text string | source string |
|---|---|
and node embeddings, opti- mizing a large-margin ranking loss to minimize the shortest-path distance between predicted and actual parents. •QEN : Wang et al. (2022) propose the Quadru- ple Evaluation Network (QEN), a taxonomy com- pletion framework using term descriptions, pre- trained language models, and code attenti... | https://arxiv.org/abs/2505.19838v1 |
are a supertype of the child concept . Do not add additional comments or information , only return the output in the described format . --- Input description . Context: List of existing parent-child ( supertype-subtype ) relations in the taxonomy . Child: Child concept ( subtype ) that you need to place in a taxonomy .... | https://arxiv.org/abs/2505.19838v1 |
to foods to make them taste sweeter Interpretation: The child concept " sweetening " refers to an additive that enhances the sweetness of food , which is a common theme in the context taxonomy . Reasoning: Let 's think step by step in order to // Output Reasoning: Let 's think step by step in order to identify the chil... | https://arxiv.org/abs/2505.19838v1 |
and then mixing it with sour cream . Parents: dish --- Context: ```powdered sugar , icing sugar flavorer , sugar syrup , sugar syrup sugar , brown sugar sugar , beet sugar ``` Child: granulated sugar Description: granulated sugar is sugar in the form of small grains Reasoning: Let 's think step by step in order to Reas... | https://arxiv.org/abs/2505.19838v1 |
// Output Reasoning: Let 's think step by step in order to find the parents of " sweetening ". We can see that " sweetening " is something added to foods to make them taste sweeter , which is related to the concept of " sweet " as a characteristic of food . Although there is no direct match in the context , we can rela... | https://arxiv.org/abs/2505.19838v1 |
poached egg , chicken marengo , casserole , terrine , macedoine , pizza , meatball , welsh rarebit , osso buco , kishke , chicken paprika , carbonnade flamande , shirred egg , scampi , mold , taco , pork and beans , bitok , french toast , burrito , scrapple , haggis , pheasant under glass , maryland chicken , beef bour... | https://arxiv.org/abs/2505.19838v1 |
Since granulated sugar is a type of sugar , we can look for candidates that are more specific types of granulated sugar . Leaf: No Children: sugarloaf --- Context: ```wine ( Non-Leaf ), blush wine ( Leaf ) wine ( Non-Leaf ), canary wine ( Leaf ) beverage ( Non-Leaf ), wine ( Non-Leaf ) vinegar ( Non-Leaf ), wine vinega... | https://arxiv.org/abs/2505.19838v1 |
, groats , criollo Parent: sweetening Description: sweetening is something added to foods to make them taste sweeter Interpretation: " Sweetening " is an ingredient or substance added to food to make it sweeter , which is a characteristic of some foods . Reasoning: Let 's think step by step in order to // Output Reason... | https://arxiv.org/abs/2505.19838v1 |
of cider that is fermented . Interpretation: " hard cider " is a specific type of cider that is fermented , which makes it an alcoholic drink . Parents: cider --- Context: ```sauce , plum sauce pudding , pease pudding dessert , pudding dish , pudding pudding , carrot pudding ``` Child: plum pudding Description: plum pu... | https://arxiv.org/abs/2505.19838v1 |
taxonomy . Description: Description of the parent concept . Interpretation: Description of the child concept in relation to the taxonomy . --- Follow the following format . Reasoning: Let 's think step by step in order to ${ produce the children }. We ... Leaf: Whether the parent concept should be added as a leaf ( has... | https://arxiv.org/abs/2505.19838v1 |
schnitzel , kabob , beef wellington , risotto , paella , tempura , special , souffle , mousse , fish stick , tostada , frog legs , chili , snack food , ramekin , ham and eggs , boiled egg , chicken provencale , rissole , pilaf , applesauce , moo goo gai pan , kedgeree , stew , tossed salad , molded salad , chicken sala... | https://arxiv.org/abs/2505.19838v1 |
and meatballs , poi , jambalaya , roulade , swiss steak , tamale pie , bacon and eggs , enchilada , barbecue , meat loaf , patty , lobster thermidor , potpie , coquilles saint jacques , sauerbraten , coq au vin , sauerkraut , tetrazzini , moussaka , refried beans , fondue , dolmas , steak au poivre , viand , sukiyaki ,... | https://arxiv.org/abs/2505.19838v1 |
