models. *arXiv preprint arXiv:2409.01579*, 2024a.
+  - Taolin Zhang, Dongyang Li, Qizhou Chen, Chengyu Wang, Longtao Huang, Hui Xue, Xiaofeng He, and Jun Huang. R 4: Reinforced retriever-reorder-responder for retrieval-augmented large language models. In *ECAI 2024*, pp. 2314–2321. IOS Press, 2024b.
+  - Yuxiang Zheng, Dayuan Fu, Xiangkun Hu, Xiaojie Cai, Lyumanshan Ye, Pengrui Lu, and Pengfei Liu. Deepresearcher: Scaling deep research via reinforcement learning in real-world environments. *arXiv preprint arXiv:2504.03160*, 2025.
+  - Kun Zhu, Xiaocheng Feng, Xiyuan Du, Yuxuan Gu, Weijiang Yu, Haotian Wang, Qianglong Chen, Zheng Chu, Jingchang Chen, and Bing Qin. An information bottleneck perspective for effective noise filtering on retrieval-augmented generation. In *Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)*, pp. 1044–1069, 2024.
+
+{13}------------------------------------------------
+
+<span id="page-13-0"></span>Compress the information in the retrieved documents into a 2-sentence summary that could be used to answer the question. If the documents do not contain relevant information, simply output " ".
+
+Question: {*Question*}
+
+**709**
+
+**724**
+
+Retrieved documents: {*Documents*}
+
+Compressed documents:
+
+Figure 5: Prompt template used to generate the summary.
+
+### <span id="page-13-1"></span>[Instruction] Answer the question.
+
+IMPORTANT: Respond ONLY with the exact answer in the same format as the examples. Do NOT add any extra text, explanations, or punctuation. Do NOT include "Answer:" or any similar prefix in your response.
+
+### [Examples]
+
+Question: Which major Russian city borders the body of water in which Saaremaa is located?
+
+Answer: Saint Petersburg
+
+Question: Who was thee first president of the association that wrote the code of ethics for psychology?
+
+Answer: G. Stanley Hall
+
+Question: Where did the Baldevins bryllup director die?
+
+Answer: Copenhagen
+
+### [Current Question]
+
+**{***Summary of the retrieved documents***}** 
+
+Question: {*Question*}
+
+Answer:
+
+Figure 6: Prompt template for LLM QA.
+
+## A USE OF LLMS
+
+Large language models (LLMs) were employed solely as writing assistants to enhance the language, improve clarity, and check grammatical correctness. They were not used to generate research ideas, design or implement methodologies, conduct data analysis, or produce any of the results presented in this work. The authors assume full responsibility for the entire content of the paper.
+
+## B PROMPT TEMPLATES
+
+Figure [5](#page-13-0) displays the prompt employed to train the compressor model for generating a summary of the retrieved documents, conditioned on the given query. This prompt is notably concise. For end-task answer generation, the prompt provided to the LLM M is illustrated in Figure [6,](#page-13-1) which incorporates few-shot in-context examples, the (generated summary of) retrieved documents, and the question.
+
+## C CASE STUDY
+
+To conduct an in-depth analysis of the advantages of our compressor, we performed case studies on one single-hop QA dataset (NQ) and one multi-hop QA dataset (2Wiki), with the results presented in Table [4](#page-14-0) and Table [5,](#page-15-0) respectively. For each case, we compared the summaries generated by off-the-shelf Qwen2.5-1.5B-Instruct, RECOMP, and our method CORE based on the same set of documents, as well as the predicted answers generated by the LLM after prepending these summaries. As shown in the tables, although the summaries produced by off-the-shelf Qwen2.5-1.5B
+
+{14}------------------------------------------------
+
+### Table 4: Case study on NQ dataset.
+
+**758**
+
+**759**
+
+**761**
+
+**764**
+
+**779 780 781**
+
+# <span id="page-14-0"></span>Question: when did the us stop drafting for the vietnam war? Gold answer: [1973]
+
+### Top-5 documents:
+
+last men conscripted, who were born in 1952 and who reported for duty in June 1973. On February 2, 1972, a drawing was held to determine draft priority numbers for men born in 1953, but in early 1973 it was announced by Secretary of Defense Melvin Laird that no further draft orders would be issued. In March 1973, 1974, and 1975, the Selective Service assigned draft priority numbers for all men born in 1954, 1955, and 1956, in case the draft was extended, but it never was. Command Sergeant Major Jeff Mellinger, believed to be the last drafted enlisted ranked.
+
+The Gates Commission issued its report in February 1970, describing how adequate military strength could be maintained without having conscription. The existing draft law was expiring at the end of June 1971, but the Department of Defense and Nixon administration decided the draft needed to continue for at least some time. In February 1971, the administration requested of Congress a two-year extension of the draft, to June 1973. Senatorial opponents of the war wanted to reduce this to a one-year extension, or eliminate the draft altogether, or tie the draft renewal to a timetable for troop withdrawal from Vietnam; Senator.
+
+draft operated from 1940 until 1946 when further inductions were suspended, and its legislative authorization expired without further extension by Congress in 1947. During this time, more than 10 million men had been inducted into military service. However, the Selective Service System remained intact. Scattered opposition was encountered especially in the northern cities where some African-Americans protested the system. The tiny Nation of Islam was at the forefront, with many Black Muslims jailed for refusing the draft, and their leader Elijah Muhammed was sentenced to federal prison for 5 years for inciting draft resistance. Organized draft resistance also developed in.
+
+was the first time any form of student deferment was used. During the Korean War a student carrying at least twelve semester hours was spared until the end of his current semester. The United States breathed easier with the Korean War Armistice on July 27, 1953; however, technology brought new promises and threats. U.S. air and nuclear power fueled the Eisenhower doctrine of "massive retaliation". This strategy demanded more machines and fewer foot soldiers, so the draft slipped to the back burner. However, the head of the SSS, Maj. Gen. Hershey, urged caution fearing the conflict looming in Vietnam. In.
+
+policy of using the draft as force to compel "voluntary" enlistment was unique in U.S. history. Previous drafts had not aimed at encouraging individuals to sign up in order to gain preferential placement or less dangerous postings. However, the incremental buildup of Vietnam without a clear threat to the country bolstered this. Some estimates suggest conscription encompassed almost one-third of all eligible men during the period of 1965–69. This group represented those without exemption or resources to avoid military service. During the active combat phase, the possibility of avoiding combat by selecting their service and military specialty led as many.
+
+| Compressor                                 | Generated summary                                                                                                                                                                                                                                                                                                               | Predicted<br>answer |
+|--------------------------------------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|---------------------|
+| Off-the-shell<br>Qwen2.5-<br>1.5B-Instruct | The US stopped drafting for the Vietnam War in late 1972 after<br>the draft expired at the end of June 1971.                                                                                                                                                                                                                    | late 1972           |
+| RECOMP                                     | The U.S. stopped drafting for the Vietnam War after the<br>Selective Service System was officially abolished in December<br>1972 due to growing public opposition and the expiration of the<br>draft's legal authority in 1971. While the draft was formally<br>ended, some individuals were still consripted into the military | December<br>1972    |
+| Ours                                       | The U.S. stopped drafting for the Vietnam War in 1973 after<br>announcing the decision by Secretary of Defense Melvin Laird<br>earlier that year. Although the Selective Service System was<br>later updated to include draft priority numbers, this decision<br>marked the end of the draft's use for national service.        | 1973                |
+
+{15}------------------------------------------------
+
+### Table 5: Case study on 2Wiki dataset.
+
+**812 813**
+
+**814 815**
+
+**817**
+
+**829 830**
+
+**834**
+
+### <span id="page-15-0"></span>Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? ´ Gold answer: [Charles Armand Rene de La Tr ´ emoille] ´
+
+Top-5 documents:
+
+as at Versailles: he was brigadier of cavalry (January 1709), first gentleman of the King's chamber (June 1709), governor of Thouars (July 1709), and Marechal de camp (February 1719). His sister ´ Marie Armande Victoire de La Tremoille married Emmanuel Th ´ eodose de La Tour d'Auvergne. ´ On 13 April 1706 he married Marie-Madeleine Motier de La Fayette (1691–1717), the daughter of Rene-Armand, marquis de La Fayette and Marie-Madeleine de Marillac, and granddaughter of the author Marie-Madeleine Pioche de la Vergne, comtesse de la Fayette. They had one child, Charles Armand Rene de La Tr ´ emoille, born in 1708. Charles Louis Bretagne de La ´
+
+Charles Bretagne Marie de La Tremoille Charles Bretagne Marie de La Tr ´ emoille (24 March ´ 1764 – 10 November 1839), 8th duc de Thouars, was a French soldier and the son of Jean Bretagne Charles de La Tremoille and his wife, Marie-Maximilienne, princess of Salm-Kyrburg. La ´ Tremoille married Louise-Emmanuelle de Ch ´ atillon in 1781. She was a grand daughter of Louis ˆ Cesar de La Baume Le Blanc, the famous writer. The couple had one daughter: At the outbreak ´ of the French Revolution, he sided with King Louis XVI, as did the rest of his family. In 1789, La Tremoille and his parents ´
+
+Jean Bretagne Charles de La Tremoille Jean-Bretagne-Charles de La Tr ´ emoille (5 February 1737 ´ – 19 May 1792), 7th duc de Thouars, was a French soldier and the son of Charles Armand Rene´ de La Tremoille and his wife, Marie Hortense de La Tour d'Auvergne. La Tr ´ emoille began his ´ military career in the Seven Years' War as the head of the regiment of Aquitaine-Cavalry. He was seriously wounded and afterwards promoted to the rank of sergeant, and later of brigadier, of the armies of the king. During the French Revolution, La Tremoille remained faithful to King Louis ´ XVI. He emigrated in
+
+Charles Louis Bretagne de La Tremoille Charles Louis Bretagne de La Tr ´ emoille (15 March ´ 1683 – 9 October 1719), 6th duke of Thouars, was the son of Charles Belgique Hollande de La Tremoille and Madeleine de Cr ´ equy, daughter and heiress of Charles III de Cr ´ equy. He became ´ duke of Thouars upon his father's death in 1709. He was also count of Laval and Montfort and inherited the ancient French claim to the Kingdom of Naples. The Neapolitan claim gave the family the rank of "princes etrangers" at the French court. The Duke made a career in the ´ military as well
+
+de Saint-Didier. They had two daughters: Following her death, La Tremoille married for a third ´ time in 1830 to Valentine Eugenie Jos ´ ephine Walsh de Serrant. They had two children: La ´ Tremoille died shortly after his son's birth, in 1839. Charles Bretagne Marie de La Tr ´ emoille ´ Charles Bretagne Marie de La Tremoille (24 March 1764 ��� 10 November 1839), 8th duc de ´ Thouars, was a French soldier and the son of Jean Bretagne Charles de La Tremoille and his wife, ´ Marie-Maximilienne, princess of Salm-Kyrburg. La Tremoille married Louise-Emmanuelle de ´ Chatillon in 1781. She was a grand daughter of Louis C ˆ esar de ´
+
+| Compressor                                 | Generated summary                                                                                                                                                                                                                                                                                                           | Predicted<br>answer                             |
+|--------------------------------------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-------------------------------------------------|
+| Off-the-shell<br>Qwen2.5-<br>1.5B-Instruct | Charles Bretagne Marie de La Tremoille's paternal grandfather<br>´<br>was Jean Bretagne Charles de La Tremoille. ´                                                                                                                                                                                                          | Jean Bretagne<br>Charles de La<br>Tremoille ´   |
+| RECOMP                                     | Charles Bretagne Marie de La Tremoille's paternal grandfather<br>´<br>was Charles Louis Bretagne de La Tremoille, who was the son<br>´<br>of Charles Belgique Hollande de La Tremoille and Madeleine<br>´<br>de Crequy. This makes him the direct paternal grandfather of<br>´<br>Charles Bretagne Marie de La Tremoille. ´ | Charles Louis<br>Bretagne de<br>La Tremoille ´  |
+| Ours                                       | ´<br>Charles Bretagne Marie de La Tremoille's paternal grandfather<br>is Charles Armand Rene de La Tr<br>´<br>emoille, her father's father, the<br>´<br>7th duc de Thouars.                                                                                                                                                 | Charles<br>Armand Rene´<br>de La<br>Tremoille ´ |
+
+**861 863**
+
+**854 855 856**
+
+{16}------------------------------------------------
+
+**869**
+
+**884**
+
+**904**
+
+**906**
+
+Table 6: Failure Cases on NQ dataset.
+
+<span id="page-16-0"></span>
+
+| Question                                                                                      | Generated Summary                                                                                                                                                                                                                                                            | Predicted<br>Answer | Gold<br>Answer  |
+|-----------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|---------------------|-----------------|
+| who is the<br>president of<br>usa right now                                                   | The current president of the USA is **Donald<br>Trump**, who was elected on January 20, 2017, and<br>is the 45th president. He is a businessman and reality<br>TV personality, and his inauguration marked a shift in<br>the country's political landscape.                  | Joe Biden           | Donald<br>Trump |
+| who has been<br>ranked no. 1<br>in the latest<br>football<br>rankings<br>announced by<br>fifa | The latest FIFA football rankings, announced in 2018,<br>have ranked France as the top team, after winning the<br>2018 World Cup. This change was implemented using<br>the Elo rating system, and the rankings were<br>introduced to better reflect football team strengths. | France              | Germany         |
+
+are concise, they largely fail to capture key information relevant to answering the question. In contrast, RECOMP demonstrates better summarization capability but is prone to being overwhelmed by lengthy documents, resulting in misjudgments and even generating misleading information—such as the statement in Table [4:](#page-14-0) "*The U.S. stopped drafting for the Vietnam War after the Selective Service System was officially abolished in December 1972*"—which leads the downstream LLM to produce the incorrect answer "1972". Our method, CORE, accurately extracts answer-critical information from lengthy documents, exemplified by the summary: "*The U.S. stopped drafting for the Vietnam War in 1973 after announcing the decision by Secretary of Defense Melvin Laird earlier that year*", thereby enabling the LLM to generate the correct answer "1973". This indicates that our compressor, trained with an end-task target-oriented reinforcement learning optimization strategy, can produce document summaries that are most helpful for answering the given question while effectively filtering out irrelevant information.
+
+To further understand the limitations of our approach, we present two failure cases from the NQ dataset where the model provided incorrect answers based on our generated summaries. As shown in Table [6,](#page-16-0) the first case reveals that although the summary contained the key information required for the correct answer, the downstream LLM still produced an error, potentially due to its over-reliance on parametric knowledge. In the second case, the summary itself omitted critical information needed to answer the question, which likely led to the incorrect response.
+
+## D IMPACT OF DIFFERENT COMPRESSORS ON PERFORMANCE
+
+In our previous experiments, we employed Qwen2.5-1.5B as the initial model to train our compressor. In this section, we utilize two additional models—Llama3.2-1B and Llama3.2-3B—as starting points to train our compressor and the baseline compressor, respectively. The experimental results are presented in Table [7](#page-17-0) and Table [8.](#page-17-1) As shown in the results, our method CORE continues to achieve lossless compression with both models, maintaining a high token compression ratio while exhibiting no performance degradation in terms of Exact Match (EM) and F1 score compared to uncompressed RAG. Furthermore, under both new model configurations, our approach consistently outperforms the baseline methods, indicating that its superiority is not dependent on a specific model architecture and thus demonstrates strong robustness.
+
+We also observe that our method adheres to a form of scaling law: the compressor trained using the 3B model outperforms the one trained with the 1B model. Specifically, the 1B compressor improves performance by 1–4 EM points over the uncompressed baseline, while the 3B compressor yields gains of 3–9 EM points.
+
+{17}------------------------------------------------
+
+<span id="page-17-0"></span>Table 7: Open-domain QA results using Qwen2.5-14B-Instruct as the downstream LLM (M). The reported token counts represent the length of in-context documents, excluding few-shot examples. RECOMP and our method CORE are both trained using **llama3.2-1B-Instruct**.
