text string | source string |
|---|---|
Kt, making exhaustive prompting infeasible. This motivated a retrieval-based approach: selecting a small set of candidate buckets from the current codebook against which an LLM can assess a new idea. Our next prototype achieved this by adding a semantic similarity-based candidate bucket set re- trieval mechanism. We us... | https://arxiv.org/abs/2505.16232v1 |
dropping them as noise or outliers (§5.4). These constraints discourage us from using algo- rithms like DBScan (singleton and rare buckets are likely to be marked as noise) (Ester et al., 1996) and HDBScan (minimum cluster size is 2) (Campello et al., 2013), and our experiments also corrobo- rate their poor performance... | https://arxiv.org/abs/2505.16232v1 |
metric, the MUSERAG models sus- tain human-like performance while the K-means and agglomerative algorithms suffer dramatically and systematically. Specifically, against a human- human AMI of 0.66 [0 .64,0.68], the llama3.3 LLM with CoT prompting achieves the best AMI among the MUSERAG models at 0.59 [0.55,0.64], while ... | https://arxiv.org/abs/2505.16232v1 |
quality ratings ( beaty18 :r= 0 .77 [0.71,0.83], P < 0.001;silvia17 :r= 0.54 [0.41,0.65], P < 0.001;beaty12 :r= 0.42 [0.27,0.55], P < 0.001). The mohr16 dataset contains rating-based originality scores, which correlate strongly with our frequency-based origi- nality scores ( r= 0.42 [0.35,0.49], P < 0.001). This datase... | https://arxiv.org/abs/2505.16232v1 |
paper, future appli- cations of frequency-based originality scoring sys- tems should carefully consider demographic fair- ness and accessibility. Differences in language use across cultural or educational backgrounds may affect bucketing judgments—particularly in non-English settings—potentially introducing un- intende... | https://arxiv.org/abs/2505.16232v1 |
enhance creativity in social networks. arXiv preprint arXiv:2410.15264 . Roger E Beaty and Dan R Johnson. 2021. Automating creativity assessment with SemDis: An open plat- form for computing semantic distance. Behavior Research Methods , 53(2):757–780. Roger E Beaty, Yoed N Kenett, Alexander P Chris- tensen, Monica D R... | https://arxiv.org/abs/2505.16232v1 |
Arts , 15(4):645.Ruth B Ekstrom, John W French, Harry H Harman, and D Dermen. 1976. Manual for kit of factor-referenced tests. Princeton, NJ: Educational Testing Service , 586:1989–1995. Martin Ester, Hans-Peter Kriegel, Jörg Sander, and Xiaowei Xu. 1996. A density-based algorithm for discovering clusters in large spat... | https://arxiv.org/abs/2505.16232v1 |
Ding, and 1 others. 2024b. DALK: Dynamic co-augmentation of LLMs and KG to answer Alzheimer’s disease questions with scien- tific literature. arXiv preprint arXiv:2405.04819 . Tian Liang, Zhiwei He, Wenxiang Jiao, Xing Wang, Yan Wang, Rui Wang, Yujiu Yang, Shuming Shi, and Zhaopeng Tu. 2023. Encouraging divergent think... | https://arxiv.org/abs/2505.16232v1 |
, 52(1):213–221. Yanxin Shen, Lun Wang, Chuanqi Shi, Shaoshuai Du, Yiyi Tao, Yixian Shen, and Hang Zhang. 2024. Com- parative analysis of listwise reranking with large lan- guage models in limited-resource language contexts. arXiv preprint arXiv:2412.20061 . 11 Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Nar... | https://arxiv.org/abs/2505.16232v1 |
28th Inter- national Conference on Intelligent User Interfaces , pages 75–78. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Day- iheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, and 41 others. 2025... | https://arxiv.org/abs/2505.16232v1 |
Prompt (CoT Condition) You are an idea bucket annotator for ideas generated for the object {object_name} in Guilford’s Alternative Uses Test. You will be given an input_idea to annotate against up to {comparison_k} comparison_ideas, given to you in a dictionary format with key-value pairs of comparison_idea_ID: compari... | https://arxiv.org/abs/2505.16232v1 |
truths. Values denote mean ± half-width of the 95% C.I. (N= 109 ). Model AMI NMI Pearson’s rSpearman’s ρ ICC (3,1) llama3.3 CoT 0.57±0.04 0.84±0.02 0 .76±0.08 0 .74±0.09 0.74±0.09 qwen3 CoT 0.54±0.04 0 .83±0.02 0 .74±0.09 0 .73±0.09 0 .74±0.09 phi4 CoT 0.56±0.03 0 .79±0.01 0 .67±0.10 0 .68±0.10 0 .67±0.10 llama3.3 Base... | https://arxiv.org/abs/2505.16232v1 |
arXiv:2505.16234v1 [cs.CL] 22 May 2025LIFEB ENCH : Evaluating Length Instruction Following in Large Language Models Wei Zhang1∗, Zhenhong Zhou1∗, Junfeng Fang3∗, Rongwu Xu4∗, Kun Wang2∗, Yuanhe Zhang1,Rui Wang1,Ge Zhang5,Xinfeng Li2, Li Sun6,Lingjuan Lyu7,Yang Liu2,Sen Su1† 1BUPT,2NTU,3NUS,4THU,5M-A-P,6NCEPU,7Sony AI A... | https://arxiv.org/abs/2505.16234v1 |