, mother 's milk , acidophilus milk , skim milk , corn sugar , lump sugar , caramel , granulated sugar , sugarloaf , beet sugar , brown sugar Parent: sweetening Description: sweetening is something added to foods to make them taste sweeter Interpretation: Sweetening is something added to foods to make them taste sweete... | https://arxiv.org/abs/2505.19838v1 |
think step by step in order to // Output Reasoning: Let 's think step by step in order to find the most specific parent concepts of " baking ingredients ". We can observe that " baking ingredients " is already a category in the given taxonomy , and it has several subtypes such as " baking powder ", " baking soda ", " p... | https://arxiv.org/abs/2505.19838v1 |
Leaf ) ``` Candidates: spices and seasonings , oils Parent: baking ingredients Description: baking ingredients Interpretation: The child concept " baking ingredients " refers to a category of ingredients used in baking , which is a part of the broader topic of cooking ingredients . Previous Reasoning: Reasoning: Let 's... | https://arxiv.org/abs/2505.19838v1 |
arXiv:2505.19848v1 [cs.CL] 26 May 2025Improving Multilingual Math Reasoning for African Languages Odunayo Ogundepo1,2∗, Akintunde Oladipo1,2∗, Kelechi Ogueji1,2∗, Esther Adenuga1∗, David Ifeoluwa Adelani3,4∗, Jimmy Lin2. 1The African Research Collective,∗Masakhane NLP,2University of Waterloo, 3Mila, McGill University,4... | https://arxiv.org/abs/2505.19848v1 |
math datasets, and monolingual vs multilingual training. Through a series of tar- geted experiments and ablations, we quantify how each approach impacts model performance across dimensions such as reasoning accuracy, resource efficiency, generalization, linguistic coherence, and cross-lingual transfer. Some key finding... | https://arxiv.org/abs/2505.19848v1 |
to generate responses in specific target languages (Aryabumi et al., 2024; Üstün et al., 2024; Odumakinde et al., 2024; Dang et al., 2024), but this not been explored for African languages. Sample efficient methods that aim to squeeze out optimal performance from existing data have also been explored to significant suc... | https://arxiv.org/abs/2505.19848v1 |
generate the personas, and supply text articles from Wikipedia and WURA (Oladipo et al., 2023a). Note that the persona is asked to be generated in English lan- guage and we only use the first 200 words from the article page. The specific prompt details can be found in the prompt in Appendices B.1 and B.2. We deduplicat... | https://arxiv.org/abs/2505.19848v1 |
24.0 65.6 51.6 39.4 31.8 9 Llama 3.1 8b Instruct ✓AfriPersona-Instruct 30,000 19.6 23.6 32.8 50.8 21.6 69.2 53.2 38.7 29.7 10 Llama 3.1 8b Instruct ✗ BigMath Translated 30,000 10.4 9.6 14.0 13.6 7.2 25.2 18.4 11.2 8.96 11 Llama 3.1 8b Instruct ✗ Open Instruct Translated 30,000 46.4 37.6 49.6 64.0 36.0 74.0 52.0 51.4 46... | https://arxiv.org/abs/2505.19848v1 |
response styles across languages and models. Instead, we implemented an LLM-as-judge evaluation frame- work (Stephan et al., 2025), which provides a more robust assessment method. This approach prompts a larger and more capable language model to deter- mine whether a generated response is correct with respect to the gr... | https://arxiv.org/abs/2505.19848v1 |
Ko, 2024). Inspired by these findings, we examine how different loss masking strategies impact model performance in low-resource settings. Specifi- cally, we evaluate the effect of masking instruc- tion and prompt tokens during training loss com- putation. Our investigation aims to determine whether restricting loss ca... | https://arxiv.org/abs/2505.19848v1 |
the intended meaning of the phrases or sentences. These lexical issues, coupled with grammatical errors, led to unnatural phrasing. Some prompts were ambiguous due to various fac- tors, including unrelated adjacent sentences, miss- ing leading questions that would provide context, incorrect word order, and various lexi... | https://arxiv.org/abs/2505.19848v1 |
, 51.4 overall), it is competitive with GPT-4’s performance ( Row 1 , 40.4 overall and 30.9 on African languages). This highlights the effec- tiveness of native-language synthetic data, partic- ularly in closing the gap to frontier models using relatively modest resources. Training on all available data (i.e., both dir... | https://arxiv.org/abs/2505.19848v1 |
to Unseen Languages To evaluate the generalisation, we test performance on a held-out set of African languages not seen dur- ing fine-tuning. Table 3 presents accuracy scores on 10 such languages for various models and fine- tuning settings. We observe that zero-shot performance improves significantly with task-specifi... | https://arxiv.org/abs/2505.19848v1 |
