text string | source string |
|---|---|
Weizhu Chen. 2023. Learn- ing from mistakes makes llm better reasoner. ArXiv preprint , abs/2310.20689. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Sub- biah, Jared D Kaplan, Prafulla Dhariwal, Arvind Nee- lakantan, Pranav Shyam, Girish Sastry, and Amanda Askell. 2020. Language models are few-shot learners. InProceed... | https://arxiv.org/abs/2505.22131v1 |
ArXiv preprint , abs/2308.09583. Yun Luo, Zhen Yang, Fandong Meng, Yafu Li, Jie Zhou, and Yue Zhang. 2023b. An empirical study of catastrophic forgetting in large language mod- els during continual fine-tuning. ArXiv preprint , abs/2308.08747. Lucie Charlotte Magister, Jonathan Mallinson, Jakub Adamek, Eric Malmi, and ... | https://arxiv.org/abs/2505.22131v1 |
Lv, Da Pan, Dian Wang, and Dong Yan. 2023b. Baichuan 2: Open large-scale language models. ArXiv preprint , abs/2309.10305. An Yang, Baosong Yang, Beichen Zhang, and Binyuan Hui. 2025. Qwen2.5 technical report. Arxiv preprint , abs/2412.15115. Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui C... | https://arxiv.org/abs/2505.22131v1 |
(Superior) and LEMA baselines, highlighting its robustness and ef- fectiveness regardless of model size or pretraining specialization. A.4 Prompts Used in EULER In this section, we introduce the prompts used in EULER. As shown in Table 5, 0-shot and3-shot represent the prompts used by the EULER model during inference i... | https://arxiv.org/abs/2505.22131v1 |
mistakenly assumes that both regular and overtime hours total “45 hours”, resulting in a calculated salary that far exceeds the ground truth. The SFT model, while correctly distinguishing between regular and over- time hours, miscalculates the overtime duration as “15 hours”, leading to an erroneous final answer. In co... | https://arxiv.org/abs/2505.22131v1 |
is : boxed{ Eliza’s earnings for this week are $580$ dollars }. Erroneous Solution Generated by EULER: Error-Example1: Eliza works 45 hours in total, and she is paid $10 per hour for all 45 hours, so her total earnings for the week are 45 hours multiplied by $10, which equals $450 . Error-Example2: Eliza works 45 hours... | https://arxiv.org/abs/2505.22131v1 |
Improving Brain-to-Image Reconstruction via Fine-Grained Text Bridging Runze Xia, Shuo Feng, Renzhi Wang, Congchi Yin, Xuyun Wen, Piji Li∗ {xiarunze, fengshuo, rzhwang, congchiyin, wenxuyun, pjli }@nuaa.edu.cn College of Artificial Intelligence, Nanjing University of Aeronautics and Astronautics, Nanjing, China MIIT Ke... | https://arxiv.org/abs/2505.22150v1 |
on the left side and the other on the right side of the room. Additionally , there is a bottle placed on the left side of the sink.Figure 1: The top section is the illustration of cognitive as- sumptions during scene observation. The bottom section demonstrates a comparison of the granularity between the detail-enhance... | https://arxiv.org/abs/2505.22150v1 |
( FgB2I ), which employs detail- enhanced text as supplements for Brain-to-Image recon- struction. We use three rewards to guide the decoding of text descriptions and employ Reinforce Algorithm to ad- dress indifferentiable problems in model co-training stage. • We conduct extensive experiments on multiple existing Bra... | https://arxiv.org/abs/2505.22150v1 |
note that the number of fMRI voxels varies across par- ticipants, so we construct a linear layer for each participant to map the fMRI signals to a unified dimensionality. For the fMRI signals of participant s, they are first passed through a linear layer to obtain the fMRI embedding Z∈Rl×d, where ldenotes the length of... | https://arxiv.org/abs/2505.22150v1 |