follow longer length instructions, we formulate our investigation around three core research questions (RQs): •RQ1: To what extent do current LLMs accurately follow explicit length instructions, and what factors may lead to length instructions following failures? •RQ2: How capable is the current LLM at outputting long ... | https://arxiv.org/abs/2505.16234v1 |
generate extended texts [ 52,68]. Various benchmarks have been proposed to evaluate comprehension over long contexts, including those based on synthetic data [99,56,65,43] and human-annotated real-world tasks [ 6,7,59,92,3]. Long-text generation, which is a classic task [ 10,19,104] in natural language generation (NLG)... | https://arxiv.org/abs/2505.16234v1 |
evaluation metrics allow comparison across models by quantifying both the magnitude and directionality of deviations. Section 4.3 provides a detailed discussion of these metrics. Data Collection and Statistics. Selecting NLG tasks tailored to evaluating length instruction following is crucial for constructing LIFEB ENC... | https://arxiv.org/abs/2505.16234v1 |
•Creative Generation Task: The Creative Generation task requires the model to produce creative text, such as academic papers, novels, and technical reports. The complexity and richness of the content can be adjusted to accommodate different length constraints. We collect data from seven datasets [ 82,75,31,42,51] and c... | https://arxiv.org/abs/2505.16234v1 |
count of the output. Length Score. The effectiveness of Length Deviation may be degraded when aggregating multiple samples, as positive and negative deviations can offset each other. Furthermore, Length Deviation is sensitive to outliers, where a few extreme values can disproportionately skew the aggregated results. To... | https://arxiv.org/abs/2505.16234v1 |
23% 57.7 89.9 89.6 GLM-4-9B-Chat 9B ✗ 40% 35.2 95.9 68.3 Mistral-7B-Instruct-v0.2 7B ✗ 84% 26.7 85.9 63.0 Long-Text Enhanced Models LongWriter-Llama3.1-8B 8B ✗ 102% 31.9 84.1 79.0 LongWriter-GLM4-9B-Chat 9B ✗ 52% 37.4 90.9 86.8 Suri-I-ORPO 7B ✗ 506% 27.4 79.7 75.2 6 Proprietary Models Open-Source Models Long-Text Enhan... | https://arxiv.org/abs/2505.16234v1 |
instruction requests outputs up to the model’s capacity. We exclude summarization tasks from consideration as they do not provide sufficiently long inputs for meaningful evaluation 7 0Token Count40000 30000 20000 1000060000 50000 Claude-3.7-Sonnet DeepSeek-R1DeepSeek-V3Gemini-2.0-Flash Gemini-2.0 -Flash-ThinkingGemini-... | https://arxiv.org/abs/2505.16234v1 |
do not, suggesting that it assumes compliance rather than measuring the true output length. Experimental details and comprehensive results for all models are provided in Appendix F. Sensitivity to Input Length. Section 5.2 reveals that summarization tasks yield the lowest perfor- mance. Since inputs of such tasks are t... | https://arxiv.org/abs/2505.16234v1 |
constraints. 6 Conclusion, Limitations and Future Insights In this paper, we introduce LIFEB ENCH , a comprehensive benchmark for evaluating the ability of LLMs to follow length instructions across diverse tasks, languages, and a broad range of length constraints. Empirical analysis shows that current LLMs are generall... | https://arxiv.org/abs/2505.16234v1 |
large language models. Natural Language Processing Journal , page 100143, 2025. [12] ByteDance. Doubao-1.5-Pro. https://seed.bytedance.com/zh/special/doubao_1_ 5_pro/ , 2025. [13] Y . Chang, X. Wang, J. Wang, Y . Wu, L. Yang, K. Zhu, H. Chen, X. Yi, C. Wang, Y . Wang, et al. A survey on evaluation of large language mod... | https://arxiv.org/abs/2505.16234v1 |
Greene. Sms spam filtering: Methods and data. Expert Systems with Applications , 39(10):9899–9908, 2012. [26] L. Dong, J. Mallinson, S. Reddy, and M. Lapata. Learning to paraphrase for question answering. InProceedings of the 2017 Conference on Empirical Methods in Natural Language Processing , pages 875–886, 2017. [27... | https://arxiv.org/abs/2505.16234v1 |
read and comprehend. Advances in neural information processing systems , 28, 2015. [43] C.-P. Hsieh, S. Sun, S. Kriman, S. Acharya, D. Rekesh, F. Jia, and B. Ginsburg. Ruler: What’s the real context size of your long-context language models? In First Conference on Language Modeling , 2024. [44] J. Huang and K. C.-C. Ch... | https://arxiv.org/abs/2505.16234v1 |