Fadaee, and Sara Hooker. 2024. The multilingual alignment prism: Aligning global and local preferences to reduce harm. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages 12027– 12049, Miami, Florida, USA. Association for Computa- tional Linguistics.Ife Adebara, AbdelRahim E... | https://arxiv.org/abs/2505.19848v1 |
(Volume 1: Long Papers) , pages 575–593, Toronto, Canada. Association for Computational Linguistics. Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plap- pert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training verifiers to ... | https://arxiv.org/abs/2505.19848v1 |
Amar Budhiraja, Kalika Bali, and Monojit Choudhury. 2020. The state and fate of linguistic diversity and inclusion in the NLP world. InProceedings of the 58th Annual Meeting of the Associ- ation for Computational Linguistics , pages 6282–6293, Online. Association for Computational Linguistics. Jared Kaplan, Sam McCandl... | https://arxiv.org/abs/2505.19848v1 |
reinforced evol-instruct. ArXiv preprint , abs/2308.09583. Niklas Muennighoff, Thomas Wang, Lintang Sutawika, Adam Roberts, Stella Biderman, Teven Le Scao, M Saiful Bari, Sheng Shen, Zheng-Xin Yong, Hai- ley Schoelkopf, et al. 2022. Crosslingual general- ization through multitask finetuning. arXiv preprint arXiv:2211.0... | https://arxiv.org/abs/2505.19848v1 |
pages 158–168, Singapore. Association for Computa- tional Linguistics. Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. Advances in neural infor... | https://arxiv.org/abs/2505.19848v1 |
Shubham Toshniwal, Ivan Moshkov, Sean Narenthi- ran, Daria Gitman, Fei Jia, and Igor Gitman. 2024. Openmathinstruct-1: A 1.8 million math instruction tun- ing dataset. arXiv preprint arXiv: Arxiv-2402.10176 . Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soum... | https://arxiv.org/abs/2505.19848v1 |
in Appendix B.6. To ensure high accuracy, we calibrate the prompt using a couple of samples from different languages. B Prompts B.1 Persona Generation Prompt Generate diverse user personas from an African context based on a given piece of text, such as a Wikipedia page or a news article, focusing on individuals who mig... | https://arxiv.org/abs/2505.19848v1 |
or goals. 3. **Create Similar Personas from Different African Communities**: - Maintain the essence of the original persona but adapt cultural, geographic, or linguistic attributes to fit another African context. 4. **Diversity of Contexts**: - Ensure a variety of regions, languages, and cultural perspectives are repre... | https://arxiv.org/abs/2505.19848v1 |
to the created math problem. 4. Provide the prompt JSON format, each with fields: "prompt", "language" 5. Your output should be in the provided language B.5 Math Response Generation Provide a step-by-step solution to the given math problem in the language of the problem and write the final answer in a new line. Note: E... | https://arxiv.org/abs/2505.19848v1 |
IEEE TRANSACTIONS, VOL. XX, NO. X, 20XX 1 Beyond Specialization: Benchmarking LLMs for Transliteration of Indian Languages Gulfarogh Azam, Mohd Sadique, Saif Ali, Mohammad Nadeem, Erik Cambria, Shahab Saquib Sohail, Mohammad Sultan Alam Abstract —Transliteration, the process of mapping text from one script to another, ... | https://arxiv.org/abs/2505.19851v1 |
sentences for 12 South Asian languages. Dakshina enabled baseline systems for single- word and sentence-level transliteration, but its scale (around 300k word pairs) is limited relative to the diversity of Indic languages. Recently,theAksharantarprojectvastlyexpanded available data, releasing a 26-million example trans... | https://arxiv.org/abs/2505.19851v1 |
paper is organized as follows. Section 2 discusses the methodology adopted. Section 3 presents the results and discussion of the current work. Limitations of the study are outlined in Section 4. Section 5 presents the concluding remarks and future directions. 2 Methodology The process employed in our methodology work c... | https://arxiv.org/abs/2505.19851v1 |
The breakdown of word pairs the final test set for each language is as given in Table1. IEEE TRANSACTIONS, VOL. XX, NO. X, 20XX 3 Figure 1: System workflow depicting data preprocessing, model inference, evaluation metrics, and error analysis steps. Language Dakshina AK-Freq AK-NEF AK-NEI Total Bengali 9198 2270 1060 16... | https://arxiv.org/abs/2505.19851v1 |