+
+|                      | NQ         |             |         |           | TriviaQA   |            |        | HotpotQA |       |        | 2WikiMultihopQA |       |  |
+|----------------------|------------|-------------|---------|-----------|------------|------------|--------|----------|-------|--------|-----------------|-------|--|
+|                      | EM         | F1          | # tok   | EM        | F1         | # tok      | EM     | F1       | # tok | EM     | F1              | # tok |  |
+| No Retrieval         | 0.2136     | 0.3097      | 0       | 0.5323    | 0.5998     | 0          | 0.2105 | 0.2948   | 0     | 0.2611 | 0.2951          | 0     |  |
+| RAG without compre   | ession     |             |         |           |            |            |        |          |       |        |                 |       |  |
+| Top1 Document        | 0.3446     | 0.4441      | 142     | 0.6082    | 0.6870     | 143        | 0.2920 | 0.3893   | 147   | 0.2679 | 0.3187          | 153   |  |
+| Top3 Documents       | 0.3778     | 0.4845      | 427     | 0.6260    | 0.7102     | 430        | 0.3164 | 0.4187   | 442   | 0.2789 | 0.3358          | 460   |  |
+| Top5 Documents       | 0.3803     | 0.4916      | 712     | 0.6410    | 0.7248     | 715        | 0.3299 | 0.4369   | 737   | 0.2964 | 0.3521          | 766   |  |
+| Top10 Documents      | 0.3867     | 0.5003      | 1425    | 0.6440    | 0.7292     | 1428       | 0.3395 | 0.4488   | 1471  | 0.3104 | 0.3675          | 1531  |  |
+| Compression of top 3 | 5 docs     |             |         |           |            |            |        |          |       |        |                 |       |  |
+| llama3.2-1B          | 0.3147     | 0.4227      | 64      | 0.5552    | 0.6415     | 60         | 0.2648 | 0.3639   | 58    | 0.2498 | 0.3003          | 61    |  |
+| Deepseek-V3 (671B)   | 0.3773     | 0.5039      | 54      | 0.6528    | 0.7433     | 51         | 0.3359 | 0.4483   | 48    | 0.2507 | 0.3031          | 45    |  |
+| RECOMP (1B)          | 0.3410     | 0.4655      | 57      | 0.6071    | 0.6880     | 48         | 0.2987 | 0.4121   | 49    | 0.3045 | 0.3653          | 33    |  |
+| CORE (1B)            | 0.3947     | 0.4923      | 47      | 0.6483    | 0.7287     | 43         | 0.3344 | 0.4454   | 45    | 0.3378 | 0.3969          | 34    |  |
+| Compression of top   | 10 docs (w | vith the co | mpresso | r trained | on top 5 d | locs)      |        |          |       |        |                 |       |  |
+| llama3.2-1B          | 0.3141     | 0.4228      | 62      | 0.5651    | 0.6512     | <b>5</b> 8 | 0.2663 | 0.3661   | 56    | 0.2493 | 0.3006          | 61    |  |
+| Deepseek-V3 (671B)   | 0.3779     | 0.5107      | 56      | 0.6529    | 0.7445     | 53         | 0.3462 | 0.4569   | 50    | 0.2900 | 0.3464          | 40    |  |
+| RECOMP (1B)          | 0.3421     | 0.4661      | 59      | 0.6095    | 0.6917     | 52         | 0.2982 | 0.4105   | 55    | 0.3072 | 0.3681          | 44    |  |
+| CORE (1B)            | 0.4033     | 0.5033      | 47      | 0.6521    | 0.7296     | 45         | 0.3412 | 0.4500   | 48    | 0.3586 | 0.4162          | 42    |  |
+
+<span id="page-17-1"></span>Table 8: Open-domain QA results using Qwen2.5-14B-Instruct as the downstream LLM (M). The reported token counts represent the length of in-context documents, excluding few-shot examples. RECOMP and our method CORE are both trained using **llama3.2-3B-Instruct**.
+
+|                      |            | NQ          |         |            | TriviaQA   |       |        | HotpotQA |       |        | 2WikiMultihopQA |       |  |
+|----------------------|------------|-------------|---------|------------|------------|-------|--------|----------|-------|--------|-----------------|-------|--|
+|                      | EM         | F1          | # tok   | EM         | F1         | # tok | EM     | F1       | # tok | EM     | F1              | # tok |  |
+| No Retrieval         | 0.2136     | 0.3097      | 0       | 0.5323     | 0.5998     | 0     | 0.2105 | 0.2948   | 0     | 0.2611 | 0.2951          | 0     |  |
+| RAG without compre   | ession     |             |         |            |            |       |        |          |       |        |                 |       |  |
+| Top1 Document        | 0.3446     | 0.4441      | 142     | 0.6082     | 0.6870     | 143   | 0.2920 | 0.3893   | 147   | 0.2679 | 0.3187          | 153   |  |
+| Top3 Documents       | 0.3778     | 0.4845      | 427     | 0.6260     | 0.7102     | 430   | 0.3164 | 0.4187   | 442   | 0.2789 | 0.3358          | 460   |  |
+| Top5 Documents       | 0.3803     | 0.4916      | 712     | 0.6410     | 0.7248     | 715   | 0.3299 | 0.4369   | 737   | 0.2964 | 0.3521          | 766   |  |
+| Top10 Documents      | 0.3867     | 0.5003      | 1425    | 0.6440     | 0.7292     | 1428  | 0.3395 | 0.4488   | 1471  | 0.3104 | 0.3675          | 1531  |  |
+| Compression of top : | 5 docs     |             |         |            |            |       |        |          |       |        |                 |       |  |
+| llama3.2-3B          | 0.3252     | 0.4334      | 60      | 0.5650     | 0.6521     | 59    | 0.2772 | 0.3809   | 58    | 0.2485 | 0.2995          | 60    |  |
+| Deepseek-V3 (671B)   | 0.3773     | 0.5039      | 54      | 0.6528     | 0.7433     | 51    | 0.3359 | 0.4483   | 48    | 0.2507 | 0.3031          | 45    |  |
+| RECOMP (3B)          | 0.3657     | 0.4912      | 55      | 0.6183     | 0.6920     | 47    | 0.3025 | 0.4238   | 52    | 0.3274 | 0.3806          | 42    |  |
+| CORE (3B)            | 0.4310     | 0.5234      | 32      | 0.6650     | 0.7306     | 38    | 0.3507 | 0.4736   | 51    | 0.3905 | 0.4474          | 40    |  |
+| Compression of top   | 10 docs (w | vith the co | mpresso | or trained | on top 5 a | locs) |        |          |       |        |                 |       |  |
+| llama3.2-3B          | 0.3318     | 0.4359      | 61      | 0.5720     | 0.6588     | 57    | 0.2791 | 0.3854   | 60    | 0.2491 | 0.3011          | 59    |  |
+| Deepseek-V3 (671B)   | 0.3779     | 0.5107      | 56      | 0.6529     | 0.7445     | 53    | 0.3462 | 0.4569   | 50    | 0.2900 | 0.3464          | 40    |  |
+| RECOMP (3B)          | 0.3682     | 0.4963      | 52      | 0.6205     | 0.6973     | 44    | 0.3077 | 0.4261   | 54    | 0.3312 | 0.3869          | 50    |  |
+| CORE (3B)            | 0.4526     | 0.5467      | 33      | 0.6736     | 0.7404     | 37    | 0.3693 | 0.4926   | 51    | 0.4071 | 0.4633          | 48    |  |
+
+{18}------------------------------------------------
+
+<span id="page-18-1"></span>**979**
+
+**994**
+
+<span id="page-18-0"></span>Table 9: Zero-Shot Evaluation on HotpotQA of Models Trained on Natural Questions.
+
+|                 | EM    | F1    | #tok |
+|-----------------|-------|-------|------|
+| No Retrieval    | 21.05 | 29.48 | 0    |
+| Full Documents  | 32.99 | 43.69 | 737  |
+| BM25            | 24.18 | 35.73 | 71   |
+| NoiseFilter-IB  | 27.97 | 38.62 | 38   |
+| RECOMP          | 28.96 | 39.95 | 56   |
+| CORE            | 33.67 | 45.06 | 36   |
+| RECOMP-Transfer | 26.68 | 37.29 | 58   |
+| CORE-Transfer   | 31.25 | 42.84 | 35   |
+|                 |       |       |      |
+
+Table 10: Evaluation on Noisy Natural Questions.
+
+|                | EM    | F1    | #tok |
+|----------------|-------|-------|------|
+| full documents | 35.21 | 45.38 | 1427 |
+| RECOMP         | 33.29 | 43.90 | 59   |
+| CORE           | 38.19 | 48.85 | 48   |
+
+## E CROSS-DATASET GENERALIZATION PERFORMANCE
+
+To verify the universal compression capability of our method, we directly transfer the model trained on the single-hop question answering dataset NQ to the multi-hop dataset HotpotQA for evaluation. The results, shown in Table [9,](#page-18-0) indicate that our transferred model achieves nearly lossless performance compared to using full documents without compression, while substantially outperforming the transfer results of the RECOMP baseline. Moreover, although both our method and the baseline underperform relative to models trained directly on the target HotpotQA dataset, our approach exhibits a smaller performance drop and demonstrates greater robustness compared to the baseline.
+
+## F ROBUSTNESS AGAINST NOISY CONTEXTS
+
+To evaluate the robustness of our approach against adversarial retrievals and noisy contexts, we constructed a noisy version of the NQ dataset. For each question, we constructed the input context by combining the top-3 passages retrieved by the DPR retriever with 7 randomly selected passages from the Wikipedia corpus to serve as irrelevant/noisy information. This resulted in a context of 10 passages, which were then shuffled to randomize the order. We then compared the performance of our method against the full-document baseline. Experimental results are presented in the table [10.](#page-18-1) In the "full documents" setting, the downstream LLM directly uses all these 10 passages to answer the question, whereas in our method, the compressor first summarizes the context, and the LLM then generates an answer based on the compressed content. The model we used was trained on the standard NQ dataset without any such noise augmentation. Our method not only matches but slightly surpasses the performance of using all documents, demonstrating its strong noise resistance and ability to extract key information from cluttered contexts. In addition, we compared our approach with the RECOMP baseline, and our method consistently outperforms it, reaffirming the superior compression capability and robustness of our model. Furthermore, our method achieves a high compression rate, condensing the source content from 1,427 tokens to just 48.

iclr26/6FyYLNnDL6/parse_report.json

@@ -0,0 +1,76 @@
+{
+  "paper_id": "6FyYLNnDL6",
+  "pipeline": "marker_non_llm_v3",
+  "parser": "marker_single",
+  "formats": [
+    "markdown",
+    "chunks"
+  ],
+  "llm_enabled": false,
+  "pdf_path": "/network/scratch/j/jianan.zhao/ReviewAgent/data/processed_papers/iclr26_20260429_1952_all/raw/6FyYLNnDL6.pdf",
+  "pdf_sha256": "cd0d01f4ebcf5121f4548f67afae4c5f0961192feefd83de0506d1ab28cd3933",
+  "bytes": 2288371,
+  "source": "https://openreview.net/pdf?id=6FyYLNnDL6",
+  "page_count": 19,
+  "ok": true,
+  "elapsed_seconds": 223.47,
+  "paper2markdown_v3": {
+    "paper_id": "6FyYLNnDL6",
+    "pipeline": "Paper2Markdown-V3",
+    "ok": true,
+    "page_count": 19,
+    "chunk_count": 106,
+    "main_body_chunk_count": 60,
+    "appendix_chunk_count": 38,
+    "reference_chunk_count": 8,
+    "model_text_chars": 43162,
+    "raw_markdown_chars": 88191,
+    "sanitized_chars": 79686,
+    "page_provenance": {
+      "min_page": 1,
+      "max_page": 19,
+      "invalid_count": 0
+    },
+    "marker_block_type_counts": {
+      "Caption": 2,
+      "Equation": 4,
+      "Figure": 1,
+      "FigureGroup": 4,
+      "Footnote": 2,
+      "ListGroup": 3,
+      "PageFooter": 19,
+      "PageHeader": 19,
+      "SectionHeader": 37,
+      "Table": 3,
+      "TableGroup": 7,
+      "Text": 330
+    },
+    "asset_count_raw": 5,
+    "asset_count_model_kept": 5,
+    "asset_count_rejected": 0,
+    "asset_reject_reasons": {
+      "kept": 5
+    },
+    "artifact_leak_audit": {
+      "ok": true,
+      "hits": {
+        "Anonymous Authors": [],
+        "ACKNOWLEDGMENT": [],
+        "OpenReview": [],
+        "\"accept_label\"": [],
+        "\"decision\"": [],
+        "\"decision_tier\"": [],
+        "\"source_status\"": [],
+        "Meta-review": [],
+        "Official Review": [],
+        "official_reviews": [],
+        "meta_reviews": [],
+        "suggested_verdict_score": []
+      },
+      "artifact_count": 2
+    },
+    "default_model_input": "model_text_v3.txt",
+    "appendix_input": "appendix_text_v3.txt",
+    "reference_input": "reference_text_v3.txt"
+  }
+}

iclr26/6FyYLNnDL6/reference_chunks.jsonl

@@ -0,0 +1,8 @@
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0060", "section": "REFERENCES", "page_start": 11, "page_end": 11, "type": "ListGroup", "text": "Zhiwei Cao, Qian Cao, Yu Lu, Ningxin Peng, Luyang Huang, Shanbo Cheng, and Jinsong Su. Retaining key information under high compression ratios: Query-guided compressor for llms. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pp. 12685–12695, 2024. Mingyang Chen, Tianpeng Li, Haoze Sun, Yijie Zhou, Chenzheng Zhu, Haofen Wang, Jeff Z Pan, Wen Zhang, Huajun Chen, Fan Yang, et al. Learning to reason with search for llms via reinforcement learning. arXiv preprint arXiv:2503.19470 , 2025a. Yiqun Chen, Lingyong Yan, Weiwei Sun, Xinyu Ma, Yi Zhang, Shuaiqiang Wang, Dawei Yin, Yiming Yang, and Jiaxin Mao. Improving retrieval-augmented generation through multi-agent reinforcement learning. arXiv preprint arXiv:2501.15228 , 2025b. Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat-Seng Chua, and Qing Li. A survey on rag meeting llms: Towards retrieval-augmented large language models. In Proceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining , pp. 6491–6501, 2024. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948 , 2025. Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa. Constructing a multi-hop qa dataset for comprehensive evaluation of reasoning steps. arXiv preprint arXiv:2011.01060 , 2020. Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. Unsupervised dense information retrieval with contrastive learning. arXiv preprint arXiv:2112.09118 , 2021. Huiqiang Jiang, Qianhui Wu, Xufang Luo, Dongsheng Li, Chin-Yew Lin, Yuqing Yang, and Lili Qiu. Longllmlingua: Accelerating and enhancing llms in long context scenarios via prompt compression. In Proceedings of the 62nd Annual Meeting of the Association for Computational Lin guistics (Volume 1: Long Papers) , pp. 1658–1677, 2024. Bowen Jin, Jinsung Yoon, Jiawei Han, and Sercan O Arik. Long-context llms meet rag: Overcoming challenges for long inputs in rag. arXiv preprint arXiv:2410.05983 , 2024a. Jiajie Jin, Yutao Zhu, Yujia Zhou, and Zhicheng Dou. Bider: Bridging knowledge inconsistency for efficient retrieval-augmented llms via key supporting evidence. In Findings of the Association for Computational Linguistics ACL 2024 , pp. 750–761, 2024b. Mandar Joshi, Eunsol Choi, Daniel S Weld, and Luke Zettlemoyer. Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension. arXiv preprint arXiv:1705.03551 , 2017.", "source": "marker_v2", "marker_block_id": "/page/10/ListGroup/318"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0061", "section": "REFERENCES", "page_start": 12, "page_end": 12, "type": "ListGroup", "text": "Hoyoun Jung and Kyung-Joong Kim. Discrete prompt compression with reinforcement learning. IEEE Access , 12:72578–72587, 2024. Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick SH Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. Dense passage retrieval for open-domain question answering. In EMNLP (1) , pp. 6769–6781, 2020. Zixuan Ke, Weize Kong, Cheng Li, Mingyang Zhang, Qiaozhu Mei, and Michael Bendersky. Bridging the preference gap between retrievers and llms. arXiv preprint arXiv:2401.06954 , 2024. Sangryul Kim and James Thorne. Context filtering with reward modeling in question answering. In Proceedings of the 31st International Conference on Computational Linguistics , pp. 11048– 11055, 2025. Mandar Kulkarni, Praveen Tangarajan, Kyung Kim, and Anusua Trivedi. Reinforcement learning for optimizing rag for domain chatbots. arXiv preprint arXiv:2401.06800 , 2024. Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, et al. Natural questions: a benchmark for question answering research. Transactions of the Association for Computational Linguistics , 7:453–466, 2019. Sha Li and Naren Ramakrishnan. Oreo: A plug-in context reconstructor to enhance retrievalaugmented generation. In Proceedings of the 2025 International ACM SIGIR Conference on Innovative Concepts and Theories in Information Retrieval (ICTIR) , pp. 238–253, 2025. Xiaoxi Li, Jiajie Jin, Guanting Dong, Hongjin Qian, Yutao Zhu, Yongkang Wu, Ji-Rong Wen, and Zhicheng Dou. Webthinker: Empowering large reasoning models with deep research capability. arXiv preprint arXiv:2504.21776 , 2025. Xinze Li, Sen Mei, Zhenghao Liu, Yukun Yan, Shuo Wang, Shi Yu, Zheni Zeng, Hao Chen, Ge Yu, Zhiyuan Liu, et al. Rag-ddr: Optimizing retrieval-augmented generation using differentiable data rewards. arXiv preprint arXiv:2410.13509 , 2024a. Zhonghao Li, Xuming Hu, Aiwei Liu, Kening Zheng, Sirui Huang, and Hui Xiong. Refiner: Restructure retrieval content efficiently to advance question-answering capabilities. arXiv preprint arXiv:2406.11357 , 2024b. Xi Victoria Lin, Xilun Chen, Mingda Chen, Weijia Shi, Maria Lomeli, Richard James, Pedro Rodriguez, Jacob Kahn, Gergely Szilvasy, Mike Lewis, et al. Ra-dit: Retrieval-augmented dual instruction tuning. In The Twelfth International Conference on Learning Representations , 2023. Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437 , 2024. Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. Lost in the middle: How language models use long contexts. arXiv preprint arXiv:2307.03172 , 2023. Maxime Louis, Herve D´ ejean, and St ´ ephane Clinchant. Pisco: Pretty simple compression for ´ retrieval-augmented generation. arXiv preprint arXiv:2501.16075 , 2025. Shengyu Mao, Yong Jiang, Boli Chen, Xiao Li, Peng Wang, Xinyu Wang, Pengjun Xie, Fei Huang, Huajun Chen, and Ningyu Zhang. Rafe: Ranking feedback improves query rewriting for rag. arXiv preprint arXiv:2405.14431 , 2024. Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh Tiwary, Rangan Majumder, and Li Deng. Ms marco: A human-generated machine reading comprehension dataset. 2016. Ori Ram, Yoav Levine, Itay Dalmedigos, Dor Muhlgay, Amnon Shashua, Kevin Leyton-Brown, and Yoav Shoham. In-context retrieval-augmented language models. Transactions of the Association for Computational Linguistics , 11:1316–1331, 2023.", "source": "marker_v2", "marker_block_id": "/page/11/ListGroup/346"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0062", "section": "REFERENCES", "page_start": 13, "page_end": 13, "type": "ListGroup", "text": "David Rau, Shuai Wang, Herve D´ ejean, and St ´ ephane Clinchant. Context embeddings for efficient ´ answer generation in rag. arXiv preprint arXiv:2407.09252 , 2024. Stephen E Robertson, Steve Walker, Susan Jones, Micheline M Hancock-Beaulieu, Mike Gatford, et al. Okapi at TREC-3 . British Library Research and Development Department, 1995. Shivam Shandilya, Menglin Xia, Supriyo Ghosh, Huiqiang Jiang, Jue Zhang, Qianhui Wu, Victor Ruhle, and Saravan Rajmohan. Taco-rl: Task aware prompt compression optimization with rein- ¨ forcement learning. In Findings of the Association for Computational Linguistics: ACL 2025 , pp. 1582–1597, 2025. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300 , 2024. Weijia Shi, Sewon Min, Michihiro Yasunaga, Minjoon Seo, Rich James, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. Replug: Retrieval-augmented black-box language models. arXiv preprint arXiv:2301.12652 , 2023. Aditi Singh, Abul Ehtesham, Saket Kumar, and Tala Talaei Khoei. Agentic retrieval-augmented generation: A survey on agentic rag. arXiv preprint arXiv:2501.09136 , 2025. Huatong Song, Jinhao Jiang, Yingqian Min, Jie Chen, Zhipeng Chen, Wayne Xin Zhao, Lei Fang, and Ji-Rong Wen. R1-searcher: Incentivizing the search capability in llms via reinforcement learning. arXiv preprint arXiv:2503.05592 , 2025. Chenyuan Wu, Ninglu Shao, Zheng Liu, Shitao Xiao, Chaozhuo Li, Chen Zhang, Senzhang Wang, and Defu Lian. Lighter and better: Towards flexible context adaptation for retrieval augmented generation. In Proceedings of the Eighteenth ACM International Conference on Web Search and Data Mining , pp. 271–280, 2025. Fangyuan Xu, Weijia Shi, and Eunsol Choi. Recomp: Improving retrieval-augmented lms with context compression and selective augmentation. In The Twelfth International Conference on Learning Representations , 2024. Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W Cohen, Ruslan Salakhutdinov, and Christopher D Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question answering. arXiv preprint arXiv:1809.09600 , 2018. Qianchi Zhang, Hainan Zhang, Liang Pang, Hongwei Zheng, and Zhiming Zheng. Adacomp: Extractive context compression with adaptive predictor for retrieval-augmented large language models. arXiv preprint arXiv:2409.01579 , 2024a. Taolin Zhang, Dongyang Li, Qizhou Chen, Chengyu Wang, Longtao Huang, Hui Xue, Xiaofeng He, and Jun Huang. R 4: Reinforced retriever-reorder-responder for retrieval-augmented large language models. In ECAI 2024 , pp. 2314–2321. IOS Press, 2024b. Yuxiang Zheng, Dayuan Fu, Xiangkun Hu, Xiaojie Cai, Lyumanshan Ye, Pengrui Lu, and Pengfei Liu. Deepresearcher: Scaling deep research via reinforcement learning in real-world environments. arXiv preprint arXiv:2504.03160 , 2025. Kun Zhu, Xiaocheng Feng, Xiyuan Du, Yuxuan Gu, Weijiang Yu, Haotian Wang, Qianglong Chen, Zheng Chu, Jingchang Chen, and Bing Qin. An information bottleneck perspective for effective noise filtering on retrieval-augmented generation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pp. 1044–1069, 2024.", "source": "marker_v2", "marker_block_id": "/page/12/ListGroup/327"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0063", "section": "REFERENCES", "page_start": 14, "page_end": 14, "type": "Text", "text": "Compress the information in the retrieved documents into a 2-sentence summary that could be used to answer the question. If the documents do not contain relevant information, simply output \" \".", "source": "marker_v2", "marker_block_id": "/page/13/Text/1"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0064", "section": "REFERENCES", "page_start": 14, "page_end": 14, "type": "Text", "text": "Question: { Question }", "source": "marker_v2", "marker_block_id": "/page/13/Text/2"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0065", "section": "REFERENCES", "page_start": 14, "page_end": 14, "type": "Text", "text": "Retrieved documents: { Documents }", "source": "marker_v2", "marker_block_id": "/page/13/Text/3"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0066", "section": "REFERENCES", "page_start": 14, "page_end": 14, "type": "Text", "text": "Compressed documents:", "source": "marker_v2", "marker_block_id": "/page/13/Text/4"}
+{"paper_id": "6FyYLNnDL6", "chunk_id": "6FyYLNnDL6:0067", "section": "REFERENCES", "page_start": 14, "page_end": 14, "type": "Caption", "text": "Figure 5: Prompt template used to generate the summary.", "source": "marker_v2", "marker_block_id": "/page/13/Caption/5"}

iclr26/6FyYLNnDL6/reference_text_v3.txt

@@ -0,0 +1,23 @@
+[p. 11 | section: REFERENCES | type: ListGroup]
+Zhiwei Cao, Qian Cao, Yu Lu, Ningxin Peng, Luyang Huang, Shanbo Cheng, and Jinsong Su. Retaining key information under high compression ratios: Query-guided compressor for llms. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pp. 12685–12695, 2024. Mingyang Chen, Tianpeng Li, Haoze Sun, Yijie Zhou, Chenzheng Zhu, Haofen Wang, Jeff Z Pan, Wen Zhang, Huajun Chen, Fan Yang, et al. Learning to reason with search for llms via reinforcement learning. arXiv preprint arXiv:2503.19470 , 2025a. Yiqun Chen, Lingyong Yan, Weiwei Sun, Xinyu Ma, Yi Zhang, Shuaiqiang Wang, Dawei Yin, Yiming Yang, and Jiaxin Mao. Improving retrieval-augmented generation through multi-agent reinforcement learning. arXiv preprint arXiv:2501.15228 , 2025b. Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat-Seng Chua, and Qing Li. A survey on rag meeting llms: Towards retrieval-augmented large language models. In Proceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining , pp. 6491–6501, 2024. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948 , 2025. Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa. Constructing a multi-hop qa dataset for comprehensive evaluation of reasoning steps. arXiv preprint arXiv:2011.01060 , 2020. Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. Unsupervised dense information retrieval with contrastive learning. arXiv preprint arXiv:2112.09118 , 2021. Huiqiang Jiang, Qianhui Wu, Xufang Luo, Dongsheng Li, Chin-Yew Lin, Yuqing Yang, and Lili Qiu. Longllmlingua: Accelerating and enhancing llms in long context scenarios via prompt compression. In Proceedings of the 62nd Annual Meeting of the Association for Computational Lin guistics (Volume 1: Long Papers) , pp. 1658–1677, 2024. Bowen Jin, Jinsung Yoon, Jiawei Han, and Sercan O Arik. Long-context llms meet rag: Overcoming challenges for long inputs in rag. arXiv preprint arXiv:2410.05983 , 2024a. Jiajie Jin, Yutao Zhu, Yujia Zhou, and Zhicheng Dou. Bider: Bridging knowledge inconsistency for efficient retrieval-augmented llms via key supporting evidence. In Findings of the Association for Computational Linguistics ACL 2024 , pp. 750–761, 2024b. Mandar Joshi, Eunsol Choi, Daniel S Weld, and Luke Zettlemoyer. Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension. arXiv preprint arXiv:1705.03551 , 2017.
+
+[p. 12 | section: REFERENCES | type: ListGroup]
+Hoyoun Jung and Kyung-Joong Kim. Discrete prompt compression with reinforcement learning. IEEE Access , 12:72578–72587, 2024. Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick SH Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. Dense passage retrieval for open-domain question answering. In EMNLP (1) , pp. 6769–6781, 2020. Zixuan Ke, Weize Kong, Cheng Li, Mingyang Zhang, Qiaozhu Mei, and Michael Bendersky. Bridging the preference gap between retrievers and llms. arXiv preprint arXiv:2401.06954 , 2024. Sangryul Kim and James Thorne. Context filtering with reward modeling in question answering. In Proceedings of the 31st International Conference on Computational Linguistics , pp. 11048– 11055, 2025. Mandar Kulkarni, Praveen Tangarajan, Kyung Kim, and Anusua Trivedi. Reinforcement learning for optimizing rag for domain chatbots. arXiv preprint arXiv:2401.06800 , 2024. Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, et al. Natural questions: a benchmark for question answering research. Transactions of the Association for Computational Linguistics , 7:453–466, 2019. Sha Li and Naren Ramakrishnan. Oreo: A plug-in context reconstructor to enhance retrievalaugmented generation. In Proceedings of the 2025 International ACM SIGIR Conference on Innovative Concepts and Theories in Information Retrieval (ICTIR) , pp. 238–253, 2025. Xiaoxi Li, Jiajie Jin, Guanting Dong, Hongjin Qian, Yutao Zhu, Yongkang Wu, Ji-Rong Wen, and Zhicheng Dou. Webthinker: Empowering large reasoning models with deep research capability. arXiv preprint arXiv:2504.21776 , 2025. Xinze Li, Sen Mei, Zhenghao Liu, Yukun Yan, Shuo Wang, Shi Yu, Zheni Zeng, Hao Chen, Ge Yu, Zhiyuan Liu, et al. Rag-ddr: Optimizing retrieval-augmented generation using differentiable data rewards. arXiv preprint arXiv:2410.13509 , 2024a. Zhonghao Li, Xuming Hu, Aiwei Liu, Kening Zheng, Sirui Huang, and Hui Xiong. Refiner: Restructure retrieval content efficiently to advance question-answering capabilities. arXiv preprint arXiv:2406.11357 , 2024b. Xi Victoria Lin, Xilun Chen, Mingda Chen, Weijia Shi, Maria Lomeli, Richard James, Pedro Rodriguez, Jacob Kahn, Gergely Szilvasy, Mike Lewis, et al. Ra-dit: Retrieval-augmented dual instruction tuning. In The Twelfth International Conference on Learning Representations , 2023. Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437 , 2024. Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. Lost in the middle: How language models use long contexts. arXiv preprint arXiv:2307.03172 , 2023. Maxime Louis, Herve D´ ejean, and St ´ ephane Clinchant. Pisco: Pretty simple compression for ´ retrieval-augmented generation. arXiv preprint arXiv:2501.16075 , 2025. Shengyu Mao, Yong Jiang, Boli Chen, Xiao Li, Peng Wang, Xinyu Wang, Pengjun Xie, Fei Huang, Huajun Chen, and Ningyu Zhang. Rafe: Ranking feedback improves query rewriting for rag. arXiv preprint arXiv:2405.14431 , 2024. Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh Tiwary, Rangan Majumder, and Li Deng. Ms marco: A human-generated machine reading comprehension dataset. 2016. Ori Ram, Yoav Levine, Itay Dalmedigos, Dor Muhlgay, Amnon Shashua, Kevin Leyton-Brown, and Yoav Shoham. In-context retrieval-augmented language models. Transactions of the Association for Computational Linguistics , 11:1316–1331, 2023.
+
+[p. 13 | section: REFERENCES | type: ListGroup]
+David Rau, Shuai Wang, Herve D´ ejean, and St ´ ephane Clinchant. Context embeddings for efficient ´ answer generation in rag. arXiv preprint arXiv:2407.09252 , 2024. Stephen E Robertson, Steve Walker, Susan Jones, Micheline M Hancock-Beaulieu, Mike Gatford, et al. Okapi at TREC-3 . British Library Research and Development Department, 1995. Shivam Shandilya, Menglin Xia, Supriyo Ghosh, Huiqiang Jiang, Jue Zhang, Qianhui Wu, Victor Ruhle, and Saravan Rajmohan. Taco-rl: Task aware prompt compression optimization with rein- ¨ forcement learning. In Findings of the Association for Computational Linguistics: ACL 2025 , pp. 1582–1597, 2025. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300 , 2024. Weijia Shi, Sewon Min, Michihiro Yasunaga, Minjoon Seo, Rich James, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. Replug: Retrieval-augmented black-box language models. arXiv preprint arXiv:2301.12652 , 2023. Aditi Singh, Abul Ehtesham, Saket Kumar, and Tala Talaei Khoei. Agentic retrieval-augmented generation: A survey on agentic rag. arXiv preprint arXiv:2501.09136 , 2025. Huatong Song, Jinhao Jiang, Yingqian Min, Jie Chen, Zhipeng Chen, Wayne Xin Zhao, Lei Fang, and Ji-Rong Wen. R1-searcher: Incentivizing the search capability in llms via reinforcement learning. arXiv preprint arXiv:2503.05592 , 2025. Chenyuan Wu, Ninglu Shao, Zheng Liu, Shitao Xiao, Chaozhuo Li, Chen Zhang, Senzhang Wang, and Defu Lian. Lighter and better: Towards flexible context adaptation for retrieval augmented generation. In Proceedings of the Eighteenth ACM International Conference on Web Search and Data Mining , pp. 271–280, 2025. Fangyuan Xu, Weijia Shi, and Eunsol Choi. Recomp: Improving retrieval-augmented lms with context compression and selective augmentation. In The Twelfth International Conference on Learning Representations , 2024. Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W Cohen, Ruslan Salakhutdinov, and Christopher D Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question answering. arXiv preprint arXiv:1809.09600 , 2018. Qianchi Zhang, Hainan Zhang, Liang Pang, Hongwei Zheng, and Zhiming Zheng. Adacomp: Extractive context compression with adaptive predictor for retrieval-augmented large language models. arXiv preprint arXiv:2409.01579 , 2024a. Taolin Zhang, Dongyang Li, Qizhou Chen, Chengyu Wang, Longtao Huang, Hui Xue, Xiaofeng He, and Jun Huang. R 4: Reinforced retriever-reorder-responder for retrieval-augmented large language models. In ECAI 2024 , pp. 2314–2321. IOS Press, 2024b. Yuxiang Zheng, Dayuan Fu, Xiangkun Hu, Xiaojie Cai, Lyumanshan Ye, Pengrui Lu, and Pengfei Liu. Deepresearcher: Scaling deep research via reinforcement learning in real-world environments. arXiv preprint arXiv:2504.03160 , 2025. Kun Zhu, Xiaocheng Feng, Xiyuan Du, Yuxuan Gu, Weijiang Yu, Haotian Wang, Qianglong Chen, Zheng Chu, Jingchang Chen, and Bing Qin. An information bottleneck perspective for effective noise filtering on retrieval-augmented generation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pp. 1044–1069, 2024.