W. Xu. Ai awareness. arXiv preprint arXiv:2504.20084 , 2025. [61] Z. Li, X. Jiang, L. Shang, and H. Li. Paraphrase generation with deep reinforcement learning. InProceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , pages 3865–3878, 2018. [62] X. Liang, H. Wang, Y . Wang, S. Song, J. ... | https://arxiv.org/abs/2505.16234v1 |
American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages 839–849, 2016. [76] R. Nallapati, B. Zhou, C. dos Santos, C. Gulcehre, and B. Xiang. Abstractive text summa- rization using sequence-to-sequence rnns and beyond. In Proceedings of the 20th SIGNLL Conference on Computa... | https://arxiv.org/abs/2505.16234v1 |
2020. [95] I. Sutskever, O. Vinyals, and Q. V . Le. Sequence to sequence learning with neural networks. Advances in neural information processing systems , 27, 2014. [96] A. Talmor, J. Herzig, N. Lourie, and J. Berant. Commonsenseqa: A question answering challenge targeting commonsense knowledge. In Proceedings of the ... | https://arxiv.org/abs/2505.16234v1 |
Methods in Natural Language Processing , pages 4393–4479, 2022. [109] Z. Yang, F. Xiong, P. Li, and Y . Liu. Rethinking long context generation from the continual learning perspective. In Proceedings of the 31st International Conference on Computational Linguistics , pages 1922–1933, 2025. [110] S. Yao, H. Chen, A. W. ... | https://arxiv.org/abs/2505.16234v1 |
of NLP Tasks to Length Paradigms . . . . . . . . . . . . . . . . . . . . . 22 B.2 Data Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 B.3 Refinement Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 B.4 Examples of Tasks . . . . . . . . . . . . . . . .... | https://arxiv.org/abs/2505.16234v1 |
. . . . . . . . . . . . . . 46 F.2 Marker-Based Length Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . 46 G Details of Input Characteristic Analysis 51 G.1 Task Type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51 G.2 Input Length . . . . . . . . . . . . . . . . . . . . .... | https://arxiv.org/abs/2505.16234v1 |
Improving Length Instruction Following During Post-training . . . . . . . . . . . . 79 N Length Instruction Following under Output Format Constraints 80 19 A Comparison of LIFEB ENCH with Existing Benchmarks In this section, we review ten representative benchmarks on long-form text generation or long-context modeling a... | https://arxiv.org/abs/2505.16234v1 |
content scenarios, typically including explicit length instructions as part of their evaluation setup. We contrast LIFEB ENCH with seven representative benchmarks in this area. Existing long-text benchmarks generally prioritize generation quality in extended contexts but neglect comprehensive evaluations across shorter... | https://arxiv.org/abs/2505.16234v1 |
A defining characteristic of these tasks is their adaptability to different use requirements. Representative examples include open-ended question answering and text completion. In open-ended question answering, models can extend answers with additional explanations, while in text writing, they can generate outputs rang... | https://arxiv.org/abs/2505.16234v1 |
associated representative NLP tasks, followed by a comprehensive overview of the data collection process, including dataset composition and provenance, preprocessing and data cleaning strategies, and the subtypes represented within each category. Question Answering (QA) Task. This task involves answering open-ended que... | https://arxiv.org/abs/2505.16234v1 |
shorter summaries whose lengths can vary considerably—from concise overviews to more detailed condensations—aligning naturally with the Flexible-Length Paradigm . Representative NLP tasks include Text Summarization andSentence Compression , both aiming to preserve key information while reducing text length. To support ... | https://arxiv.org/abs/2505.16234v1 |
control output length meaningfully, while the latter is designed as a multiple-choice task, constraining the response format and limiting the scope for open-ended explanations. Nonetheless, reasoning itself inherently suits the Flexible- Length Paradigms : some inferences can be expressed succinctly, while others benef... | https://arxiv.org/abs/2505.16234v1 |
it for Prompt-based Creative Writing . •CNN/DailyMail [ 42]: News summaries from articles; we applie it to News Writing , prompting models to reconstruct or expand original news content. The Chinese data are drawn from the following three datasets: 10https://www.imsdb.com/ 11https://www.reddit.com/r/WritingPrompts/ 25 ... | https://arxiv.org/abs/2505.16234v1 |