Mistral-Large 2402 It is a major language model from Mistral AI and is mainly designed to support reasoning as well as instruction-following IEEE TRANSACTIONS, VOL. XX, NO. X, 20XX 4 tasks for complex problems [ 26]. It is a dense, decoder-only transformer model and has 32K token context window. The modelalsoguarantees... | https://arxiv.org/abs/2505.19851v1 |
learning rate multiplier (value = 1.5), remaining parameters were kept at their default values. 2.6 Performance metric Inourstudy,wehaveusedTop-1AccuracyandCharacterError Rate (CER) to identify the model correctness. Top-1 Accuracy refers to the percentage of instances where the most probable output (i.e., the model’s... | https://arxiv.org/abs/2505.19851v1 |
and IndicXlit accuracy = 47.95%). When considering only the general purpose LLMs, we found that GPT-4.5 performed the best followed by the other GPT models (GPT-4o and GPT-4.1 showed almost similar perfor- mance). Gemma-3 and Mistral-Large exhibited similar perfor- manceandwereleastaccurate. Withremarkableperformances,... | https://arxiv.org/abs/2505.19851v1 |
HindiDakshina 69.15 0.093 70.72 0.088 67.26 0.098 52.12 0.149 50.69 0.149 60.07 0.121 71.02 0.087 AK-Freq 50.65 0.112 55.26 0.104 49.88 0.116 36.10 0.157 43.81 0.138 55.60 0.106 56.24 0.098 AK-NEF 60.10 0.109 58.26 0.121 58.87 0.116 55.57 0.126 46.14 0.155 55.32 0.132 60.95 0.104 AK-NEI 70.42 0.074 73.05 0.068 70.85 0.... | https://arxiv.org/abs/2505.19851v1 |
0.090 65.94 0.083 60.14 0.096 51.00 0.124 36.03 0.160 64.11 0.096 66.08 0.083 UrduDakshina 49.96 0.177 54.19 0.165 48.40 0.184 36.36 0.244 36.28 0.232 41.56 0.200 51.31 0.168 AK-Freq - - - - - - - - - - - - - - AK-NEF 57.35 0.114 62.87 0.094 55.39 0.119 52.82 0.126 46.75 0.188 48.65 0.151 57.23 0.113 AK-NEI 53.58 0.112... | https://arxiv.org/abs/2505.19851v1 |
languages, specialized IndicXlit model exhibited better performance and made fewer mistakes (average CER 0.239 in Marathi and 0.264 in Gujarati). Refer to Figure 7for detailed error distribution. 3.4 Robustness The inclusion of reasonable noise in the data enabled us to test how well these models work under real-world ... | https://arxiv.org/abs/2505.19851v1 |
Hindi 30 0.208 35 0.197 12 0.275 41 0.198 Kannada 15 0.264 22 0.232 17 0.217 26 0.241 Malayalam 16 0.318 23 0.328 12 0.341 35 0.261 Marathi 26 0.253 29 0.249 18 0.263 43 0.191 Panjabi 19 0.287 19 0.294 10 0.315 25 0.251 Tamil 19 0.218 22 0.218 20 0.232 37 0.172 Telugu 24 0.235 26 0.213 20 0.225 35 0.205 Urdu 36 0.236 3... | https://arxiv.org/abs/2505.19851v1 |
Mehta, N. Gupta, and V. K. Valloli, “Indian language transliteration using deep learning,” in 2018 IEEE Re- cent Advances in Intelligent Computational Systems (RAICS) . IEEE, 2018, pp. 103–107. [10]Y. Merhav and S. Ash, “Design challenges in named entity transliteration,”in Proceedingsofthe27thInternationalConfer- ence... | https://arxiv.org/abs/2505.19851v1 |
B. Coleman, Y. Gao, B. Mustafa, I. Barr, E. Parisotto, D. Tian, M. Eyal, C. Cherry, J.-T. Peter, D. Sinopalnikov, S. Bhupatiraju, R. Agarwal, M. Kazemi, D. Malkin, R. Kumar, D. Vilar, I. Brusilovsky, J. Luo, A. Steiner, A. Friesen, A. Sharma, A. Sharma, A. M. Gilady, A. Goedeckemeyer, A. Saade, A. Feng, A. Kolesnikov, ... | https://arxiv.org/abs/2505.19851v1 |
arXiv:2505.19862v1 [cs.CL] 26 May 2025REA-RL : Reflection-Aware Online Reinforcement Learning for Efficient Large Reasoning Models Hexuan Deng1Wenxiang Jiao Xuebo Liu1∗Jun Rao1Min Zhang1 1Institute of Computing and Intelligence, Harbin Institute of Technology, Shenzhen, China {hxuandeng,wenxiangjiaonju,rao7jun}@gmail.c... | https://arxiv.org/abs/2505.19862v1 |
miles ... ... to 45 miles away from home. 1974 tokens in total<think>Alright, let's try to figure out how far John is from home after his driving and his trip back. ... ... spent 2 hours in traf fic and 0.5 hours driving at 30 mph, so that's 2.5 hours. So, the remaining time he can spend driving at 80 mph is 4 - 2.5 = ... | https://arxiv.org/abs/2505.19862v1 |
his 4-hour journey . ### Step 1: Calculate the Distance ... ... 3. **Driving at 80 mph for the Remaining Half-Hour :** ... ... = 180 - 55 = 125 \, \text{miles} \] \[ \boxed{125 \text{ miles}} \]103 tokens 403 tokens 15 tokensFigure 1: Overthinking and non-reflective cases from GSM8k. The left column shows the output of... | https://arxiv.org/abs/2505.19862v1 |