could recognize the objects within the image and to eval- uate the semantic similarity between the text and the image. Since the decoded text cannot provide gradient information for model training, we utilize the reinforce algorithm (Sutton et al., 1999) to design a loss function that would achieve these objectives. Th... | https://arxiv.org/abs/2505.22150v1 |
trade-off factors that balance the impor- tance of each reward in the training process, with L1,L2, and L3calculated using formula (1). Text-Bridged Brian-to-Image Reconstruction Fine-grained text decoded from brain signals enhances diffusion-model reconstruction via semantic control but can compromise color or structu... | https://arxiv.org/abs/2505.22150v1 |
the LDM (Takagi & Nishimoto, 2023) and BrainDiffuser (Ozcelik & VanRullen, 2023) methods, using the same evaluation metrics as theirs. The values presented in the table are the mean of the assessment results from four participants. MethodLow-Level High-Level PixCorr ↑ SSIM↑ Alex(2) ↑ Alex(5) ↑ Incep↑ CLIP↑ Eff↓ SwA V↓ ... | https://arxiv.org/abs/2505.22150v1 |
inaccuracies that can sometimes limit the effective- ness of detail enhancement in image captions, though more advanced models can partially mitigate this issue. Main Results We apply fine-grained text descriptions to the reconstruction results of a total of three methods, LDM, BrainDiffuser, and MindEye. We first cond... | https://arxiv.org/abs/2505.22150v1 |
through semantic decoding. We ac- knowledge the limitations of current fMRI data and the chal- lenges in decoding neural signals, which inform our future work to improve the granularity of signal decoding. FgB2I provides new perspectives on understanding and reconstruct- ing the intricate process of human visual cognit... | https://arxiv.org/abs/2505.22150v1 |
Retrieved from https://doi.org/10.1145/3581783.3613832 doi: 10.1145/3581783.3613832 Luck, S. J., & Ford, M. A. (1998). On the role of selective attention in visual perception. Proceedings of the National Academy of Sciences ,95(3), 825–830. Mnih, V ., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, M. ... | https://arxiv.org/abs/2505.22150v1 |
AAAI con- ference on artificial intelligence, AAAI 2022, thirty- fourth conference on innovative applications of artifi- cial intelligence, IAAI 2022, the twelveth symposium on educational advances in artificial intelligence, EAAI 2022 virtual event, february 22 - march 1, 2022 (pp. 5350–5358). AAAI Press. Retrieved fr... | https://arxiv.org/abs/2505.22150v1 |
arXiv:2505.22156v1 [cs.CL] 28 May 2025InComeS: Integrating Compression and Selection Mechanisms into LLMs for Efficient Model Editing Shuaiyi Li1∗Zhisong Zhang2,†Yang Deng3Chenlong Deng2Tianqing Fang2 Hongming Zhang2Haitao Mi2Dong Yu2Wai Lam1,† {sli, wlam}@se.cuhk.edu.hk ,zhisonzhang@tencent.com 1The Chinese University... | https://arxiv.org/abs/2505.22156v1 |
Mech- anisms), a novel framework for efficient and scalable model editing. InComeS adopts context compression techniques to condense the representation of each edit into the KV cache of special gist tokens, which can be cached and reused for computational efficiency. While gisting [29] was originally developed to compr... | https://arxiv.org/abs/2505.22156v1 |
edit batch. Given a batch of editing information 2 Messi wins 2022 World Cup <GIST> …… …… The UK Capital is London <GIST> The UK Capital is London <GIST>Please unveil the UK capital : …… …. .... …. …. …. ….…. …. …. …. …. ….…. …. …. …. …. ….0.1 0.05 0.05 0.1 0.3 0.4…. …. …. …. …. ….…. …. …. …. …. ….𝒈𝟓EditsQuery : …… C... | https://arxiv.org/abs/2505.22156v1 |