component specifies the task type and the desired model generation, while the [Requirement] sets the length constraints of the output. This dual- component framework precise control over both task appropriateness and output flexibility. For each task category or subtype, unique [Instruction] and[Requirement] templates ... | https://arxiv.org/abs/2505.16234v1 |
novels. An Example Data of Summarization Task [Instruction] Summarize this book: I was born in Tuckahoe, near Hillsborough, and about twelve miles from Easton, in Talbot county, Maryland. I have no accurate knowledge of my age, never having seen any authentic record containing it. By far the larger part of the slaves k... | https://arxiv.org/abs/2505.16234v1 |
they sometimes rely on LLMs to create drafts. An Example Data of Creative Generation Task [Instruction] You are given an excerpt from a script: PEDDLER: Oh I come from a land From a faraway place Where the caravan camels roam Where they cut off your ear /Where it’s flat and immense 29 If they don’t like your face /And ... | https://arxiv.org/abs/2505.16234v1 |
(Text type: Review ). •Text Classification on Emails [ 94]: A dataset of email exchanges among journalists; 9 samples were selected (Text type: Email ). •Hate Speech [ 24]: Tweets collected from Twitter15containing hate speech; 6 samples were selected (Text type: Tweets ). The Chinese label-task data were collected fro... | https://arxiv.org/abs/2505.16234v1 |
words long. Table 8: Refinement Template for Label. C.2 LIFEB ENCH -REFACTOR For tasks under the Adaptive-Length Paradigm , where the output length is closely tied to the input length, we categorize them as Refactor Tasks . This category includes classic natural language generation tasks such as Machine Translation ,Te... | https://arxiv.org/abs/2505.16234v1 |
Commentary Writing 2 Total (Creative Generation) 20 33 D Detailed Experimental Setup In this section, we describe the experimental settings in detail, some of which are not presented due to the page limitation of the main paper. In Section D.1, we introduce the model types, sources, input and output window sizes used i... | https://arxiv.org/abs/2505.16234v1 |
newline and tab characters. Formally, for a given output, we first count the number of Chinese characters using the Unicode range [\u4e00-\u9fff] , and then count English words via the regular expression \b[a-zA-Z0-9’]+\b . The final word count is the sum of these two components, which allows us to maintain excellent c... | https://arxiv.org/abs/2505.16234v1 |
ranging from 0 to 100. The computation of LSvaries 35 according to the specific constraint type as follows: LSE=100×ek1·LD, ifLD < 0 100×e−k2·LD,ifLD≥0, LSM=100, ifLD < 0 100×e−k2·LD,ifLD≥0, LSL=100×ek1·LD,ifLD < 0 100, ifLD≥0,(3) where LSE,LSM, and LSLcorrespond to the Equal To ,At Most , and At Least control metho... | https://arxiv.org/abs/2505.16234v1 |
As depicted, all three constraint methods yield a perfect score when outputs fully comply with the length instructions. Under the Equal To setting, an LDof−50% yields a score of 8.2, whereas an LDof+50% results in a higher score of 22.3, clearly demonstrating the intended asymmetric treatment At extreme deviations ( e.... | https://arxiv.org/abs/2505.16234v1 |
Score of 53.6, reflecting a failure to follow length instructions in certain cases—specifically, generating overlong outputs that substantially inflate Length Deviation . This highlights a key limitation in length instruction following. For length constraints greater than 2048, almost all models reach Length Score scor... | https://arxiv.org/abs/2505.16234v1 |
55.1 53.4 49 30 32.1 42.5 42.1 21.9 44 LD 35% 27% 31% 27% 30% 97% 119% 70% 43% 47% 53% Gemini-2.5-ProLS 80.9 76.1 74.4 72.8 67.4 49.5 37 48 54.3 39.4 60 LD 17% 9% 10% 12% 15% 40% 57% 44% 37% 35% 28% Doubao-1.5-ProLS 31.4 42.6 46.8 65.5 64.2 60.6 63.5 57.3 42.4 12.3 48.7 LD 30% 24% 21% 12% 12% 14% 13% 24% 27% 52% 23% Do... | https://arxiv.org/abs/2505.16234v1 |
98.5 89.2 91.6 94.5 100 100 100 100 97 LD 1% 1% 1% 7% 5% 3% 0% 0% 0% 0% 2% o1-miniLS 95.5 92.2 91.1 85 75.5 75.4 86 96.2 99.7 100 89.7 LD 4% 5% 6% 11% 19% 18% 10% 3% 0% 0% 8% o3-miniLS 99.2 99.2 99.7 100 99.9 99.5 97.9 99.5 100 100 99.5 LD 0% 0% 0% 0% 0% 0% 1% 0% 0% 0% 0% Claude-3.7-SonnetLS 90.5 92.1 93.9 94.3 99.5 99... | https://arxiv.org/abs/2505.16234v1 |