These models explicitly generate intermediate reasoning steps, i.e., test-time scaling [4,27], before providing a final answer, incorporating self-reflection and error correction to enhance their performance on complex reasoning tasks [ 20,26]. However, this enhanced capability comes at the cost of slower and more verb... | https://arxiv.org/abs/2505.19862v1 |
similar phenomenon in Deepseek- R1-Distill-Qwen-7B (R1-7B): the model tends to engage in excessive reflection after completing 3 MethodGSM8K Math500 Gaokao23 Amc23 Olympiad Aime24 Average Acc↑ TR↓ Acc↑ TR↓ Acc↑ TR↓ Acc↑ TR↓ Acc↑ TR↓ Acc↑ TR↓ Acc↑ TR↓ Original 91.66 100 1344 92.00 100 3893 81.82 100 3785 88.12 100 5840 ... | https://arxiv.org/abs/2505.19862v1 |
represent revision using our method without and with the gold answer as input, respectively. Our baseline for comparison is “Fixed Trunc ( G)”, which denotes fixed truncation of the original generated response, retaining n%of the thinking tokens and truncating the response at the nearest newline. Subsequently, similar ... | https://arxiv.org/abs/2505.19862v1 |
guiding the direction of online RL optimization, we introduce an efficient reflection model that provides revisions online (§4.2). Finally, we refine the reward to include a reflection reward to penalize non-reflective behavior (§4.3). The workflow is illustrated in Figure 2. 4.1 Online Reinforcement Learning GRPO trai... | https://arxiv.org/abs/2505.19862v1 |
and post- revision responses (highlighted in blue), we can decompose the reward into two components. Formally, letaiandar irepresent the advantages of the original response siand revised response sr i, respectively. The first component is the average reward across all tokens in both responses, formulated as (ar i+ ai)/... | https://arxiv.org/abs/2505.19862v1 |
process, we follow R1 to use a temperature of 0.6 and a top_p of 0.95. Finally, the reflection model is fine-tuned from Qwen2.5-7B-Instruct for one epoch. During sequential revision, we set the temperature to 0 for the reflection model. Online RL training. “MReflect ” represents online revision in §4.2. “ RLen” represe... | https://arxiv.org/abs/2505.19862v1 |
LenMReflect 89.23 37.80 91.20 48.66 78.44 52.23 86.56 59.65 57.19 70.37 41.67 84.15 74.05 58.81 GRPO R RLen+Reflect 92.72 67.07 91.40 69.52 80.52 72.54 85.62 75.64 53.93 83.04 47.92 91.37 75.35 76.53 GRPO R RLen+Reflect MReflect 89.99 53.12 89.60 61.79 81.30 61.19 85.62 71.39 55.11 77.81 42.92 88.45 74.09 73.86Budget: ... | https://arxiv.org/abs/2505.19862v1 |
75.82 76.36 54.15 GRPO R RLen+Reflect MReflect 89.99 52.08 91.00 60.78 82.08 55.85 89.38 67.76 59.11 73.11 51.25 81.09 77.13 65.11 Table 3: Results of our proposed reflection model. “32B Revise” refers to the “Model Revise” results using a 32B LLM from Table 1. “7B Revise” uses our trained 7B revision model. “Fixed Tru... | https://arxiv.org/abs/2505.19862v1 |
of question answering, and the right plot shows the average token consumption per answer. We present results for the GRPO baseline using only a length reward and our three proposed methods. exhibiting minimal performance degradation, demonstrating that our model achieves comparable results at a substantially reduced co... | https://arxiv.org/abs/2505.19862v1 |
length reduction, showing similar trends between accuracy and response length. Both reflection model and reflection reward enhance reflection. Across the three easier datasets, our three approaches demonstrate an average increase of 49% in reflection probability and 7.5% performance improvement compared to GRPO RLen, d... | https://arxiv.org/abs/2505.19862v1 |
poorer parallelism and longer runtime. However, considering our significant improvement and the cost-free improvement of reflection reward, this is acceptable. 10 References [1]Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Sh... | https://arxiv.org/abs/2505.19862v1 |
Pro- ceedings of the Neural Information Processing Systems Track on Datasets and Bench- marks 1, NeurIPS Datasets and Benchmarks 2021, December 2021, virtual , 2021. URL https://datasets-benchmarks-proceedings.neurips.cc/paper/2021/hash/ be83ab3ecd0db773eb2dc1b0a17836a1-Abstract-round2.html . [14] Edward J. Hu, Yelong ... | https://arxiv.org/abs/2505.19862v1 |
test-time scaling. arXiv preprint arXiv:2501.19393 , 2025. URL https://arxiv.org/abs/ 2501.19393 . [28] Tergel Munkhbat, Namgyu Ho, Seo Hyun Kim, Yongjin Yang, Yujin Kim, and Se-Young Yun. Self-Training Elicits Concise Reasoning in Large Language Models. arXiv preprint arXiv:2502.20122 , 2025. URL https://arxiv.org/abs... | https://arxiv.org/abs/2505.19862v1 |