+
+[p. 14 | section: REFERENCES | type: Text]
+Compress the information in the retrieved documents into a 2-sentence summary that could be used to answer the question. If the documents do not contain relevant information, simply output " ".
+
+[p. 14 | section: REFERENCES | type: Text]
+Question: { Question }
+
+[p. 14 | section: REFERENCES | type: Text]
+Retrieved documents: { Documents }
+
+[p. 14 | section: REFERENCES | type: Text]
+Compressed documents:
+
+[p. 14 | section: REFERENCES | type: Caption]
+Figure 5: Prompt template used to generate the summary.

iclr26/6FyYLNnDL6/sanitization_report.json

@@ -0,0 +1,59 @@
+{
+  "paper_id": "6FyYLNnDL6",
+  "pipeline": "Paper2Markdown-V3",
+  "ok": true,
+  "page_count": 19,
+  "chunk_count": 106,
+  "main_body_chunk_count": 60,
+  "appendix_chunk_count": 38,
+  "reference_chunk_count": 8,
+  "model_text_chars": 43162,
+  "raw_markdown_chars": 88191,
+  "sanitized_chars": 79686,
+  "page_provenance": {
+    "min_page": 1,
+    "max_page": 19,
+    "invalid_count": 0
+  },
+  "marker_block_type_counts": {
+    "Caption": 2,
+    "Equation": 4,
+    "Figure": 1,
+    "FigureGroup": 4,
+    "Footnote": 2,
+    "ListGroup": 3,
+    "PageFooter": 19,
+    "PageHeader": 19,
+    "SectionHeader": 37,
+    "Table": 3,
+    "TableGroup": 7,
+    "Text": 330
+  },
+  "asset_count_raw": 5,
+  "asset_count_model_kept": 5,
+  "asset_count_rejected": 0,
+  "asset_reject_reasons": {
+    "kept": 5
+  },
+  "artifact_leak_audit": {
+    "ok": true,
+    "hits": {
+      "Anonymous Authors": [],
+      "ACKNOWLEDGMENT": [],
+      "OpenReview": [],
+      "\"accept_label\"": [],
+      "\"decision\"": [],
+      "\"decision_tier\"": [],
+      "\"source_status\"": [],
+      "Meta-review": [],
+      "Official Review": [],
+      "official_reviews": [],
+      "meta_reviews": [],
+      "suggested_verdict_score": []
+    },
+    "artifact_count": 2
+  },
+  "default_model_input": "model_text_v3.txt",
+  "appendix_input": "appendix_text_v3.txt",
+  "reference_input": "reference_text_v3.txt"
+}

iclr26/6FyYLNnDL6/sanitized_v3.txt

@@ -0,0 +1,377 @@
+{0}
+## ABSTRACT
+Retrieval-Augmented Generation (RAG) has emerged as a promising approach to enhance the timeliness of knowledge updates and the factual accuracy of responses in large language models. However, incorporating a large number of retrieved documents significantly increases input length, leading to higher computational costs. Existing approaches to document compression tailored for RAG often degrade task performance, as they typically rely on predefined heuristics in the absence of clear compression guidelines. These heuristics fail to ensure that the compressed content effectively supports downstream tasks. To address these limitations, we propose CORE, a novel method for lossless context compression in RAG. CORE is optimized end-to-end and does not depend on predefined compression labels, which are often impractical to obtain. Instead, it leverages downstream task performance as a feedback signal, iteratively refining the compression policy to enhance task effectiveness. Extensive experiments across four datasets demonstrate the effectiveness of CORE. With a high compression ratio of 3%, CORE not only prevents performance degradation compared to including full documents (i.e., without compression) but also improves the average Exact Match (EM) score by 3.3 points. The code for CORE is available at <
+## 1 INTRODUCTION
+Large language models (LLMs) have undergone rapid development in recent years, significantly enhancing performance across various language tasks due to their emergent capabilities in semantic understanding and reasoning. Nevertheless, LLMs still face challenges in updating knowledge and providing factual responses [\(Fan et al., 2024\)](#page-10-0). To address these issues, Retrieval-Augmented Generation (RAG) has emerged as a promising approach. By retrieving the most relevant documents from external knowledge bases and prepending them as contextual information to the original input, RAG substantially improves LLM performance on knowledge-intensive tasks [\(Ram et al., 2023\)](#page-11-0).
+While RAG enhances performance, its effectiveness is closely tied to the number of retrieved documents used, since a broader context increases the probability of encompassing critical evidence. As illustrated in Figure [1,](#page-1-0) performance was weakest without any retrieved documents (i.e., without RAG). Accuracy improved consistently as more documents were added to the context, ultimately exceeding the no-RAG baseline by over 10 Exact Match (EM) points. However, this performance gain came with two significant limitations: (1) a substantial increase in computational cost from processing a larger number of context tokens [\(Xu et al., 2024\)](#page-12-0), and (2) the model's difficulty in effectively leveraging all provided documents, often resulting in the omission of key information located in the middle of the context [\(Liu et al., 2023\)](#page-11-1).
+These limitations have motivated recent research efforts aimed at compressing the retrieved context [\(Jin et al., 2024b;](#page-10-1) [Wu et al., 2025;](#page-12-1) [Jin et al., 2024a;](#page-10-2) [Zhang et al., 2024a\)](#page-12-2). Prominent approaches include document summarization [\(Xu et al., 2024\)](#page-12-0), key information extraction [\(Cao et al., 2024;](#page-10-3) [Xu](#page-12-0) [et al., 2024\)](#page-12-0), the construction of key supporting evidence [\(Jin et al., 2024b\)](#page-10-1), and noise filtering based on information theory [\(Zhu et al., 2024\)](#page-12-3). Despite recent progress, these methods have several notable shortcomings. First, compression often results in a performance trade-off. For instance, RECOMP
+{1}------------------------------------------------
+**059**
+**061**
+**072 073 074**
+**079**
+**094**
+<span id="page-1-0"></span>![](_page_1_Figure_1.jpeg)
+![](_page_1_Figure_2.jpeg)
+Figure 1: Performance evolution with an increasing number of retrieved documents on two datasets. Traditional RAG requires more documents for better performance, while our method achieves comparable or superior results with significant token compression.
+[\(Xu et al., 2024\)](#page-12-0) suffers a 3–5 point drop in EM score (Table [1\)](#page-5-0), making it unsuitable for accuracysensitive applications. Second, most compression methods are heuristic in nature. The models are typically trained to generate summaries that are generally good but not necessarily useful for the downstream answer-generation LLM. This limitation arises from the lack of an ideal supervisory signal that defines what an optimal summary should be for the downstream task. This fundamental gap hinders the end-to-end optimization of existing methods. Finally, some compression models [\(Zhu et al., 2024\)](#page-12-3) have parameter counts comparable to the LLM that performs the end task, resulting in substantial computational costs that undermine the efficiency gains of compression.
+Addressing these critical shortcomings requires a method that aligns compression with downstream task requirements, thereby minimizing performance trade-offs. To bridge this gap, we propose CORE, a novel method designed to achieve lossless context compression for RAG. Unlike previous compression methods, CORE is optimized in an end-to-end manner. Since obtaining predefined summary labels for supervision is impractical, we instead use downstream task performance as a feedback signal to evaluate the compression model's output. This feedback enables iterative refinement of the compression policy, guiding the compression model toward improved downstream performance. To this end, we employ Group Relative Policy Optimization (GRPO), a technique particularly well-suited for this purpose [\(Liu et al., 2024;](#page-11-2) [Shao et al., 2024;](#page-12-4) [Chen et al., 2025a\)](#page-10-4). In our framework, the accuracy of the downstream QA task is formalized as a reward, with the compression policy optimized through group-wise relative comparisons. Furthermore, our compression model is substantially smaller than the downstream LLM, which significantly reduces the computational overhead associated with encoding retrieved documents.
+We evaluate CORE on four benchmark datasets: two single-hop QA datasets (*Natural Questions* and *TriviaQA*) and two multi-hop datasets (*HotpotQA* and *2WikiMultihopQA*). As shown in Table [1,](#page-5-0) CORE achieves state-of-the-art performance across all baselines. With a compression ratio of 3%, our approach not only avoids performance degradation compared to prepending full documents but also improves the average EM score by 3.3 points. We further demonstrate two key advantages of our approach. First, the effectiveness of CORE is not tied to a specific model architecture, as demonstrated by the fact that lossless compression can be achieved when various models are trained as compressors (Figure [3\)](#page-7-0). Second, the compressor exhibits strong transferability: a compression model trained using feedback from one LLM generalizes effectively to other LLMs (Table [2\)](#page-7-1). Finally, an in-depth case study (Tables [4](#page-14-0) and [5\)](#page-15-0) provides a qualitative analysis of the benefits of CORE.
+## 2 CORE-RAG
+This section introduces our proposed method, COmpression via REinforcement learning (CORE), which is shown in Figure [2.](#page-2-0) First, we provide an overview of the entire workflow. Then, we detail the end-to-end training strategy for the compression model, which is designed to drastically reduce the number of document tokens while preserving task performance.
+{2}------------------------------------------------
+<span id="page-2-0"></span>![](_page_2_Figure_1.jpeg)
+Figure 2: Overview of our method CORE. The upper section illustrates the inference pipeline. The lower section depicts the end-to-end training method for the compression model.
+#### 2.1 Problem Formulation
+We adopt the same problem formulation as prior work (Xu et al., 2024). Given an input question q, a target output y, and a set of k retrieved documents D, our objective is to compress D with respect to q into a summary s that preserves the most useful information while using significantly fewer tokens than D. This summary s is then prepended to the original input q and fed into an LLM to generate the final answer for the downstream task. This overall pipeline is illustrated in the upper half of Figure 2. The process involves two key components: a compression model $\pi_{\theta} \colon (q, D) \mapsto s$ and a large language model $M \colon (s, q) \mapsto \hat{y}$ , which generates the predicted answer $\hat{y}$ . We treat M as a black-box system and focus exclusively on training the compressor $\pi_{\theta}$ . The compressor itself is also a language model, but it is intentionally designed to be significantly smaller than M to reduce the computational cost of encoding the retrieved documents.
+#### 2.2 Training the Compressor
+Our compressor is designed to generate document summaries that are highly useful to the LLM (M) for downstream tasks. This objective is challenging because the criteria for an effective summary are task-dependent, and direct supervision is unavailable. We therefore formulate this as an end-to-end training problem and employ reinforcement learning to optimize the compressor without relying on pre-defined compression labels. The overall architecture of our training framework is illustrated in the lower portion of Figure 2. The following sections describe the key components of our approach: distillation warm-up, policy optimization, and reward calculation.
+### 2.2.1 DISTILLATION FOR WARM-START
+Due to the limited parameter size of our compression model, its capability for question-focused document summarization is constrained. To provide a strong initial policy for RL and ensure training stability, we employ knowledge distillation from a teacher model to initialize our compressor. Specifically, we first utilize a large-scale language model (DeepSeek-V3) as the teacher to generate summaries of retrieved documents related to the given question. We then evaluate the performance of the downstream LLM (M) on the QA task under two conditions: (1) with the teacher-generated summary $\hat{s}$ prepended to the input question q, and (2) with the original question alone. The corresponding performance scores are denoted as $p_{\text{summary}}$ and $p_{\text{original}}$ , respectively. By comparing these results, we retain instances where $p_{\text{summary}} > p_{\text{original}}$ , indicating that the summary enhances RAG performance. We also retain cases where $p_{\text{original}} = 1$ (i.e., the model produces a fully correct answer without the summary) and $p_{\text{summary}} < p_{\text{original}}$ ; for these, we set the target summary $\hat{s}$ to an empty string. All other instances are discarded. The resulting filtered and modified dataset is denoted as
+{3}------------------------------------------------
+$\mathcal{X}_f$ , which is used for supervised fine-tuning of the compression model. The fine-tuning objective is defined as:
+$\mathcal{L}_{\text{distill}} = \frac{1}{|\mathcal{X}_f|} \sum_{(q,D,\hat{s}) \in \mathcal{X}_f} \mathcal{L}_{\text{CE}}(\pi_{\theta}(q,D),\hat{s}), \tag{1}$
+where $\pi_{\theta}(q, D)$ denotes the output of the compression model and $\mathcal{L}_{CE}$ is the cross-entropy loss. This distillation process yields a robust initialization for RL and promotes stability in subsequent training.
+#### 2.2.2 END-TO-END TRAINING WITH RL
+Following the distillation phase, the compressor possesses a preliminary compression capability. However, as summaries from even the largest teacher models are not guaranteed to be optimal for the downstream task, further end-to-end optimization is necessary. We therefore formulate this optimization as a reinforcement learning problem. In this framework, the compressor functions as a policy that generates a summary from an input question and its corresponding documents. A reward function, which directly reflects performance on the downstream task (e.g., question-answering accuracy), then evaluates the summary's quality. The objective is to optimize the compressor's parameters to maximize the expected cumulative reward, thereby directly aligning its outputs with the downstream task's objectives.
+Specifically, we employ Group Relative Policy Optimization (GRPO) (Shao et al., 2024) algorithm. Unlike Proximal Policy Optimization (PPO), which trains a separate critic model, GRPO estimates the baseline directly from a group of rollouts. Given an existing policy, $\pi_{\theta_{\text{old}}}$ , and a reference policy, $\pi_{\theta_{\text{ref}}}$ , the GRPO objective maximizes the compressor policy $\pi_{\theta}$ using G rollouts $\tau = \{s_i\}_{i=1}^G \sim \pi_{\theta_{\text{old}}}(\cdot|x)$ , for each input $x \sim \mathcal{D}$ :
+$$\mathcal{J}(\theta) = \mathbb{E}_{x \sim \mathcal{D}, \ \{s_i\}_{i=1}^G \sim \pi_{\theta_{\text{old}}}(\cdot \mid x)}$$
+$$\frac{1}{G} \sum_{i=1}^G \left[ \min \left( \frac{\pi_{\theta}(s_i \mid x)}{\pi_{\theta_{\text{old}}}(s_i \mid x)} A_i, \ \text{clip} \left( \frac{\pi_{\theta}(s_i \mid x)}{\pi_{\theta_{\text{old}}}(s_i \mid x)}, 1 - \epsilon, 1 + \epsilon \right) A_i \right) - \beta \mathbb{D}_{\text{KL}} \left( \pi_{\theta} \parallel \pi_{\theta_{\text{ref}}} \right) \right],$$
+$$(2)$$
+where $A_i = (r_i - \text{mean}(\{r_j\}_{j=1}^G))/\text{std}(\{r_j\}_{j=1}^G)$ represents the normalized advantage of the i-th rollout within the group, $\epsilon$ is the clipping ratio, and $\beta$ is the coefficient for the KL divergence penalty. The inclusion of the KL divergence term ensures that the updated policy does not deviate significantly from the reference policy.
+#### 2.2.3 REWARD CALCULATION
+**Generating End-Task Output.** It is important to note that the reward is not computed directly from the compressor's output summary. Instead, the summary s generated by the compressor is prepended to the original input question q, and this combined input is fed into the LLM $M:(s,q)\mapsto \hat{y}$ to produce a predicted answer $\hat{y}$ . The reward is then calculated by comparing $\hat{y}$ to the gold answer y. Throughout the training process, the parameters of M remain fixed and are not updated.
+**Computing Rewards.** We design simple rule-based rewards based on end-task performance to guide the compressor's improvement, which consists of two components:
+• EM Reward ( $r_{\rm EM}$ ). We employ EM as the main reward function, which is a widely adopted metric for evaluating the accuracy of QA tasks. The EM reward is binary: it yields a value of 1 if the generated answer perfectly matches the ground truth, and 0 otherwise.
+$$r_{\rm EM} = \begin{cases} 1 & \text{if } y = \hat{y}, \\ 0 & \text{otherwise.} \end{cases}$$
+(3)
+• F1 Reward $(r_{\rm F1})$ . Since exact matches occur infrequently in practice, relying solely on EM rewards leads to sparse reward signals. Furthermore, the EM metric fails to distinguish between partially correct answers, as all non-exact matches receive zero reward. Therefore, we introduce F1 reward, which provides a finer-grained evaluation by measuring the degree of partial match.