85.9 LD 136% 85% 40% 77% 21% 10% 23% 21% 5% 2% 42% LongWriter-Llama3.1-8BLS 63.2 70 75.4 85.7 88.9 95.6 95.9 85.2 86.4 94.2 84.1 LD 713% 66% 31% 34% 9% 22% 63% 65% 35% 9% 105% LongWriter-GLM4-9BLS 79.9 90.1 97 97.1 85.5 86 95.7 93.4 88.8 95.8 90.9 LD 61% 23% 3% 2% 13% 10% 11% 13% 14% 4% 15% Suri-I-ORPOLS 53.6 70.7 78.1... | https://arxiv.org/abs/2505.16234v1 |
84.2 LD 0% 0% 0% 0% 0% 0% -3% -16% -19% -43% -8% Qwen3-32BLS 100 99.7 98.3 99.8 99.7 97.9 97.4 86.5 63.1 34.3 87.4 LD 0% 0% 0% 0% 0% -1% -1% -5% -13% -28% -5% Qwen3-32B-ThinkingLS 100 100 100 100 100 100 98 77.5 65.3 37.2 87.8 LD 0% 0% 0% 0% 0% 0% -1% -7% -11% -24% -4% Qwen3-235B-A22BLS 99.6 99.5 98.9 98.5 99.4 98.1 95... | https://arxiv.org/abs/2505.16234v1 |
struggle with very short or ultra-long constraints. Even long-text enhanced models often fail to meet these extremes and may sacrifice short length instruction following ability. These results show that precise length instruction following—especially at the extreme constraints—remains an open problem for current LLMs. ... | https://arxiv.org/abs/2505.16234v1 |
the maximum output length claimed by models often differs substantially from actual performance in ultra-long generation scenarios. 44 Table 14: Length Deviation andLength Score for extended length constraints. Models MetricEqual To At Most At Least 16k 32k 16k 32k 16k 32k GPT-4o miniLD -94% - -95% - -94% - LS 1.1 - 10... | https://arxiv.org/abs/2505.16234v1 |
an intuitive indication of whether LLMs are aware that their outputs fail to meet length instructions. If the generated output is objectively insufficient yet the model reports it as adequate, this suggests a lack of length awareness or a form of deceptive reporting. In contrast, if the model acknowledges the insuffici... | https://arxiv.org/abs/2505.16234v1 |
/uni00000035/uni00000048/uni00000044/uni0000004f/uni00000003/uni00000029/uni0000004c/uni00000057 /uni00000035/uni00000048/uni00000053/uni00000052/uni00000055/uni00000057/uni00000048/uni00000047/uni00000003/uni00000029/uni0000004c/uni00000057 /uni00000035/uni00000048/uni00000049/uni00000048/uni00000055/uni00000048/uni00... | https://arxiv.org/abs/2505.16234v1 |
Figure 13: Empirical Cumulative Distribution Function (ECDF) of marker positions in generated sequences: The x-axis represents fixed-length bins partitioning the generated output into five equal- length segments (each covering 20% of the total sequence). The y-axis reports the cumulative proportion of markers that fall... | https://arxiv.org/abs/2505.16234v1 |
addition, QA tasks demonstrate the highest average length scores overall, indicating that models find it comparatively easier to manage length constraints in scenarios involving direct answers rather than extensive text condensation. /uni0000002a/uni00000033/uni00000037/uni00000010/uni00000017/uni00000052/uni00000003/u... | https://arxiv.org/abs/2505.16234v1 |
different length constraints under the Equal To control method. /uni0000002a/uni00000033/uni00000037/uni00000010/uni00000017/uni00000052/uni00000003/uni00000050/uni0000004c/uni00000051/uni0000004c/uni0000002a/uni00000033/uni00000037/uni00000010/uni00000017/uni00000052/uni00000052/uni00000014/uni00000010/uni00000050/uni... | https://arxiv.org/abs/2505.16234v1 |
22% 23% -1% -38% -65% -83% o1-miniLD-EN 3% 7% 8% 9% 13% 28% 56% 28% -51% -85% LD-CN -6% -5% 3% 29% 23% 17% 10% -21% -50% -78% o3-miniLD-EN 1% 1% 1% 0% 1% 1% 5% 7% 4% -80% LD-CN -1% -2% -6% -9% -10% -9% -6% 1% -8% -47% Claude-3.7-SonnetLD-EN 41% -1% -5% 3% 1% -6% -13% -23% -36% -48% LD-CN 109% 38% 30% 29% 2% -5% -14% -1... | https://arxiv.org/abs/2505.16234v1 |
Suri-I-ORPOLD-EN 467% 250% 74% 153% 88% 47% 29% -3% -43% -67% LD-CN 3808% 1654% 1017% 552% 513% 312% 297% 107% 27% -27% 57 H Lazy Strategy Analysis In this section, we further identify and analyze several Lazy Generation Strategies that language models employ when instructed to produce long-length outputs. Such lazy st... | https://arxiv.org/abs/2505.16234v1 |
the storyline. Please let me know how you would like to proceed! (Generated By GPT-4o) Example 3: 抱歉，我无法满足该请求。(Generated By GPT-4o mini) Example 4: 抱歉，根据当前平台的限制，我无法一次性生成8192字的社会评论。不 过，我可以为您提供一篇详细且有深度的评论，涵盖您提供的信息。如果您需要， 请让我知道！(Generated By o1-mini) Refusal with Attempt. The following examples illustrate Refusal with Att... | https://arxiv.org/abs/2505.16234v1 |
Attempt responses, though these occurrences remain infrequent. Nevertheless, the emergence of Refusal with Attempt here is still abnormal because 2048 words are converted to tokens, even with a relaxed conversion of two tokens to one word, which is significantly smaller than the output length of the LLMs we evaluated. ... | https://arxiv.org/abs/2505.16234v1 |