Yang, Beichen Zhang, Binyuan Hui, Bofei Gao, Bowen Yu, Chengpeng Li, Dayiheng Liu, Jianhong Tu, Jingren Zhou, Junyang Lin, et al. Qwen2.5-Math Technical Report: Toward Mathematical Expert Model via Self-Improvement. arXiv preprint arXiv:2409.12122 , 2024. URL https://arxiv.org/abs/2409.12122 . [43] Edward Yeo, Yuxuan T... | https://arxiv.org/abs/2505.19862v1 |
align with the Gold Answer. For each of the {N} parts, please reply in format: [1]. Think: [Explanation for label choice] Label: Reasoning/Right Result/Wrong Result [2]. Think: [Explanation for label choice] Label: Reasoning/Right Result/Wrong Result ... The First Prompt for Detecting Overthinking without Gold Answer *... | https://arxiv.org/abs/2505.19862v1 |
the middle of its response. For each of the {N} parts, please reply in format: [1]. Think/Result [2]. Think/Result ... The Prompt for Math Evaluation Please reason step by step, and put your final answer within \boxed{}. Question: {Question} 16 B Further Analysis B.1 Case Study for REA-RL To illustrate how our method w... | https://arxiv.org/abs/2505.19862v1 |
\boxed{ }. </think> **Final Answer:** The minimum possible value of the sum is \boxed{ }.3638 tokensAdvantage: 0.95-0.45 2307 tokens 35 tokensAdvantage: 0.95 Advantage: 0.95+0.45 Advantage: -1.10Question: Determine the minimum possible value of the sum , where , , and are positive real numbers. Answer: Figure 4: Case S... | https://arxiv.org/abs/2505.19862v1 |
51.10 86.88 63.27 53.93 71.67 48.75 84.92 72.11 58.43 GRPO R Len+Reflect 92.95 89.63 91.60 81.74 81.30 84.68 83.12 87.00 55.11 88.38 44.58 95.52 74.78 87.83 w/D0.1 92.34 80.26 91.60 76.54 79.22 78.87 84.06 82.64 53.78 87.91 40.00 93.25 73.50 83.25 w/D0.4 93.25 108.46 88.00 93.24 80.52 96.51 81.88 92.82 52.59 95.43 37.5... | https://arxiv.org/abs/2505.19862v1 |
5841 54.58 100 9018 78.25 100 4189 Model Reflect 92.42 72.73 93.00 78.17 80.26 75.46 87.19 84.01 57.93 86.77 54.58 94.72 77.56 81.98 + Gold 91.74 68.75 91.60 69.77 79.74 70.51 85.94 77.27 57.19 79.27 54.58 93.70 76.80 76.54 GRPO R LenMReflect 89.23 100 496 92.40 100 1756 79.74 100 1798 88.12 100 3273 60.74 100 4795 47.... | https://arxiv.org/abs/2505.19862v1 |
arXiv:2505.19866v1 [cs.AI] 26 May 2025HS-ST AR: Hierarchical Sampling for Self-Taught Reasoners via Difficulty Estimation and Budget Reallocation Feng Xiong1*, Hongling Xu∗, Yifei Wang1, Runxi Cheng2, Yong Wang1†, Xiangxiang Chu1 1Alibaba Group2Tsinghua University {jingxun.xf,wangyong.lz}@alibaba-inc.com Abstract Self-... | https://arxiv.org/abs/2505.19866v1 |
diffi- culty level of problems are most beneficial for self- taught reasoners? Intuitively, problems that are too simple provide limited learning value, while those that are overly challenging may either waste sam- pling resources by requiring numerous attempts to generate correct responses or be beyond the model’s cap... | https://arxiv.org/abs/2505.19866v1 |
dynamically reallocate sampling budgets toward high-utility prob- lems, significantly enhancing training effec- tiveness under a fixed sampling budget. •Extensive experiments across seven reason- ing benchmarks and various backbone LLMs demonstrate the superiority of our HS-ST AR. Further analyses confirm the effective... | https://arxiv.org/abs/2505.19866v1 |
more candidate responses. Notably, DPO-Boundary-SR significantly achieves the best performance across all benchmarks, with an av- erage score of 38.3%. These results reinforce that boundary-level problems are more sampling- efficient, indicating strategically prioritizing them is key to enhance self-training. 3 Methodo... | https://arxiv.org/abs/2505.19866v1 |
jointly captures the model’s ability to solve a given problem by evaluating both the accuracy of the final answer and the soundness of the reasoning process, thereby providing an effective estimate of problem difficulty even with limited responses. 3.2 Phase 2 : Re-Sampling Building on the insights from Sec. 2, which h... | https://arxiv.org/abs/2505.19866v1 |
first perform a warm-up training using synthetic solutions. Specifically, we leverage the MATH dataset (Hendrycks et al., 2021) and prompt gpt-4o- 2024-08-06 to systematically rewrite each solution in a step-by-step format, then organize these steps separated by " \n\n ". The resulting model, denoted asM0, serves as th... | https://arxiv.org/abs/2505.19866v1 |