+$$r_{\rm FI} = \frac{2 \times I_N}{P_N + R_N},\tag{4}$$
+where $P_N$ denotes the number of tokens in the predicted answer, $R_N$ denotes the number of tokens in the gold answer, and $I_N$ is the number of intersecting tokens between the two answers.
+{4}------------------------------------------------
+The final reward function combines these reward signals through weighted summation:
+$$r = r_{\rm EM} + \alpha \cdot r_{\rm Fl},\tag{5}$$
+where α ∈ (0, 1] is a hyperparameter that controls the relative contribution of the F1 reward.
+### 2.2.4 TRAINING TEMPLATE
+**221**
+**224**
+**256**
+Figure [5](#page-13-0) displays the prompt employed to train the compressor model for generating a summary of the retrieved documents, conditioned on the given question. This prompt is notably concise. For end-task answer generation, the prompt provided to the LLM M is illustrated in Figure [6,](#page-13-1) which incorporates few-shot in-context examples, the (generated summary of) retrieved documents, and the question.
+### 2.2.5 EFFICIENCY ANALYSIS
+Training Efficiency. Since our method employs reinforcement learning for training, it incurs greater time and computational costs compared to approaches that do not utilize reinforcement learning [\(Xu et al., 2024;](#page-12-0) [Cao et al., 2024\)](#page-10-3). However, our training process only optimizes a lightweight compressor model with relatively few parameters, while the larger generator LLM responsible for producing task answers remains fixed and is not updated during training. This design ensures high training efficiency—for instance, training one epoch takes approximately 2 hours using eight H20 GPUs, and convergence is typically achieved within just two epochs. In contrast, other reinforcement learning-based methods, such as ReSearch [\(Chen et al., 2025a\)](#page-10-4) and R1-Searcher [\(Song et al.,](#page-12-5) [2025\)](#page-12-5), require direct fine-tuning of the large generator LLM, leading to considerably higher training time and resource consumption. Furthermore, it is important to emphasize that our method exhibits strong generalization capability. As shown in Section 3.3, a model trained only once demonstrates broad applicability, thereby reducing the need for frequent retraining and further lowering the overall training cost.
+Inference Efficiency. Our method significantly enhances inference efficiency. In contrast to RAG approaches that do not employ a compressor—and thus require the generator LLM to directly encode lengthy documents, often spanning thousands of tokens—our approach introduces a lightweight compressor that processes long documents and summarizes them into compact representations of only a few dozen tokens before feeding them to the generator LLM. Since the compressor is an order of magnitude smaller in parameter size than the generator LLM, it substantially reduces the encoding time that would otherwise be incurred by the generator, leading to notable gains in inference efficiency. It is also important to note that the use of reinforcement learning does not adversely affect inference efficiency, as it is only involved during the training phase.
+## 3 EXPERIMENTS
+### 3.1 EXPERIMENTAL SETTINGS
+Datasets and Evaluation Metrics. We evaluate our method on four benchmark datasets: two single-hop question-answering datasets, Natural Questions (NQ) [\(Kwiatkowski et al., 2019\)](#page-11-3) and TriviaQA [\(Joshi et al., 2017\)](#page-10-5), as well as two multi-hop question-answering datasets, HotpotQA [\(Yang et al., 2018\)](#page-12-6) and 2WikiMultihopQA [\(Ho et al., 2020\)](#page-10-6). Results are reported on the test sets of Natural Questions and TriviaQA, as well as the development sets of HotpotQA and 2WikiMultihopQA. Following RECOMP [\(Xu et al., 2024\)](#page-12-0), the performance is measured using Exact Match and token-level F1 scores, while efficiency is assessed by the number of tokens provided in the context.
+Compression Model (πθ). We trained our compression model using Qwen2.5-1.5B-Instruct to generate summaries of the retrieved documents. To evaluate the effect of using different models as compressors, we also trained compressors using Llama3.2-1B-Instruct and Llama3.2-3B-Instruct (Section [3.3\)](#page-7-0).
+Large Language Model (M). We use Qwen2.5-14B-Instruct as the primary LLM model M to generate predicted answers which are used to guide the training of the compressor. To evaluate the generalization ability of our method, we also transfer to another LLM model, LLama3.1-8B-Instruct (Table [2\)](#page-7-1).
+{5}------------------------------------------------
+<span id="page-5-0"></span>Table 1: Open-domain QA results using Qwen2.5-14B-Instruct as the downstream LLM (M). The reported token counts represent the length of in-context documents, excluding few-shot examples. RECOMP, NoiseFilter-IB and our method CORE are all trained using Qwen2.5-1.5B-Instruct.
+| | | NQ | | 7 | TriviaQA | <b>L</b> | F | IotpotQ. | A | 2Wiki | Multiho | pQA |
+|-------------------------|---------|------------|----------|-----------|-----------|-----------|-------|----------|-------|-------|---------|-------|
+| | EM | F1 | # tok | EM | F1 | # tok | EM | F1 | # tok | EM | F1 | # tok |
+| No Retrieval | 21.36 | 30.97 | 0 | 53.23 | 59.98 | 0 | 21.05 | 29.48 | 0 | 26.11 | 29.51 | 0 |
+| RAG without compression | on | | | | | | | | | | | |
+| Top1 Document | 34.46 | 44.41 | 142 | 60.82 | 68.70 | 143 | 29.20 | 38.93 | 147 | 26.79 | 31.87 | 153 |
+| Top3 Documents | 37.78 | 48.45 | 427 | 62.60 | 71.02 | 430 | 31.64 | 41.87 | 442 | 27.89 | 33.58 | 460 |
+| Top5 Documents | 38.03 | 49.16 | 712 | 64.10 | 72.48 | 715 | 32.99 | 43.69 | 737 | 29.64 | 35.21 | 766 |
+| Top10 Documents | 38.67 | 50.03 | 1425 | 64.40 | 72.92 | 1428 | 33.95 | 44.88 | 1471 | 31.04 | 36.75 | 1531 |
+| Compression of top 5 do | cuments | ; | | | | | | | | | | |
+| BM25 | 25.23 | 36.47 | 37 | 55.36 | 63.90 | 39 | 24.18 | 35.73 | 71 | 25.42 | 30.29 | 68 |
+| Qwen2.5-1.5B | 31.94 | 43.03 | 36 | 57.99 | 66.70 | 30 | 27.36 | 37.47 | 33 | 25.93 | 31.18 | 32 |
+| DeepSeek-V3 (671B) | 37.73 | 50.39 | 54 | 64.13 | 73.20 | 50 | 33.59 | 44.83 | 48 | 27.99 | 32.67 | 92 |
+| RECOMP-Abs (1.5B) | 34.18 | 46.26 | 58 | 60.31 | 68.50 | 53 | 28.96 | 39.95 | 56 | 30.25 | 36.73 | 52 |
+| RECOMP-Ext (1.5B) | 33.84 | 46.05 | 56 | 60.18 | 68.39 | 48 | 29.93 | 41.09 | 45 | 30.78 | 37.07 | 51 |
+| NoiseFilter-IB (1.5B) | 35.15 | 45.94 | 48 | 59.51 | 68.15 | 35 | 27.97 | 38.62 | 38 | 27.85 | 34.69 | 40 |
+| LongLLMLingua (1.5B) | 33.65 | 43.15 | 152 | 58.96 | 66.82 | 148 | 28.03 | 38.49 | 149 | 29.37 | 33.62 | 153 |
+| QGC (1.5B) | 36.23 | 45.88 | 49 | 61.02 | 68.45 | 47 | 29.16 | 40.05 | 45 | 31.14 | 36.83 | 51 |
+| CORE (1.5B) | 41.02 | 50.40 | 46 | 65.63 | 72.55 | 32 | 33.67 | 45.06 | 36 | 36.72 | 42.05 | 49 |
+| Compression of top 10 a | locumen | ts (with t | the comp | ressor ti | rained or | ı top 5 d | locs) | | | | | |
+| BM25 | 25.91 | 36.88 | 38 | 55.28 | 63.16 | 37 | 23.49 | 35.01 | 68 | 25.61 | 30.54 | 65 |
+| Qwen2.5-1.5B | 32.94 | 44.84 | 40 | 58.45 | 67.31 | 33 | 28.17 | 38.48 | 36 | 26.22 | 31.57 | 34 |
+| DeepSeek-V3 (671B) | 37.79 | 51.07 | 56 | 65.29 | 74.45 | 53 | 34.62 | 45.69 | 50 | 29.00 | 34.64 | 40 |
+| RECOMP-Abs (1.5B) | 34.40 | 46.93 | 59 | 61.42 | 69.88 | 52 | 31.54 | 42.92 | 52 | 31.98 | 38.16 | 49 |
+| RECOMP-Ext (1.5B) | 33.96 | 46.34 | 60 | 61.03 | 69.51 | 50 | 31.92 | 43.18 | 55 | 32.52 | 38.87 | 44 |
+| NoiseFilter-IB (1.5B) | 35.36 | 46.24 | 50 | 59.92 | 68.32 | 38 | 28.21 | 38.83 | 38 | 28.63 | 35.16 | 42 |
+| LongLLMLingua (1.5B) | 33.78 | 43.37 | 154 | 59.17 | 66.97 | 150 | 28.33 | 38.95 | 148 | 29.62 | 34.11 | 151 |
+| QGC (1.5B) | 36.03 | 45.62 | 50 | 61.23 | 68.74 | 49 | 29.12 | 39.63 | 46 | 31.71 | 37.52 | 50 |
+| CORE (1.5B) | 41.88 | 51.26 | 52 | 66.76 | 73.64 | 33 | 34.68 | 46.35 | 37 | 37.99 | 43.28 | 48 |
+**Retrieval Corpus and Retrievers.** Following previous studies (Xu et al., 2024), we use the Wikipedia corpus from December 20, 2018, as the retrieval source for all four datasets. The articles are segmented into non-overlapping 100-word documents. To ensure that our method is not dependent on a specific retriever, we experiment with several mainstream retrievers. Specifically, we use DPR (Karpukhin et al., 2020) for NQ, a hybrid of DPR and BM25 (Robertson et al., 1995) for TriviaQA, and the Contriever model (Izacard et al., 2021) trained on the MS MARCO dataset (Nguyen et al., 2016) for HotpotQA and 2WikiMultihopQA.
+**Baselines.** To evaluate the effectiveness of our method, we compared it against various baselines. First, we evaluated the uncompressed approach—retaining the original in-context RALM setup—by prepending the top 1, 3, 5, and 10 retrieved documents to the prompt. We also tested alternative compression methods, including the traditional BM25 algorithm (which ranks sentences by their similarity to the input), off-the-shelf Qwen2.5-1.5B-Instruct model (with comparable parameter size to our approach), and DeepSeek-V3 model (671B parameters, far exceeding our compressor's capacity). Furthermore, we included state-of-the-art context compression methods for RAG, RECOMP (Xu et al., 2024), NoiseFilter-IB (Zhu et al., 2024), LongLLMLingua (Jiang et al., 2024) and QGC (Cao et al., 2024). For RECOMP, we evaluated both its abstractive and extractive variants. To ensure a fair comparison, all trainable approaches were all trained using the same model.
+**Implementation Details.** For the distillation warm-up phase, we perform full-parameter supervised fine-tuning on the off-the-shell language model for two epochs using LLaMA-Factory $^1$ . This warmed-up model then serves as the initializer for the subsequent reinforcement learning phase. We adopt the Verl framework $^2$ for RL training. The initialized compression model is trained for two epochs on each dataset. Training is conducted on eight NVIDIA H20 GPUs using full parameter GRPO optimization, with a learning rate of 1e-5, a batch size of 256, five rollouts per sample, and a KL loss coefficient of 0.001. The downstream LLM (M) used for reward generation is served using the vLLM inference engine during RL training.
+<sup>&</sup>lt;sup>1</sup>
+<sup>&</sup>lt;sup>2</sup>
+{6}------------------------------------------------
+### 3.2 OVERALL PERFORMANCE
+**329**
+**334 335 336**
+**364**
+**368 369 370**
+The detailed comparison results are presented in Table [1.](#page-5-0) We evaluate the following approaches: traditional RAG without compression using the top 1, 3, 5, and 10 documents prepended to the original input; BM25-based compression; off-the-shelf LLMs (Qwen2.5-1.5B-Instruct and DeepSeek-V3); and state-of-the-art trainable RAG compressors (RECOMP-Abstractive, RECOMP-Extractive, and NoiseFilter-IB). For fair comparison, all trainable methods were trained using the same backbone model, Qwen2.5-1.5B-Instruct. These compressors were trained on five-document inputs, and we report their performance both on in-domain five-document compression and out-of-domain generalization to ten-document compression.
+RAG versus No RAG. As shown in Table [1,](#page-5-0) prepending the original input with retrieved documents yields a substantial improvement over the no-retrieval baseline, indicating that these documents provide valuable information for generating the answer. Performance improves as the number of retrieved documents increases from 1 to 10, albeit with diminishing returns—a trend attributable to the decreasing relevance of lower-ranked documents. However, this performance gain comes at the cost of a significant increase in the number of additional tokens the LLM must encode, rising from 0 to over 1,400 tokens.
+Lossless Compression of CORE. Our compressor was trained using a context of five documents. As presented in Table [1,](#page-5-0) compared to prepending the full content of all five documents to the original input (i.e., the uncompressed baseline), our method achieves a high compression rate of approximately 6% *with no loss in performance*. Remarkably, on all four datasets, the approach not only maintains performance but also enhances EM by 1 to 7 points. For example, on NQ, the model achieves an EM of 41.02 with compressed input, surpassing the uncompressed score of 38.03.
+Comparison with Compression Baselines. When compressing the top five documents, all baseline compression methods result in performance degradation to varying degrees compared to the uncompressed baseline. Specifically, BM25 leads to a substantial performance drop. Using the offthe-shelf Qwen2.5-1.5B-Instruct model for compression yields better results than BM25 but still underperforms relative to the uncompressed baseline. Surprisingly, even the large-scale DeepSeek-V3 model (with 671B parameters) achieves only near-lossless compression on TriviaQA and HotpotQA, while performance on NQ and 2Wiki remains below the uncompressed baseline. As for the trained compressors (RECOMP, NoiseFilter-IB, LongLLMLingua, and QGC), all of them exhibit performance degradation compared to no compression. The decline generally ranges from 2 to 6 EM poin ts across nearly all datasets. In contrast, our method, CORE, delivers the best performance. It not only surpasses compression methods of comparable size by 4-5 EM points but also maintains an advantage over the hundreds-of-times-larger DeepSeek model. This clearly demonstrates the benefit and importance of end-to-end optimization.
+Generalization to 10-Doc Compression. When generalizing the trained compressors to handle the top-10 documents without retraining, the aforementioned conclusions remain valid. CORE continues to achieve lossless compression and performs best among all compression methods. On NQ, it achieves a token compression ratio of 3.6% while improving the EM by 3.2 points compared to using all ten documents. Similar trends are observed on TriviaQA, with a compression ratio of 2.3% and a gain of 2.4 EM points relative to the full-document baseline. On HotpotQA, CORE achieves a compression ratio of 2.5% and an improvement of 0.7 EM points. For 2WikiMultihopQA, it obtains a compression ratio of 3.1% along with a notable increase of 6.9 EM points.
+### 3.3 ROBUSTNESS AND GENERALIZATION ABILITY ANALYSIS
+Robustness of CORE Across Compressor Architectures. To evaluate whether the effectiveness of CORE depends on the choice of compressor model, we compared the performance of compressors trained using different model architectures—while keeping the downstream LLM (Qwen2.5-14B-Instruct) fixed. The compressors tested include LLaMA3.2-1B-Instruct, Qwen2.5-1.5B-Instruct, and LLaMA3.2-3B-Instruct, which vary in architecture and parameter count. As shown in Figure [3,](#page-7-0) the results indicate that: (1) These trained compressors consistently achieve lossless compression and outperform the uncompressed baseline (represented by the red reference line in the figure, which corresponds to prepending the full document content), confirming that our training framework is robust and not tied to a specific compressor architecture. (2) Compression performance improves as the size of the compressor model increases, consistent with scaling laws. More detailed results can be found in Tables [7](#page-17-0) and [8.](#page-17-1)
+{7}------------------------------------------------
+<span id="page-7-0"></span>![](_page_7_Figure_1.jpeg)
+Figure 3: The impact of different models used to train the compressor.