a detailed case study on the creative generation task. Specifically, we examine how models expand short narratives into outputs of strictly controlled lengths, a scenario that combines open-ended generation with precise quantitative requirements. For this analysis, we select Claude-3.7-Sonnet-Thinking . Although top-pe... | https://arxiv.org/abs/2505.16234v1 |
the beach, but after traffic delays, found only high surfer waves. " Count: 1. We 2. packed 3. snorkels 4. for 5. the 6. beach 7. but 8. after 9. traffic 10. delays 11. found 12. only 13. high 14. surfer 15. waves That’s still 15 words. Let me add one more meaningful word: "We packed snorkels for the perfect beach day,... | https://arxiv.org/abs/2505.16234v1 |
day for the beach. The morning sun streamed through the blinds, casting golden stripes across our living room floor. Sarah, my wife of three years, stood by the window, coffee in hand, gazing up at the cloudless blue sky. "What do you think? " she asked, turning to me with that hopeful smile I’d fallen in love with bac... | https://arxiv.org/abs/2505.16234v1 |
son, Ethan, six years old and bursting with energy, bounded into the room wearing his favorite shark-patterned swim trunks. (19) 8. "Are we going to the ocean? Please say yes! " (9) 9. Sarah and I exchanged glances. (6) 10. We had both taken Friday off for a long weekend, with no concrete plans. .... (1070 words) Total... | https://arxiv.org/abs/2505.16234v1 |
plan the expanded story structure: 1. Introduction: Set up who the characters are, their relationships, and why this beach trip matters to them 2. Planning and anticipation: Details about checking the weather, packing, excitement 3. The journey: Expanded section on the traffic and travel 4. Arrival: First impressions o... | https://arxiv.org/abs/2505.16234v1 |
Maya, despite still battling occasional fatigue from her treatments, had thrown herself into researching the marine life we might encounter while snorkeling. Each night, she’d send us excited texts about the different species of fish that inhabited the coastal waters, along with fascinating facts about their behaviors ... | https://arxiv.org/abs/2505.16234v1 |
been best friends since our freshman year of college when we were randomly assigned as roommates. Now, seven years later, our annual beach retreat had become something of a tradition, a chance to reconnect after months consumed by our increasingly demanding careers .... (8961 words in total) Long Constraint (8192 words... | https://arxiv.org/abs/2505.16234v1 |
fundamental limitations in present-day models’ capacity for precise length controllability, particularly within extended or complex generation tasks. /uni0000002a/uni00000033/uni00000037/uni00000010/uni00000017/uni00000052/uni00000003/uni00000050/uni0000004c/uni00000051/uni0000004c/uni0000002a/uni00000033/uni00000037/u... | https://arxiv.org/abs/2505.16234v1 |
even if not all minor aspects are covered. Deduct points only if the response goes off-topic, contains irrelevant or repeated information, or fails to respond to the main point of the user’s request. Score from highly relevant and fully applicable to completely irrelevant or inapplicable. 2. Accuracy: Score from conten... | https://arxiv.org/abs/2505.16234v1 |
multiple con- straints [ 103,41]. This raises an important question: do models resort to shortcuts—such as repetition or degenerate patterns—to satisfy the length constraint at the cost of faithfully following the original instruction, particularly under long-length requirements? Additionally, analyzing how output qual... | https://arxiv.org/abs/2505.16234v1 |
models typically outperform those in the Increase then Decrease group in length instruction following, as shown in Table 1, supporting the hypothesis that stronger instruction-following capabilities lead to more robust long-text generation. •Low Long-Text Quality : These models show a significant drop in quality under ... | https://arxiv.org/abs/2505.16234v1 |
to contain richer content, better structure, and more developed reasoning, which collectively contribute to higher quality. These results also explain why our evaluation does not directly compare generations from the same model under different length constraints—such comparisons could unfairly penalize shorter generati... | https://arxiv.org/abs/2505.16234v1 |