has already undergone. Moreover, we find that stepwise initialization not only enables format-consistent reasoning but also outperforms vanilla SFT, demonstrating its effectiveness as a lightweight and generalizable warm-up strategy. 4.3 “Zero Training” of HS-ST AR Settings. The recent emergence of DeepSeek- R1 (DeepSe... | https://arxiv.org/abs/2505.19866v1 |
(Oracle) 45.7 47.0 47.9 STAR-DPO 44.2 45.1 45.7 HS-ST AR-Acc 44.7 46.2 46.8 HS-ST AR-Reward 45.4 46.3 46.7 HS-ST AR 45.6 46.6 47.5 of R1-Zero-like training (Chu et al., 2025; Yu et al., 2025), where reinforcement learning is ap- plied directly to pre-trained models. Following this, we explore a similar “zero training” ... | https://arxiv.org/abs/2505.19866v1 |
of difficulty estima- tion, leading to better coverage of Boundary sam- ples. However, this gain in estimation accuracy introduces a trade-off. As depicted in Fig. 4b, un- der a fixed total sampling budget, allocating more Inlier Boundary Outlier020406080100Iteration 0 76.3 23.7 0.08.477.1 14.5 0.021.578.5 Inlier Bound... | https://arxiv.org/abs/2505.19866v1 |
5.1 Self-Taught Reasoners Recent studies have shown that LLMs can pro- gressively improve themselves by training on self- generated responses using SFT or DPO (Zelik- man et al., 2022; Gulcehre et al., 2023; Huang et al., 2023; Yuan et al., 2024; Li et al., 2025). In mathematical reasoning tasks, response selec- tion i... | https://arxiv.org/abs/2505.19866v1 |
ARrelies on difficulty estimation tech- niques such as reward-guided estimation to identify high-utility problems. Therefore, our framework is inherently tied to mathematical tasks, where problem difficulty is relatively well- defined. This limits the generalizability of HS- STARto other domains where difficulty estima... | https://arxiv.org/abs/2505.19866v1 |
Aditya Siddhant, Alex Ahern, Miaosen Wang, Chenjie Gu, Wolfgang Macherey, Arnaud Doucet, Orhan Firat, and Nando de Freitas. 2023. Reinforced self-training (rest) for language modeling. Preprint , arXiv:2308.08998. Chaoqun He, Renjie Luo, Yuzhuo Bai, Shengding Hu, Zhen Thai, Junhao Shen, Jinyi Hu, Xu Han, Yujie Huang, Y... | https://arxiv.org/abs/2505.19866v1 |
for instruction tuning. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) . Zhong-Zhi Li, Duzhen Zhang, Ming-Liang Zhang, Ji- axin Zhang, Zengyan Liu, Yuxuan Yao, Haotian Xu, Junhao Zheng, Pei-Jie Wan... | https://arxiv.org/abs/2505.19866v1 |
of llms via imagination, searching, and criticizing. In Advances in Neural Information Processing Systems , volume 37, pages 52723–52748. Curran Associates, Inc. Yuxuan Tong, Xiwen Zhang, Rui Wang, Ruidong Wu, and Junxian He. 2024. Dart-math: Difficulty-aware rejection tuning for mathematical problem-solving. InAdvance... | https://arxiv.org/abs/2505.19866v1 |
2024. ReST-MCTS*: LLM self-training via process reward guided tree search. In The Thirty-eighth Annual Conference on Neural Information Processing Systems . Hanning Zhang, Jiarui Yao, Chenlu Ye, Wei Xiong, and Tong Zhang. 2025a. Online-dpo-r1: Unlocking effective reasoning without the ppo overhead. Notion Blog . Jia Zh... | https://arxiv.org/abs/2505.19866v1 |
t,xandRincorr t,x in descending order 20 k←min(|Rcorr t,x|,|Rincorr t,x|) 21 Sample kpairs from top- kresponses: Dpairs t← Dpairs t∪ {(rcorr (i), rincorr (i))}k i=1 22 end 23 Update Mt+1using DPO loss on Dpairs t 24end 25return MT PURE-VR. (Cheng et al., 2025) PURE is a rein- forcement learning approach for LLM fine-tu... | https://arxiv.org/abs/2505.19866v1 |
phase, remain- ing sampling efforts are exclusively focused on Outlier samples for subsequent iterative preference optimization. Re-Sampling on Inlier + Outlier. Following prior difficulty estimation in the pre-sampling phase, re- maining sampling efforts are allocated to both In- lier and Outlier samples for subsequen... | https://arxiv.org/abs/2505.19866v1 |
to other approaches. Inlier Boundary Outlier020406080100Qwen2.5-3B 75.4 24.6 0.05.374.1 20.6 0.014.485.6 Inlier Boundary OutlierPhi-3.5-Mini-Instruct 73.6 26.4 0.06.470.2 23.4 0.010.289.8 Inlier Boundary OutlierQwen2.5-Math-7B 72.7 27.3 0.010.870.8 18.4 0.025.574.5Proportion (%)Inlier (SDE <12.5%) Boundary (SDE 12.5-87... | https://arxiv.org/abs/2505.19866v1 |
ESLM : Risk-Averse Selective Language Modeling for Efficient Pretraining Melis Ilayda Bal1∗Volkan Cevher2,3Michael Muehlebach1 1Max Planck Institute for Intelligent Systems, Tübingen, Germany 2LIONS, EPFL3AGI Foundations, Amazon Abstract Large language model pretraining is compute-intensive, yet many tokens contribute ... | https://arxiv.org/abs/2505.19893v1 |