+Table 2: Zero-shot transfer of the trained compressor to Llama-3.1-8B-Instruct.
+<span id="page-7-1"></span>
+| | NQ | | | 7 | TriviaQA | | | HotpotQA | | | 2WikiMultihopQA | | |
+|----------------------|---------|-------|-------|-------|----------|-------|-------|----------|-------|-------|-----------------|-------|--|
+| | EM | F1 | # tok | EM | F1 | # tok | EM | F1 | # tok | EM | F1 | # tok | |
+| No Retrieval | 24.04 | 34.91 | 0 | 55.64 | 62.57 | 0 | 19.93 | 27.75 | 0 | 27.64 | 31.18 | 0 | |
+| RAG without compres | ssion | | | | | | | | | | | | |
+| Top1 Document | 33.80 | 44.06 | 142 | 59.17 | 67.50 | 143 | 27.95 | 37.49 | 147 | 28.41 | 33.43 | 153 | |
+| Top3 Documents | 36.87 | 47.81 | 427 | 61.13 | 70.06 | 430 | 30.17 | 40.71 | 442 | 28.67 | 34.23 | 460 | |
+| Top5 Documents | 37.65 | 48.87 | 712 | 62.26 | 71.04 | 715 | 31.44 | 42.16 | 737 | 29.43 | 35.18 | 766 | |
+| Top10 Documents | 38.12 | 49.93 | 1425 | 63.95 | 72.71 | 1428 | 32.19 | 42.62 | 1471 | 30.45 | 36.04 | 1531 | |
+| Compression of top 5 | docume | nts | | | | | | | | | | | |
+| Qwen2.5-1.5B | 32.60 | 44.21 | 36 | 56.76 | 65.77 | 30 | 26.86 | 36.90 | 33 | 25.45 | 30.88 | 32 | |
+| DeepSeek-V3 (671B) | 37.56 | 50.11 | 54 | 62.52 | 72.34 | 50 | 33.05 | 44.25 | 48 | 28.64 | 33.87 | 92 | |
+| RECOMP-Abs (1.5B) | 33.41 | 45.50 | 58 | 58.50 | 67.37 | 53 | 28.85 | 39.76 | 56 | 31.63 | 37.81 | 52 | |
+| RECOMP-Ext (1.5B) | 33.12 | 45.06 | 60 | 57.98 | 66.84 | 55 | 29.03 | 40.04 | 52 | 31.85 | 38.02 | 55 | |
+| CORE (1.5B) | 40.72 | 50.00 | 46 | 64.08 | 71.13 | 32 | 32.17 | 43.71 | 36 | 35.99 | 41.42 | 49 | |
+| Compression of top 1 | 0 docum | ents | | | | | | | | | | | |
+| Qwen2.5-1.5B | 32.88 | 44.66 | 40 | 57.44 | 66.56 | 33 | 27.31 | 37.31 | 36 | 25.80 | 31.30 | 34 | |
+| DeepSeek-V3 (671B) | 37.49 | 51.28 | 56 | 63.79 | 73.80 | 53 | 34.24 | 45.35 | 50 | 31.45 | 37.09 | 40 | |
+| RECOMP-Abs (1.5B) | 34.18 | 46.80 | 59 | 59.69 | 68.89 | 52 | 30.17 | 41.42 | 55 | 33.61 | 39.78 | 44 | |
+| RECOMP-Ext (1.5B) | 34.06 | 46.55 | 60 | 59.33 | 68.71 | 50 | 30.52 | 41.98 | 55 | 33.52 | 39.42 | 44 | |
+| CORE (1.5B) | 41.77 | 51.27 | 52 | 65.25 | 72.45 | 33 | 33.25 | 45.09 | 37 | 37.59 | 42.87 | 48 | |
+Transferability of CORE Across Downstream LLMs. We evaluate the transferability of our trained compressor and other baseline compressors to a new downstream LLM, the LLaMA-3.1-8B model. The results are presented in Table 2. Note that all trainable compressors, including our own, were trained using feedback generated by Qwen2.5-14B-Instruct. The findings reveal that existing trained compressors (e.g., RECOMP) exhibit limited generalization capability, as indicated by a larger performance gap relative to the baseline of prepending full documents. In contrast, CORE demonstrates stronger generalization, achieving lossless compression on the new downstream LLM. Specifically, it not only maintains performance but also surpasses the full-document baseline (i.e., no compression) across all four datasets while retaining a high compression rate. These results suggest that the summaries produced by our method are inherently high-quality and preserve key information necessary for accurate answering, thereby enabling effective transfer to other LLMs.
+### 3.4 ABLATION STUDY
+Table 3 presents an ablation study on the two stages of our method: distillation and GRPO. Here, "w/o distillation" denotes training the compressor with GRPO directly, bypassing the warm-start phase, while "w/o GRPO" indicates using only the distillation step without subsequent reinforcement learning. The results demonstrate that removing either component leads to performance degradation, confirming the necessity of both. In addition, the decline is more substantial when GRPO is omitted, highlighting the crucial role of reinforcement learning in the absence of explicit supervision. Distillation injects external knowledge into the model, providing a favorable starting point for RL training and thereby enabling RL to more effectively unlock the compressor's full potential.
+{8}------------------------------------------------
+Table 3: Ablation study.
+<span id="page-8-0"></span>
+| Dataset | Metric | w/o distillation | w/o GRPO | CORE |
+|----------|--------|------------------|----------|-------|
+| | EM | 36.37 | 34.18 | 41.02 |
+| NQ | F1 | 46.91 | 46.26 | 50.40 |
+| | EM | 65.23 | 60.31 | 65.63 |
+| TQA | F1 | 72.41 | 68.50 | 72.55 |
+| | EM | 32.01 | 28.96 | 33.67 |
+| HotpotQA | F1 | 42.73 | 39.95 | 45.06 |
+| | EM | 31.40 | 30.25 | 36.72 |
+| 2Wiki | F1 | 36.89 | 36.73 | 42.05 |
+![](_page_8_Figure_3.jpeg)
+Figure 4: The impact of α.
+### 3.5 THE IMPACT OF α.
+Figure [4](#page-8-0) illustrates the performance of our method under different values of α, which controls the weighting coefficient of the F1 reward term. Setting α = 0 corresponds to using only the EM reward. Results indicate that performance improves when α > 0 across all datasets, demonstrating the effectiveness of the F1 reward in mitigating the sparsity issue associated with the EM reward. However, the optimal value of α is dataset-dependent; values between 0.2 and 0.5 generally yield strong performance.
+### 3.6 CASE STUDY
+To conduct an in-depth analysis of the advantages of our compressor, we performed case studies on one single-hop QA dataset (NQ) and one multi-hop QA dataset (2Wiki), with the results presented in Table [4](#page-14-0) and Table [5,](#page-15-0) respectively. For each case, we compared the summaries generated by off-the-shelf Qwen2.5-1.5B-Instruct, RECOMP, and our method CORE based on the same set of documents, as well as the predicted answers generated by the LLM after prepending these summaries. As shown in the tables, although the summaries produced by off-the-shelf Qwen2.5-1.5B are concise, they largely fail to capture key information relevant to answering the question. In contrast, RECOMP demonstrates better summarization capability but is prone to being overwhelmed by lengthy documents, resulting in misjudgments and even generating misleading information—such as the statement in Table [4:](#page-14-0) "*The U.S. stopped drafting for the Vietnam War after the Selective Service System was officially abolished in December 1972*"—which leads the downstream LLM to produce the incorrect answer "1972". Our method, CORE, accurately extracts answer-critical information from lengthy documents, exemplified by the summary: "*The U.S. stopped drafting for the Vietnam War in 1973 after announcing the decision by Secretary of Defense Melvin Laird earlier that year*", thereby enabling the LLM to generate the correct answer "1973". This indicates that our compressor, trained with an end-task target-oriented reinforcement learning optimization strategy, can produce document summaries that are most helpful for answering the given question while effectively filtering out irrelevant information.
+## 4 RELATED WORK
+Context Compression in RAG. RAG enhances the performance of LLMs on knowledge-intensive tasks by retrieving the most relevant documents from extensive knowledge bases and prepending them as contextual information to the original input [\(Ram et al., 2023;](#page-11-0) [Fan et al., 2024;](#page-10-0) [Lin et al.,](#page-11-6) [2023;](#page-11-6) [Shi et al., 2023\)](#page-12-8). However, this approach requires the LLM to process significantly longer token sequences, resulting in increased computational costs. To mitigate this issue, researchers have begun to explore methods for compressing retrieved documents in RAG systems [\(Xu et al., 2024;](#page-12-0) [Cao et al., 2024;](#page-10-3) [Jin et al., 2024b;](#page-10-1) [Zhu et al., 2024;](#page-12-3) [Kim & Thorne, 2025;](#page-11-7) [Rau et al., 2024;](#page-12-9) [Wu et al.,](#page-12-1) [2025;](#page-12-1) [Louis et al., 2025;](#page-11-8) [Jin et al., 2024a;](#page-10-2) [Li et al., 2024a;](#page-11-9)[b;](#page-11-10) [Zhang et al., 2024a\)](#page-12-2). For instance, [Xu et al.](#page-12-0) [\(2024\)](#page-12-0) propose compressing retrieved documents into textual summaries before in-context augmentation, training the compressor through data selection and distillation. Similarly, [Cao et al.](#page-10-3) [\(2024\)](#page-10-3) introduce a Query-Guided Compressor (QGC) that uses queries to guide the compression process, effectively preserving essential information. [Jin et al.](#page-10-1) [\(2024b\)](#page-10-1) refine retrieved documents into Key Supporting Evidence (KSE) through a combination of knowledge synthesis, supervised fine-tuning (SFT), and preference alignment. Meanwhile, [Zhu et al.](#page-12-3) [\(2024\)](#page-12-3) present an informationtheoretic approach called NoiseFilter-IB, which filters noise by maximizing the mutual information
+{9}------------------------------------------------
+**509**
+**529 530**
+**538 539** between the compressed content and the ground-truth output. Additionally, [Kim & Thorne](#page-11-7) [\(2025\)](#page-11-7) train a compressor to extract critical information using reward functions based on predefined heuristic rules. However, most of these methods are heuristic in nature, and due to the lack of ideal compression labels, the compressed content they produce cannot be guaranteed to benefit downstream LLMs. In contrast, our method, CORE, adopts an end-to-end optimization approach to address these limitations.
+Reinforcement Learning. Reinforcement learning (RL) has recently achieved notable success, enabling LLMs to develop reasoning capabilities without explicit step-by-step supervision [\(Liu et al.,](#page-11-2) [2024;](#page-11-2) [Shao et al., 2024;](#page-12-4) [Guo et al., 2025\)](#page-10-9). Building on these advances, several studies have applied RL to improve RAG [\(Ke et al., 2024\)](#page-11-11). For example, [Kulkarni et al.](#page-11-12) [\(2024\)](#page-11-12) use RL to autonomously decide whether to retrieve documents, while [Zhang et al.](#page-12-10) [\(2024b\)](#page-12-10) employ RL to optimize the ranking of retrieved documents. Similarly, [Mao et al.](#page-11-13) [\(2024\)](#page-11-13) propose a framework for training query rewriting models for RAG without relying on human annotations. MMOA-RAG [\(Chen et al., 2025b\)](#page-10-10) enhances RAG through multi-agent reinforcement learning, incorporating a query rewriter, retriever, and generator. Meanwhile, RL has been applied to address other challenges in RAG, such as enhancing the quality of retrieved content. For instance, Oreo [\(Li & Ramakrishnan, 2025\)](#page-11-14) trains a reconstructor with PPO and ROUGE-based rewards to rewrite passages for improved performance. In contrast, our work CORE tackles the distinct problem of computational efficiency. Our approach leverages the GRPO algorithm, direct task-performance rewards, and a lightweight compressor to enhance efficiency without compromising accuracy. Moreover, a line of research has utilized RL to integrate search with reasoning in a step-by-step manner [\(Singh et al., 2025\)](#page-12-11). For instance, [Chen](#page-10-4) [et al.](#page-10-4) [\(2025a\)](#page-10-4) introduce a framework called ReSearch, which trains LLMs to reason with search using RL, without requiring supervised data for reasoning steps. Related approaches include R1- Searcher [\(Song et al., 2025\)](#page-12-5), WebThinker [\(Li et al., 2025\)](#page-11-15), and DeepResearcher [\(Zheng et al., 2025\)](#page-12-12). Although these methods are end-to-end, they differ fundamentally from our problem setting. These approaches typically involve directly training the LLM generator—which tends to be a large-scale model with a high parameter count. However, such training becomes infeasible when the model is a black box (e.g., GPT-4), as internal weights or gradients are inaccessible. Furthermore, these methods introduce extensive internal thinking processes that substantially increase context length and inference time. In contrast, our approach treats the generator LLM as a fixed black-box model and trains only a lightweight plug-in compressor to produce document summaries. This design significantly improves both training and inference efficiency. A parallel line of work applies RL to prompt compression. For example, PCRL [\(Jung & Kim, 2024\)](#page-11-16) and TACO-RL [\(Shandilya et al., 2025\)](#page-12-13) learn to compress prompts via token-level keep-or-drop decisions, using the similarity between model outputs with compressed and original prompts as the reward. Our work, CORE, introduces key distinctions in both objective and methodology. First, we target the more complex problem of compressing multiple retrieved documents in RAG, rather than single prompts. Second, we employ a generative compressor that can rephrase and synthesize content, instead of making token-level binary actions. Most importantly, we optimize compression using a direct task-performance reward with GRPO, which enables us to achieve true lossless compression at significantly higher ratios—a stark contrast to the performance degradation observed in prior prompt compression methods.
+## 5 CONCLUSION
+This paper analyzes the limitations of current context compression methods for RAG. A primary challenge is the lack of optimal reference summaries for supervised learning, which often results in performance degradation in downstream tasks. To overcome this, we frame context compression as a reinforcement learning problem, utilizing downstream task performance as a reward signal to train the compression policy, thereby enabling end-to-end optimization. Extensive experiments demonstrate that our proposed method, CORE, achieves effectively lossless compression by maintaining a high compression ratio while preserving original task performance. Surprisingly, CORE not only preserves but actually enhances performance on all downstream tasks. Further in-depth analysis provides additional insights into the efficacy of our approach.
+{10}------------------------------------------------
+## REFERENCES
+- Zhiwei Cao, Qian Cao, Yu Lu, Ningxin Peng, Luyang Huang, Shanbo Cheng, and Jinsong Su. Retaining key information under high compression ratios: Query-guided compressor for llms. In *Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)*, pp. 12685–12695, 2024.
+- <span id="page-10-4"></span>Mingyang Chen, Tianpeng Li, Haoze Sun, Yijie Zhou, Chenzheng Zhu, Haofen Wang, Jeff Z Pan, Wen Zhang, Huajun Chen, Fan Yang, et al. Learning to reason with search for llms via reinforcement learning. *arXiv preprint arXiv:2503.19470*, 2025a.
+- Yiqun Chen, Lingyong Yan, Weiwei Sun, Xinyu Ma, Yi Zhang, Shuaiqiang Wang, Dawei Yin, Yiming Yang, and Jiaxin Mao. Improving retrieval-augmented generation through multi-agent reinforcement learning. *arXiv preprint arXiv:2501.15228*, 2025b.
+- <span id="page-10-0"></span>Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat-Seng Chua, and Qing Li. A survey on rag meeting llms: Towards retrieval-augmented large language models. In *Proceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining*, pp. 6491–6501, 2024.
+- <span id="page-10-9"></span>Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. *arXiv preprint arXiv:2501.12948*, 2025.
+- <span id="page-10-6"></span>Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa. Constructing a multi-hop qa dataset for comprehensive evaluation of reasoning steps. *arXiv preprint arXiv:2011.01060*, 2020.
+- Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. Unsupervised dense information retrieval with contrastive learning. *arXiv preprint arXiv:2112.09118*, 2021.
+- <span id="page-10-8"></span>Huiqiang Jiang, Qianhui Wu, Xufang Luo, Dongsheng Li, Chin-Yew Lin, Yuqing Yang, and Lili Qiu. Longllmlingua: Accelerating and enhancing llms in long context scenarios via prompt compression. In *Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)*, pp. 1658–1677, 2024.
+- <span id="page-10-2"></span>Bowen Jin, Jinsung Yoon, Jiawei Han, and Sercan O Arik. Long-context llms meet rag: Overcoming challenges for long inputs in rag. *arXiv preprint arXiv:2410.05983*, 2024a.