with disorganized or erratic distributions, reflecting less reliable behavior across constraints. K.2 Text Quality Comparison of Long-text Generation In Section K.1, we analyzed how the same model performs under varying length constraints and observed that generation quality often decreases under longer constraints. To... | https://arxiv.org/abs/2505.16234v1 |
8.50 9.30 8.30 8.30 7.80 8.05 8.90 7.95 7.90 8.15 LongWriter-Llama3.1-8B 8.30 9.25 7.85 8.10 6.45 8.60 9.10 7.75 7.90 6.90 6.85 8.30 6.25 6.75 5.95 7.45 8.50 7.10 7.05 7.45 Mistral-7B-Instruct-v0.2 7.70 7.20 7.35 7.20 6.55 7.50 7.35 7.25 7.55 6.50 7.20 7.50 7.40 7.45 6.45 7.60 7.30 7.50 7.60 6.70 o1-mini 9.40 9.40 9.10... | https://arxiv.org/abs/2505.16234v1 |
their respective versions and overall Arena Scores on the Chatbot Arena LLM Leaderboard. Model Chatbot Arena Model Version Overall Arena Score1 GPT-4o mini gpt-4o-mini-2024-07-18 1272 GPT-4o gpt-4o-2024-08-06 1265 o1-mini o1-mini 1303 o3-mini o3-mini 1305 Claude-3.7-Sonnet claude-3-7-sonnet-20250219 1290 Claude-3.7-Son... | https://arxiv.org/abs/2505.16234v1 |
up to 2048 words, lower EoS logits or probabilities correlate strongly with longer generations. However, some anomalies appear in other models: Llama-3.1-8B-Instruct shows this inverse relationship clearly only in probability, while Qwen2.5-7B-Instruct primarily demonstrates it in logits. Additionally, at very long con... | https://arxiv.org/abs/2505.16234v1 |
and assessing inherent pre-trained capabilities directly. We conducted experiments on 360 fundamental data units under an 8192-word constraint with the Equal To control method. Four representative open-source base models were tested: Llama-3.1-8B ,GLM-4-9B ,Qwen2.5-7B , and Mistral-7B-v0.1 . The experimental results ar... | https://arxiv.org/abs/2505.16234v1 |
mod- els evaluated, we observe substantial im- provements in Length Score scores, with in- creases consistently exceeding 57 points and the highest achieving 76.4—substantially outperforming the top-performing model, Gemini-2.5-Pro (39.4, as reported in Ta- ble 11). Correspondingly, significant reduc- tions in Length D... | https://arxiv.org/abs/2505.16234v1 |
(↓6.5) 38.5 (↓1.8) 20.2 (↓14.0) Gemini-2.0-Flash 44.2 (↓15.6) 51.6 (↓8.1) 34.0 (↓4.2) 51.2 (↑3.5) 52.2 (↓0.5) Gemini-2.0-Flash-Thinking 51.0 (↓2.4) 52.1 (↑3.1) 56.6 (↑26.6) 47.0 (↑14.9) 28.6 (↓13.9) Gemini-2.5-Pro 67.1 (↓5.7) 63.7 (↓3.7) 54.2 (↑4.7) 35.4 (↓1.6) 30.6 (↓17.4) Tables 20, Table 21, and Table 22 summarize t... | https://arxiv.org/abs/2505.16234v1 |
arXiv:2505.16237v1 [cs.CL] 22 May 2025Align-GRAG: Reasoning-Guided Dual Alignment for Graph Retrieval-Augmented Generation Derong Xu1,2∗, Pengyue Jia2∗, Xiaopeng Li2, Yingyi Zhang2, Maolin Wang2, Qidong Liu2, Xiangyu Zhao2,Yichao Wang3,Huifeng Guo3,Ruiming Tang3,Enhong Chen1,Tong Xu1 1University of Science and Technolo... | https://arxiv.org/abs/2505.16237v1 |
graph- based RAG (GRAG) [ 14,18–20] as an enhancement to traditional RAG methods. Instead of retrieving isolated text chunks where information is treated independently of its structural relationships, GRAG retrieves subgraphs from databases [ 18]. These subgraphs maintain the interconnected relationships between knowle... | https://arxiv.org/abs/2505.16237v1 |
knowledge, enabling the generator to produce accurate, context-aware responses grounded in the graph’s structure. Extensive experiments are conducted on GraphQA benchmark [ 18], covering tasks such as commonsense reasoning, scene graph understanding, and knowledge graph reasoning, which demonstrate the superior perform... | https://arxiv.org/abs/2505.16237v1 |
generates a reasoning chain that links the query to the answer, identifying key intermediate nodes and bridging information gaps. Next, we propose the graph aligner to achieve dual alignment with this summarized reasoning chain: (1) Graph Node Alignment : A KL divergence loss [ 32] highlights critical nodes and reasoni... | https://arxiv.org/abs/2505.16237v1 |
Rowling . We believe this can serve as an excellent label. The deeper insights provided through reasoning are subsequently utilized in the dual node and representation alignment process to train the aligner module. More details for summarization prompt are provided in Appendix G. 3.2.2 Node Alignment The Aligner module... | https://arxiv.org/abs/2505.16237v1 |