only high-risk tokens via a value-at-risk threshold. This reshapes the effective training distribution and loss by focusing computational resources on tokens with higher learning value. entropy (Shannon, 1948) or loss, and retains only the highest-risk tokens using value-at-risk (VaR) thresholding. This dynamic filteri... | https://arxiv.org/abs/2505.19893v1 |
2:foreach training iteration k= 1, . . . , K do 3: Sample a batch of tokens B={x1, . . . , x M} ∼ D . 4: Compute per-token statistics Sθk(x): */ Entropy or loss depending on the selection type Sθk(xj) =Hθk(xj)as in (i) ,(VaR -entropy) ℓθk(xj)as in (ii) ,(CVaR -loss) 5: Compute threshold SVaR θk,α←VaR α {Sθk(xj)}M j=1... | https://arxiv.org/abs/2505.19893v1 |
statistical risk that eliminates the need for an auxiliary external selector to improve computational efficiency. ESLM reshapes the loss towards high-risk tokens within each batch. Concretely, the risk is charac- terized by either (i)high predictive uncertainty ( VaR -entropy selection) or (ii)high loss impact (CVaR -l... | https://arxiv.org/abs/2505.19893v1 |
the batch—those with scores in the top (1−α)quantile. This induces an adversarial sub-distribution Qover the batch, supported only on the most challenging tokens. ESLM can then be seen as minimizing the worst-case expected loss over this ambiguity set: min θsup Q∈Pα(B;θ)Exj∼Q[ℓθ(xj)], 4 where the ambiguity set Pα(B;θ)i... | https://arxiv.org/abs/2505.19893v1 |
signal capturing the relative change in CVaR, γ >0controls adaptation rate, and εis a small constant for numerical stability. The core idea for this update rule is stabilizing CVaR: if ∆norm>0(i.e., CVaR increases), the model is encountering harder tokens, αis then decreased to include more tokens and expand the traini... | https://arxiv.org/abs/2505.19893v1 |
trade-off between computational efficiency and robustness, while remaining agnostic to training configurations. Risk-aversion in language modeling. Risk-sensitive optimization offers a principled mechanism to enhance robustness by focusing training on high-risk examples (Rockafellar et al., 2000). The CVaR objective ha... | https://arxiv.org/abs/2505.19893v1 |
focusing optimization on the high-risk tokens, eliminating redundant gradient computation. This efficiency gain holds consistently across model scales. See Appendix E.1 for the convergence of validation loss versus training FLOPs. 0.10 1.00 Log(FLOPs)×10170.400.450.500.550.600.650.70Log(Loss) Validation Loss vs. Traini... | https://arxiv.org/abs/2505.19893v1 |
it still relies on access to curated validation data and per-sample gradient estimation, which becomes impractical at larger model scales due to high memory demands. In contrast, ESLM operates without gradient tracing and scales more naturally. Moreover, GREATS performs selection at the instance level, often discarding... | https://arxiv.org/abs/2505.19893v1 |
(Appendix E.3) confirm that batch- scaled ESLM also converges faster in compute space to lower perplexity than standard CLM. These findings underscore a central trade-off: self-supervised adaptability of ESLM not only reduces redun- dancy but also unlocks efficient scaling in large-batch training by preserving learning... | https://arxiv.org/abs/2505.19893v1 |
the cost of underutilization. We find α∈[0.1,0.2] offers a favorable trade-off between compute savings and generalization. •Model size: Across 124M, 350M, and 774M GPT-2 models, ESLM consistently improves efficiency and generalization (see Figures 2-3, and Appendix E), with gains that are more pronounced in larger mode... | https://arxiv.org/abs/2505.19893v1 |
, 9(3):203–228. Ben-Tal, A., Ghaoui, L., and Nemirovski, A. (2009). Robust Optimization . Princeton Series in Applied Mathematics. Princeton University Press. Bisk, Y ., Zellers, R., Gao, J., Choi, Y ., et al. (2020). PIQA: Reasoning about physical commonsense in natural language. AAAI Conference on Artificial Intellig... | https://arxiv.org/abs/2505.19893v1 |
translation. International Conference on Learning Representations . Jiang, A. H., Wong, D. L.-K., Zhou, G., Andersen, D. G., Dean, J., Ganger, G. R., Joshi, G., Kaminksy, M., Kozuch, M., Lipton, Z. C., et al. (2019). Accelerating deep learning by focusing on the biggest losers. ArXiv , 1910.00762. Joshi, M., Choi, E., ... | https://arxiv.org/abs/2505.19893v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.