+- <span id="page-10-1"></span>Jiajie Jin, Yutao Zhu, Yujia Zhou, and Zhicheng Dou. Bider: Bridging knowledge inconsistency for efficient retrieval-augmented llms via key supporting evidence. In *Findings of the Association for Computational Linguistics ACL 2024*, pp. 750–761, 2024b.
+- <span id="page-10-5"></span>Mandar Joshi, Eunsol Choi, Daniel S Weld, and Luke Zettlemoyer. Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension. *arXiv preprint arXiv:1705.03551*, 2017.
+{11}------------------------------------------------
+<span id="page-11-12"></span><span id="page-11-11"></span><span id="page-11-7"></span><span id="page-11-3"></span>**604 605 606**
+<span id="page-11-15"></span><span id="page-11-14"></span>**617**
+<span id="page-11-10"></span><span id="page-11-9"></span>**619**
+<span id="page-11-6"></span><span id="page-11-2"></span><span id="page-11-1"></span>**634**
+<span id="page-11-13"></span><span id="page-11-8"></span><span id="page-11-5"></span><span id="page-11-0"></span>**636**
+- <span id="page-11-16"></span><span id="page-11-4"></span>Hoyoun Jung and Kyung-Joong Kim. Discrete prompt compression with reinforcement learning. *IEEE Access*, 12:72578–72587, 2024.
+- Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick SH Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. Dense passage retrieval for open-domain question answering. In *EMNLP (1)*, pp. 6769–6781, 2020.
+- Zixuan Ke, Weize Kong, Cheng Li, Mingyang Zhang, Qiaozhu Mei, and Michael Bendersky. Bridging the preference gap between retrievers and llms. *arXiv preprint arXiv:2401.06954*, 2024.
+- Sangryul Kim and James Thorne. Context filtering with reward modeling in question answering. In *Proceedings of the 31st International Conference on Computational Linguistics*, pp. 11048– 11055, 2025.
+- Mandar Kulkarni, Praveen Tangarajan, Kyung Kim, and Anusua Trivedi. Reinforcement learning for optimizing rag for domain chatbots. *arXiv preprint arXiv:2401.06800*, 2024.
+- Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, et al. Natural questions: a benchmark for question answering research. *Transactions of the Association for Computational Linguistics*, 7:453–466, 2019.
+- Sha Li and Naren Ramakrishnan. Oreo: A plug-in context reconstructor to enhance retrievalaugmented generation. In *Proceedings of the 2025 International ACM SIGIR Conference on Innovative Concepts and Theories in Information Retrieval (ICTIR)*, pp. 238–253, 2025.
+- Xiaoxi Li, Jiajie Jin, Guanting Dong, Hongjin Qian, Yutao Zhu, Yongkang Wu, Ji-Rong Wen, and Zhicheng Dou. Webthinker: Empowering large reasoning models with deep research capability. *arXiv preprint arXiv:2504.21776*, 2025.
+- Xinze Li, Sen Mei, Zhenghao Liu, Yukun Yan, Shuo Wang, Shi Yu, Zheni Zeng, Hao Chen, Ge Yu, Zhiyuan Liu, et al. Rag-ddr: Optimizing retrieval-augmented generation using differentiable data rewards. *arXiv preprint arXiv:2410.13509*, 2024a.
+- Zhonghao Li, Xuming Hu, Aiwei Liu, Kening Zheng, Sirui Huang, and Hui Xiong. Refiner: Restructure retrieval content efficiently to advance question-answering capabilities. *arXiv preprint arXiv:2406.11357*, 2024b.
+- Xi Victoria Lin, Xilun Chen, Mingda Chen, Weijia Shi, Maria Lomeli, Richard James, Pedro Rodriguez, Jacob Kahn, Gergely Szilvasy, Mike Lewis, et al. Ra-dit: Retrieval-augmented dual instruction tuning. In *The Twelfth International Conference on Learning Representations*, 2023.
+- Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report. *arXiv preprint arXiv:2412.19437*, 2024.
+- Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. Lost in the middle: How language models use long contexts. *arXiv preprint arXiv:2307.03172*, 2023.
+- Maxime Louis, Herve D´ ejean, and St ´ ephane Clinchant. Pisco: Pretty simple compression for ´ retrieval-augmented generation. *arXiv preprint arXiv:2501.16075*, 2025.
+- Shengyu Mao, Yong Jiang, Boli Chen, Xiao Li, Peng Wang, Xinyu Wang, Pengjun Xie, Fei Huang, Huajun Chen, and Ningyu Zhang. Rafe: Ranking feedback improves query rewriting for rag. *arXiv preprint arXiv:2405.14431*, 2024.
+- Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh Tiwary, Rangan Majumder, and Li Deng. Ms marco: A human-generated machine reading comprehension dataset. 2016.
+- Ori Ram, Yoav Levine, Itay Dalmedigos, Dor Muhlgay, Amnon Shashua, Kevin Leyton-Brown, and Yoav Shoham. In-context retrieval-augmented language models. *Transactions of the Association for Computational Linguistics*, 11:1316–1331, 2023.
+{12}------------------------------------------------
+<span id="page-12-13"></span>**654**
+<span id="page-12-11"></span><span id="page-12-8"></span><span id="page-12-5"></span><span id="page-12-4"></span><span id="page-12-1"></span>**656**
+<span id="page-12-12"></span><span id="page-12-10"></span><span id="page-12-6"></span><span id="page-12-3"></span><span id="page-12-2"></span><span id="page-12-0"></span>**684 685**
+- <span id="page-12-9"></span><span id="page-12-7"></span>David Rau, Shuai Wang, Herve D´ ejean, and St ´ ephane Clinchant. Context embeddings for efficient ´ answer generation in rag. *arXiv preprint arXiv:2407.09252*, 2024.
+- Stephen E Robertson, Steve Walker, Susan Jones, Micheline M Hancock-Beaulieu, Mike Gatford, et al. *Okapi at TREC-3*. British Library Research and Development Department, 1995.
+- Shivam Shandilya, Menglin Xia, Supriyo Ghosh, Huiqiang Jiang, Jue Zhang, Qianhui Wu, Victor Ruhle, and Saravan Rajmohan. Taco-rl: Task aware prompt compression optimization with rein- ¨ forcement learning. In *Findings of the Association for Computational Linguistics: ACL 2025*, pp. 1582–1597, 2025.
+- Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. *arXiv preprint arXiv:2402.03300*, 2024.
+- Weijia Shi, Sewon Min, Michihiro Yasunaga, Minjoon Seo, Rich James, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. Replug: Retrieval-augmented black-box language models. *arXiv preprint arXiv:2301.12652*, 2023.
+- Aditi Singh, Abul Ehtesham, Saket Kumar, and Tala Talaei Khoei. Agentic retrieval-augmented generation: A survey on agentic rag. *arXiv preprint arXiv:2501.09136*, 2025.
+- Huatong Song, Jinhao Jiang, Yingqian Min, Jie Chen, Zhipeng Chen, Wayne Xin Zhao, Lei Fang, and Ji-Rong Wen. R1-searcher: Incentivizing the search capability in llms via reinforcement learning. *arXiv preprint arXiv:2503.05592*, 2025.
+- Chenyuan Wu, Ninglu Shao, Zheng Liu, Shitao Xiao, Chaozhuo Li, Chen Zhang, Senzhang Wang, and Defu Lian. Lighter and better: Towards flexible context adaptation for retrieval augmented generation. In *Proceedings of the Eighteenth ACM International Conference on Web Search and Data Mining*, pp. 271–280, 2025.
+- Fangyuan Xu, Weijia Shi, and Eunsol Choi. Recomp: Improving retrieval-augmented lms with context compression and selective augmentation. In *The Twelfth International Conference on Learning Representations*, 2024.
+- Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W Cohen, Ruslan Salakhutdinov, and Christopher D Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question answering. *arXiv preprint arXiv:1809.09600*, 2018.
+- Qianchi Zhang, Hainan Zhang, Liang Pang, Hongwei Zheng, and Zhiming Zheng. Adacomp: Extractive context compression with adaptive predictor for retrieval-augmented large language models. *arXiv preprint arXiv:2409.01579*, 2024a.
+- Taolin Zhang, Dongyang Li, Qizhou Chen, Chengyu Wang, Longtao Huang, Hui Xue, Xiaofeng He, and Jun Huang. R 4: Reinforced retriever-reorder-responder for retrieval-augmented large language models. In *ECAI 2024*, pp. 2314–2321. IOS Press, 2024b.
+- Yuxiang Zheng, Dayuan Fu, Xiangkun Hu, Xiaojie Cai, Lyumanshan Ye, Pengrui Lu, and Pengfei Liu. Deepresearcher: Scaling deep research via reinforcement learning in real-world environments. *arXiv preprint arXiv:2504.03160*, 2025.
+- Kun Zhu, Xiaocheng Feng, Xiyuan Du, Yuxuan Gu, Weijiang Yu, Haotian Wang, Qianglong Chen, Zheng Chu, Jingchang Chen, and Bing Qin. An information bottleneck perspective for effective noise filtering on retrieval-augmented generation. In *Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)*, pp. 1044–1069, 2024.
+{13}------------------------------------------------
+<span id="page-13-0"></span>Compress the information in the retrieved documents into a 2-sentence summary that could be used to answer the question. If the documents do not contain relevant information, simply output " ".
+Question: {*Question*}
+**709**
+**724**
+Retrieved documents: {*Documents*}
+Compressed documents:
+Figure 5: Prompt template used to generate the summary.
+### <span id="page-13-1"></span>[Instruction] Answer the question.
+IMPORTANT: Respond ONLY with the exact answer in the same format as the examples. Do NOT add any extra text, explanations, or punctuation. Do NOT include "Answer:" or any similar prefix in your response.
+### [Examples]
+Question: Which major Russian city borders the body of water in which Saaremaa is located?
+Answer: Saint Petersburg
+Question: Who was thee first president of the association that wrote the code of ethics for psychology?
+Answer: G. Stanley Hall
+Question: Where did the Baldevins bryllup director die?
+Answer: Copenhagen
+### [Current Question]
+**{***Summary of the retrieved documents***}**
+Question: {*Question*}
+Answer:
+Figure 6: Prompt template for LLM QA.
+## A USE OF LLMS
+Large language models (LLMs) were employed solely as writing assistants to enhance the language, improve clarity, and check grammatical correctness. They were not used to generate research ideas, design or implement methodologies, conduct data analysis, or produce any of the results presented in this work. The authors assume full responsibility for the entire content of the paper.
+## B PROMPT TEMPLATES
+Figure [5](#page-13-0) displays the prompt employed to train the compressor model for generating a summary of the retrieved documents, conditioned on the given query. This prompt is notably concise. For end-task answer generation, the prompt provided to the LLM M is illustrated in Figure [6,](#page-13-1) which incorporates few-shot in-context examples, the (generated summary of) retrieved documents, and the question.
+## C CASE STUDY
+To conduct an in-depth analysis of the advantages of our compressor, we performed case studies on one single-hop QA dataset (NQ) and one multi-hop QA dataset (2Wiki), with the results presented in Table [4](#page-14-0) and Table [5,](#page-15-0) respectively. For each case, we compared the summaries generated by off-the-shelf Qwen2.5-1.5B-Instruct, RECOMP, and our method CORE based on the same set of documents, as well as the predicted answers generated by the LLM after prepending these summaries. As shown in the tables, although the summaries produced by off-the-shelf Qwen2.5-1.5B
+{14}------------------------------------------------
+### Table 4: Case study on NQ dataset.
+**758**
+**759**
+**761**
+**764**
+**779 780 781**
+# <span id="page-14-0"></span>Question: when did the us stop drafting for the vietnam war? Gold answer: [1973]
+### Top-5 documents:
+last men conscripted, who were born in 1952 and who reported for duty in June 1973. On February 2, 1972, a drawing was held to determine draft priority numbers for men born in 1953, but in early 1973 it was announced by Secretary of Defense Melvin Laird that no further draft orders would be issued. In March 1973, 1974, and 1975, the Selective Service assigned draft priority numbers for all men born in 1954, 1955, and 1956, in case the draft was extended, but it never was. Command Sergeant Major Jeff Mellinger, believed to be the last drafted enlisted ranked.
+The Gates Commission issued its report in February 1970, describing how adequate military strength could be maintained without having conscription. The existing draft law was expiring at the end of June 1971, but the Department of Defense and Nixon administration decided the draft needed to continue for at least some time. In February 1971, the administration requested of Congress a two-year extension of the draft, to June 1973. Senatorial opponents of the war wanted to reduce this to a one-year extension, or eliminate the draft altogether, or tie the draft renewal to a timetable for troop withdrawal from Vietnam; Senator.
+draft operated from 1940 until 1946 when further inductions were suspended, and its legislative authorization expired without further extension by Congress in 1947. During this time, more than 10 million men had been inducted into military service. However, the Selective Service System remained intact. Scattered opposition was encountered especially in the northern cities where some African-Americans protested the system. The tiny Nation of Islam was at the forefront, with many Black Muslims jailed for refusing the draft, and their leader Elijah Muhammed was sentenced to federal prison for 5 years for inciting draft resistance. Organized draft resistance also developed in.
+was the first time any form of student deferment was used. During the Korean War a student carrying at least twelve semester hours was spared until the end of his current semester. The United States breathed easier with the Korean War Armistice on July 27, 1953; however, technology brought new promises and threats. U.S. air and nuclear power fueled the Eisenhower doctrine of "massive retaliation". This strategy demanded more machines and fewer foot soldiers, so the draft slipped to the back burner. However, the head of the SSS, Maj. Gen. Hershey, urged caution fearing the conflict looming in Vietnam. In.
+policy of using the draft as force to compel "voluntary" enlistment was unique in U.S. history. Previous drafts had not aimed at encouraging individuals to sign up in order to gain preferential placement or less dangerous postings. However, the incremental buildup of Vietnam without a clear threat to the country bolstered this. Some estimates suggest conscription encompassed almost one-third of all eligible men during the period of 1965–69. This group represented those without exemption or resources to avoid military service. During the active combat phase, the possibility of avoiding combat by selecting their service and military specialty led as many.
+| Compressor | Generated summary | Predicted<br>answer |
+|--------------------------------------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|---------------------|
+| Off-the-shell<br>Qwen2.5-<br>1.5B-Instruct | The US stopped drafting for the Vietnam War in late 1972 after<br>the draft expired at the end of June 1971. | late 1972 |
+| RECOMP | The U.S. stopped drafting for the Vietnam War after the<br>Selective Service System was officially abolished in December<br>1972 due to growing public opposition and the expiration of the<br>draft's legal authority in 1971. While the draft was formally<br>ended, some individuals were still consripted into the military | December<br>1972 |
+| Ours | The U.S. stopped drafting for the Vietnam War in 1973 after<br>announcing the decision by Secretary of Defense Melvin Laird<br>earlier that year. Although the Selective Service System was<br>later updated to include draft priority numbers, this decision<br>marked the end of the draft's use for national service. | 1973 |
+{15}------------------------------------------------
+### Table 5: Case study on 2Wiki dataset.