in Section 4.4. 3.2.4 Graph Pruning and Representation Generation We use the trained Graph Aligner to prune irrelevant nodes, and produce graph representations that are aligned with language. The Aligner performs graph pruning across the entire dataset. To start, we introduce a hyperparameter: the number of seed nodes,... | https://arxiv.org/abs/2505.16237v1 |
±0.0095 0.5378 ±0.0044 0.7373 ±0.0064 0.5251 ±0.0052 gte-large [22] 0.8776 ±0.0095 0.8592 ±0.0074 0.5392 ±0.0013 0.7340 ±0.0044 0.5374 ±0.0038 bge-reranker-base [21] 0.8534 ±0.0159 0.8577 ±0.0029 0.5323 ±0.0052 0.7397 ±0.0012 0.5254 ±0.0010 bge-reranker-large [21] 0.8612 ±0.0184 0.8644 ±0.0060 0.5366 ±0.0045 0.7391 ±0.... | https://arxiv.org/abs/2505.16237v1 |
employs cross-modality pooling to fuse text and graph representations. Despite these efforts, their approaches remain limited. In contrast, Align-GRAG explicitly aligns the representation of graphs with language, effectively bridging the gap between these two spaces and achieving state-of-the-art results. 4.2 Ablation ... | https://arxiv.org/abs/2505.16237v1 |
reasoning over graph structures. 4.4 Evaluation of Representation Alignment 024820 40 60 80 100 Align Degree (Step)4 3 2 1 012Cosine-Similarity-score Aligned Unaligned (a) Align with Query 024820 40 60 80 100 Align Degree (Step)020406080Cosine-Similarity-score Aligned Unaligned (b) Align with Summarization 024820 40 60... | https://arxiv.org/abs/2505.16237v1 |
[ 59,31,60] have incorporated GNNs by feeding graph tokens into LLMs. Beyond basic integration, recent studies [ 61–63] have focused on tighter architectural fusion, embedding graph neural layers within transformer architec- tures to enable seamless interaction between graph reasoning and natural language understanding... | https://arxiv.org/abs/2505.16237v1 |
Alexander H Miller, and Sebastian Riedel. Language models as knowledge bases? arXiv preprint arXiv:1909.01066 , 2019. [8]Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qiang- long Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, et al. A survey on hallucination in large language models: Prin... | https://arxiv.org/abs/2505.16237v1 |
Dingkun Long, Pengjun Xie, and Meishan Zhang. Towards general text embeddings with multi-stage contrastive learning. arXiv preprint arXiv:2308.03281 , 2023. [23] Pengyue Jia, Derong Xu, Xiaopeng Li, Zhaocheng Du, Xiangyang Li, Xiangyu Zhao, Yichao Wang, Yuhao Wang, Huifeng Guo, and Ruiming Tang. Bridging relevance and ... | https://arxiv.org/abs/2505.16237v1 |
Adriana Romero, Pietro Liò, and Yoshua Bengio. Graph attention networks. In International Conference on Learning Representations , 2018. [39] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint a... | https://arxiv.org/abs/2505.16237v1 |
IEEE transactions on neural networks and learning systems , 32(1):4–24, 2020. [55] Chao Huang, Xubin Ren, Jiabin Tang, Dawei Yin, and Nitesh Chawla. Large language models for graphs: Progresses and directions. In Companion Proceedings of the ACM Web Conference 2024 , WWW ’24, page 1284–1287, New York, NY , USA, 2024. A... | https://arxiv.org/abs/2505.16237v1 |
and relational knowledge bases. arXiv preprint arXiv:2404.13207 , 2024. [70] Tianyu Fan, Jingyuan Wang, Xubin Ren, and Chao Huang. Minirag: Towards extremely simple retrieval-augmented generation. arXiv preprint arXiv:2501.06713 , 2025. [71] Lei Liang, Mengshu Sun, Zhengke Gui, Zhongshu Zhu, Zhouyu Jiang, Ling Zhong, Y... | https://arxiv.org/abs/2505.16237v1 |
where |E|is the number of edges, and Ce is a predefined per-edge cost that serves as a regularization parameter to control the subgraph’s size. In this way, we can obtain a preliminary retrieved subgraph GRetriever . B Experimental Settings Table 4: Statistics of GraphQA Benchmark [18]. Dataset ExplaGraphs SceneGraphs ... | https://arxiv.org/abs/2505.16237v1 |
2] as generators. Both reranker-based and GNN-based methods apply LoRA for fine-tuning, except for GraphToken, whose results are from G-retriever [ 18]. All methods are compared using the same training hyperparameters where applicable, ensuring fair comparability (e.g., GNN layers and LoRA rank). In the Aligner module,... | https://arxiv.org/abs/2505.16237v1 |
prune irrelevant knowledge while preserving the essential information. However, when nseedis too small (e.g., nseed= 4), performance drops significantly, suggesting that some critical information may be pruned during the process. On the other hand, increasing the number of seed nodes to 15 (utilizing 60.04% of the toke... | https://arxiv.org/abs/2505.16237v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.