+**812 813**
+**814 815**
+**817**
+**829 830**
+**834**
+### <span id="page-15-0"></span>Question: Who is Charles Bretagne Marie De La Tremoille's paternal grandfather? ´ Gold answer: [Charles Armand Rene de La Tr ´ emoille] ´
+Top-5 documents:
+as at Versailles: he was brigadier of cavalry (January 1709), first gentleman of the King's chamber (June 1709), governor of Thouars (July 1709), and Marechal de camp (February 1719). His sister ´ Marie Armande Victoire de La Tremoille married Emmanuel Th ´ eodose de La Tour d'Auvergne. ´ On 13 April 1706 he married Marie-Madeleine Motier de La Fayette (1691–1717), the daughter of Rene-Armand, marquis de La Fayette and Marie-Madeleine de Marillac, and granddaughter of the author Marie-Madeleine Pioche de la Vergne, comtesse de la Fayette. They had one child, Charles Armand Rene de La Tr ´ emoille, born in 1708. Charles Louis Bretagne de La ´
+Charles Bretagne Marie de La Tremoille Charles Bretagne Marie de La Tr ´ emoille (24 March ´ 1764 – 10 November 1839), 8th duc de Thouars, was a French soldier and the son of Jean Bretagne Charles de La Tremoille and his wife, Marie-Maximilienne, princess of Salm-Kyrburg. La ´ Tremoille married Louise-Emmanuelle de Ch ´ atillon in 1781. She was a grand daughter of Louis ˆ Cesar de La Baume Le Blanc, the famous writer. The couple had one daughter: At the outbreak ´ of the French Revolution, he sided with King Louis XVI, as did the rest of his family. In 1789, La Tremoille and his parents ´
+Jean Bretagne Charles de La Tremoille Jean-Bretagne-Charles de La Tr ´ emoille (5 February 1737 ´ – 19 May 1792), 7th duc de Thouars, was a French soldier and the son of Charles Armand Rene´ de La Tremoille and his wife, Marie Hortense de La Tour d'Auvergne. La Tr ´ emoille began his ´ military career in the Seven Years' War as the head of the regiment of Aquitaine-Cavalry. He was seriously wounded and afterwards promoted to the rank of sergeant, and later of brigadier, of the armies of the king. During the French Revolution, La Tremoille remained faithful to King Louis ´ XVI. He emigrated in
+Charles Louis Bretagne de La Tremoille Charles Louis Bretagne de La Tr ´ emoille (15 March ´ 1683 – 9 October 1719), 6th duke of Thouars, was the son of Charles Belgique Hollande de La Tremoille and Madeleine de Cr ´ equy, daughter and heiress of Charles III de Cr ´ equy. He became ´ duke of Thouars upon his father's death in 1709. He was also count of Laval and Montfort and inherited the ancient French claim to the Kingdom of Naples. The Neapolitan claim gave the family the rank of "princes etrangers" at the French court. The Duke made a career in the ´ military as well
+de Saint-Didier. They had two daughters: Following her death, La Tremoille married for a third ´ time in 1830 to Valentine Eugenie Jos ´ ephine Walsh de Serrant. They had two children: La ´ Tremoille died shortly after his son's birth, in 1839. Charles Bretagne Marie de La Tr ´ emoille ´ Charles Bretagne Marie de La Tremoille (24 March 1764 – 10 November 1839), 8th duc de ´ Thouars, was a French soldier and the son of Jean Bretagne Charles de La Tremoille and his wife, ´ Marie-Maximilienne, princess of Salm-Kyrburg. La Tremoille married Louise-Emmanuelle de ´ Chatillon in 1781. She was a grand daughter of Louis C ˆ esar de ´
+| Compressor | Generated summary | Predicted<br>answer |
+|--------------------------------------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-------------------------------------------------|
+| Off-the-shell<br>Qwen2.5-<br>1.5B-Instruct | Charles Bretagne Marie de La Tremoille's paternal grandfather<br>´<br>was Jean Bretagne Charles de La Tremoille. ´ | Jean Bretagne<br>Charles de La<br>Tremoille ´ |
+| RECOMP | Charles Bretagne Marie de La Tremoille's paternal grandfather<br>´<br>was Charles Louis Bretagne de La Tremoille, who was the son<br>´<br>of Charles Belgique Hollande de La Tremoille and Madeleine<br>´<br>de Crequy. This makes him the direct paternal grandfather of<br>´<br>Charles Bretagne Marie de La Tremoille. ´ | Charles Louis<br>Bretagne de<br>La Tremoille ´ |
+| Ours | ´<br>Charles Bretagne Marie de La Tremoille's paternal grandfather<br>is Charles Armand Rene de La Tr<br>´<br>emoille, her father's father, the<br>´<br>7th duc de Thouars. | Charles<br>Armand Rene´<br>de La<br>Tremoille ´ |
+**861 863**
+**854 855 856**
+{16}------------------------------------------------
+**869**
+**884**
+**904**
+**906**
+Table 6: Failure Cases on NQ dataset.
+<span id="page-16-0"></span>
+| Question | Generated Summary | Predicted<br>Answer | Gold<br>Answer |
+|-----------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|---------------------|-----------------|
+| who is the<br>president of<br>usa right now | The current president of the USA is **Donald<br>Trump**, who was elected on January 20, 2017, and<br>is the 45th president. He is a businessman and reality<br>TV personality, and his inauguration marked a shift in<br>the country's political landscape. | Joe Biden | Donald<br>Trump |
+| who has been<br>ranked no. 1<br>in the latest<br>football<br>rankings<br>announced by<br>fifa | The latest FIFA football rankings, announced in 2018,<br>have ranked France as the top team, after winning the<br>2018 World Cup. This change was implemented using<br>the Elo rating system, and the rankings were<br>introduced to better reflect football team strengths. | France | Germany |
+are concise, they largely fail to capture key information relevant to answering the question. In contrast, RECOMP demonstrates better summarization capability but is prone to being overwhelmed by lengthy documents, resulting in misjudgments and even generating misleading information—such as the statement in Table [4:](#page-14-0) "*The U.S. stopped drafting for the Vietnam War after the Selective Service System was officially abolished in December 1972*"—which leads the downstream LLM to produce the incorrect answer "1972". Our method, CORE, accurately extracts answer-critical information from lengthy documents, exemplified by the summary: "*The U.S. stopped drafting for the Vietnam War in 1973 after announcing the decision by Secretary of Defense Melvin Laird earlier that year*", thereby enabling the LLM to generate the correct answer "1973". This indicates that our compressor, trained with an end-task target-oriented reinforcement learning optimization strategy, can produce document summaries that are most helpful for answering the given question while effectively filtering out irrelevant information.
+To further understand the limitations of our approach, we present two failure cases from the NQ dataset where the model provided incorrect answers based on our generated summaries. As shown in Table [6,](#page-16-0) the first case reveals that although the summary contained the key information required for the correct answer, the downstream LLM still produced an error, potentially due to its over-reliance on parametric knowledge. In the second case, the summary itself omitted critical information needed to answer the question, which likely led to the incorrect response.
+## D IMPACT OF DIFFERENT COMPRESSORS ON PERFORMANCE
+In our previous experiments, we employed Qwen2.5-1.5B as the initial model to train our compressor. In this section, we utilize two additional models—Llama3.2-1B and Llama3.2-3B—as starting points to train our compressor and the baseline compressor, respectively. The experimental results are presented in Table [7](#page-17-0) and Table [8.](#page-17-1) As shown in the results, our method CORE continues to achieve lossless compression with both models, maintaining a high token compression ratio while exhibiting no performance degradation in terms of Exact Match (EM) and F1 score compared to uncompressed RAG. Furthermore, under both new model configurations, our approach consistently outperforms the baseline methods, indicating that its superiority is not dependent on a specific model architecture and thus demonstrates strong robustness.
+We also observe that our method adheres to a form of scaling law: the compressor trained using the 3B model outperforms the one trained with the 1B model. Specifically, the 1B compressor improves performance by 1–4 EM points over the uncompressed baseline, while the 3B compressor yields gains of 3–9 EM points.
+{17}------------------------------------------------
+<span id="page-17-0"></span>Table 7: Open-domain QA results using Qwen2.5-14B-Instruct as the downstream LLM (M). The reported token counts represent the length of in-context documents, excluding few-shot examples. RECOMP and our method CORE are both trained using **llama3.2-1B-Instruct**.
+| | NQ | | | | TriviaQA | | | HotpotQA | | | 2WikiMultihopQA | | |
+|----------------------|------------|-------------|---------|-----------|------------|------------|--------|----------|-------|--------|-----------------|-------|--|
+| | EM | F1 | # tok | EM | F1 | # tok | EM | F1 | # tok | EM | F1 | # tok | |
+| No Retrieval | 0.2136 | 0.3097 | 0 | 0.5323 | 0.5998 | 0 | 0.2105 | 0.2948 | 0 | 0.2611 | 0.2951 | 0 | |
+| RAG without compre | ession | | | | | | | | | | | | |
+| Top1 Document | 0.3446 | 0.4441 | 142 | 0.6082 | 0.6870 | 143 | 0.2920 | 0.3893 | 147 | 0.2679 | 0.3187 | 153 | |
+| Top3 Documents | 0.3778 | 0.4845 | 427 | 0.6260 | 0.7102 | 430 | 0.3164 | 0.4187 | 442 | 0.2789 | 0.3358 | 460 | |
+| Top5 Documents | 0.3803 | 0.4916 | 712 | 0.6410 | 0.7248 | 715 | 0.3299 | 0.4369 | 737 | 0.2964 | 0.3521 | 766 | |
+| Top10 Documents | 0.3867 | 0.5003 | 1425 | 0.6440 | 0.7292 | 1428 | 0.3395 | 0.4488 | 1471 | 0.3104 | 0.3675 | 1531 | |
+| Compression of top 3 | 5 docs | | | | | | | | | | | | |
+| llama3.2-1B | 0.3147 | 0.4227 | 64 | 0.5552 | 0.6415 | 60 | 0.2648 | 0.3639 | 58 | 0.2498 | 0.3003 | 61 | |
+| Deepseek-V3 (671B) | 0.3773 | 0.5039 | 54 | 0.6528 | 0.7433 | 51 | 0.3359 | 0.4483 | 48 | 0.2507 | 0.3031 | 45 | |
+| RECOMP (1B) | 0.3410 | 0.4655 | 57 | 0.6071 | 0.6880 | 48 | 0.2987 | 0.4121 | 49 | 0.3045 | 0.3653 | 33 | |
+| CORE (1B) | 0.3947 | 0.4923 | 47 | 0.6483 | 0.7287 | 43 | 0.3344 | 0.4454 | 45 | 0.3378 | 0.3969 | 34 | |
+| Compression of top | 10 docs (w | vith the co | mpresso | r trained | on top 5 d | locs) | | | | | | | |
+| llama3.2-1B | 0.3141 | 0.4228 | 62 | 0.5651 | 0.6512 | <b>5</b> 8 | 0.2663 | 0.3661 | 56 | 0.2493 | 0.3006 | 61 | |
+| Deepseek-V3 (671B) | 0.3779 | 0.5107 | 56 | 0.6529 | 0.7445 | 53 | 0.3462 | 0.4569 | 50 | 0.2900 | 0.3464 | 40 | |
+| RECOMP (1B) | 0.3421 | 0.4661 | 59 | 0.6095 | 0.6917 | 52 | 0.2982 | 0.4105 | 55 | 0.3072 | 0.3681 | 44 | |
+| CORE (1B) | 0.4033 | 0.5033 | 47 | 0.6521 | 0.7296 | 45 | 0.3412 | 0.4500 | 48 | 0.3586 | 0.4162 | 42 | |
+<span id="page-17-1"></span>Table 8: Open-domain QA results using Qwen2.5-14B-Instruct as the downstream LLM (M). The reported token counts represent the length of in-context documents, excluding few-shot examples. RECOMP and our method CORE are both trained using **llama3.2-3B-Instruct**.
+| | | NQ | | | TriviaQA | | | HotpotQA | | | 2WikiMultihopQA | | |
+|----------------------|------------|-------------|---------|------------|------------|-------|--------|----------|-------|--------|-----------------|-------|--|
+| | EM | F1 | # tok | EM | F1 | # tok | EM | F1 | # tok | EM | F1 | # tok | |
+| No Retrieval | 0.2136 | 0.3097 | 0 | 0.5323 | 0.5998 | 0 | 0.2105 | 0.2948 | 0 | 0.2611 | 0.2951 | 0 | |
+| RAG without compre | ession | | | | | | | | | | | | |
+| Top1 Document | 0.3446 | 0.4441 | 142 | 0.6082 | 0.6870 | 143 | 0.2920 | 0.3893 | 147 | 0.2679 | 0.3187 | 153 | |
+| Top3 Documents | 0.3778 | 0.4845 | 427 | 0.6260 | 0.7102 | 430 | 0.3164 | 0.4187 | 442 | 0.2789 | 0.3358 | 460 | |
+| Top5 Documents | 0.3803 | 0.4916 | 712 | 0.6410 | 0.7248 | 715 | 0.3299 | 0.4369 | 737 | 0.2964 | 0.3521 | 766 | |
+| Top10 Documents | 0.3867 | 0.5003 | 1425 | 0.6440 | 0.7292 | 1428 | 0.3395 | 0.4488 | 1471 | 0.3104 | 0.3675 | 1531 | |
+| Compression of top : | 5 docs | | | | | | | | | | | | |
+| llama3.2-3B | 0.3252 | 0.4334 | 60 | 0.5650 | 0.6521 | 59 | 0.2772 | 0.3809 | 58 | 0.2485 | 0.2995 | 60 | |
+| Deepseek-V3 (671B) | 0.3773 | 0.5039 | 54 | 0.6528 | 0.7433 | 51 | 0.3359 | 0.4483 | 48 | 0.2507 | 0.3031 | 45 | |
+| RECOMP (3B) | 0.3657 | 0.4912 | 55 | 0.6183 | 0.6920 | 47 | 0.3025 | 0.4238 | 52 | 0.3274 | 0.3806 | 42 | |
+| CORE (3B) | 0.4310 | 0.5234 | 32 | 0.6650 | 0.7306 | 38 | 0.3507 | 0.4736 | 51 | 0.3905 | 0.4474 | 40 | |
+| Compression of top | 10 docs (w | vith the co | mpresso | or trained | on top 5 a | locs) | | | | | | | |
+| llama3.2-3B | 0.3318 | 0.4359 | 61 | 0.5720 | 0.6588 | 57 | 0.2791 | 0.3854 | 60 | 0.2491 | 0.3011 | 59 | |
+| Deepseek-V3 (671B) | 0.3779 | 0.5107 | 56 | 0.6529 | 0.7445 | 53 | 0.3462 | 0.4569 | 50 | 0.2900 | 0.3464 | 40 | |
+| RECOMP (3B) | 0.3682 | 0.4963 | 52 | 0.6205 | 0.6973 | 44 | 0.3077 | 0.4261 | 54 | 0.3312 | 0.3869 | 50 | |
+| CORE (3B) | 0.4526 | 0.5467 | 33 | 0.6736 | 0.7404 | 37 | 0.3693 | 0.4926 | 51 | 0.4071 | 0.4633 | 48 | |
+{18}------------------------------------------------
+<span id="page-18-1"></span>**979**
+**994**
+<span id="page-18-0"></span>Table 9: Zero-Shot Evaluation on HotpotQA of Models Trained on Natural Questions.
+| | EM | F1 | #tok |
+|-----------------|-------|-------|------|
+| No Retrieval | 21.05 | 29.48 | 0 |
+| Full Documents | 32.99 | 43.69 | 737 |
+| BM25 | 24.18 | 35.73 | 71 |
+| NoiseFilter-IB | 27.97 | 38.62 | 38 |
+| RECOMP | 28.96 | 39.95 | 56 |
+| CORE | 33.67 | 45.06 | 36 |
+| RECOMP-Transfer | 26.68 | 37.29 | 58 |
+| CORE-Transfer | 31.25 | 42.84 | 35 |
+| | | | |
+Table 10: Evaluation on Noisy Natural Questions.
+| | EM | F1 | #tok |
+|----------------|-------|-------|------|
+| full documents | 35.21 | 45.38 | 1427 |
+| RECOMP | 33.29 | 43.90 | 59 |
+| CORE | 38.19 | 48.85 | 48 |
+## E CROSS-DATASET GENERALIZATION PERFORMANCE
+To verify the universal compression capability of our method, we directly transfer the model trained on the single-hop question answering dataset NQ to the multi-hop dataset HotpotQA for evaluation. The results, shown in Table [9,](#page-18-0) indicate that our transferred model achieves nearly lossless performance compared to using full documents without compression, while substantially outperforming the transfer results of the RECOMP baseline. Moreover, although both our method and the baseline underperform relative to models trained directly on the target HotpotQA dataset, our approach exhibits a smaller performance drop and demonstrates greater robustness compared to the baseline.
+## F ROBUSTNESS AGAINST NOISY CONTEXTS
+To evaluate the robustness of our approach against adversarial retrievals and noisy contexts, we constructed a noisy version of the NQ dataset. For each question, we constructed the input context by combining the top-3 passages retrieved by the DPR retriever with 7 randomly selected passages from the Wikipedia corpus to serve as irrelevant/noisy information. This resulted in a context of 10 passages, which were then shuffled to randomize the order. We then compared the performance of our method against the full-document baseline. Experimental results are presented in the table [10.](#page-18-1) In the "full documents" setting, the downstream LLM directly uses all these 10 passages to answer the question, whereas in our method, the compressor first summarizes the context, and the LLM then generates an answer based on the compressed content. The model we used was trained on the standard NQ dataset without any such noise augmentation. Our method not only matches but slightly surpasses the performance of using all documents, demonstrating its strong noise resistance and ability to extract key information from cluttered contexts. In addition, we compared our approach with the RECOMP baseline, and our method consistently outperforms it, reaffirming the superior compression capability and robustness of our model. Furthermore, our method achieves a high compression rate, condensing the source content from 1,427 tokens to just 48.