edit information, ensuring that the subsequent information selection process is seamless and well-aligned with the model’s internal representations. 3.2 Edit Selection After compressing the edit contexts, we obtain a pool of gist representations for the batch of edits. To integrate this information into the model, we i... | https://arxiv.org/abs/2505.22156v1 |
losses. This scheme increases the weights of edit-sensitive tokens to encourage the model to learn to retrieve information from the compressed edits. The loss differences are calculated with a teacher model, which is the original, unedited version of the target LM. In addition to token reweighting, we also adopt knowle... | https://arxiv.org/abs/2505.22156v1 |
mainly compare with methods designed to support batch or sequential editing. Table 1 presents our main results, which demonstrate the effectiveness of InComeS in both single-editing and batch-editing scenarios. In addition, InComeS surpasses ICL in all metrics except single 2-edits and 3-edits for Qwen2.5-7B, which sho... | https://arxiv.org/abs/2505.22156v1 |
7.88 4.23 SERAC [28] 55.56 59.23 53.67 42.34 40.33 39.39 MEND [27] 34.23 45.34 30.25 39.88 35.45 34.21 ICL 69.76 76.91 74.54 53.53 50.54 44.77 InComeS 66.464.73%↓71.247.37%↓76.54 2.68%↑55.13 3.00%↑53.48 5.82%↑47.91 7.01%↑ Table 1: Results on MQuAKE [47]. The difference between InComeS and ICL is marked. Method ModelSin... | https://arxiv.org/abs/2505.22156v1 |
[28] 89.56 / 78.32 60.56 / 40.45 92.69 / 89.61 66.59 / 51.61 ICL 93.31 / 82.95 65.81 /49.75 68.86 / 60.84 62.19 / 55.58 InComeS 91.16 / 76.81 65.15 /45.66 97.22 / 87.09 70.70 / 52.23 Base Qwen2.5-7B22.35 / 22.35 21.46 / 21.46 36.21 / 36.21 43.86 / 43.86 FT-M 98.93 /90.18 49.39 / 43.13 99.51 /92.60 50.04 / 46.41 LoRA [1... | https://arxiv.org/abs/2505.22156v1 |
2 4 67 9 11 13 15 x - Layer0.20.40.60.8Zero-gist probability (b) - Zero-gist prob on each layer 100 gist bsz 200 gist bsz 300 gist bsz400 gist bsz 500 gist bsz 0 2 4 6 8 10 12 14 16 x - T oken position0.00.20.40.60.81.0Probability (c) - Zero-gist and golden probability Golden prob - success Golden prob - fail Golden pr... | https://arxiv.org/abs/2505.22156v1 |
T=0.45. Information flow on tokens We further investigate the cross-attention patterns to understand how the model performs context selection. We measure the zero-gist and golden-gist probability (Figure 4c), and cross-attention entropy (Figure 4d) of each token from two representative examples containing a correctly (... | https://arxiv.org/abs/2505.22156v1 |
The memory formats applied by different researchers are diverse. Methods like SERAC [28], IKE [46], MeLLo [48] adopt explicit non-parametric memory , which stores specific edit instances, and a retriever that is responsible for recalling relevant edits from the memory. For example, IKE uses KNN, and SERAC applies a tra... | https://arxiv.org/abs/2505.22156v1 |
Token-based Context Compression”. In: CoRR abs/2412.17483 (2024). DOI: 10. 48550/ ARXIV. 2412. 17483 . arXiv: 2412. 17483 .URL:https: // doi. org/ 10. 48550/arXiv.2412.17483 . [7] Qingxiu Dong et al. “Calibrating Factual Knowledge in Pretrained Language Models”. In: Findings of the Association for Computational Linguis... | https://arxiv.org/abs/2505.22156v1 |
21st Conference on Computational Natural Language Learning (CoNLL 2017) . Ed. by Roger Levy and Lucia Specia. Vancouver, Canada: Association for Computational Linguistics, Aug. 2017, pp. 333–342. DOI:10.18653/v1/K17-1034 .URL:https://aclanthology. org/K17-1034 . [20] Shuaiyi Li et al. “Consecutive Batch Model Editing w... | https://arxiv.org/abs/2505.22156v1 |
2022, 7-8 April 2022, Virtual Event . Ed. by Gerardo Flores et al. V ol. 174. Proceedings of Machine Learning Research. PMLR, 2022, pp. 248–260. URL:https://proceedings.mlr.press/v174/pal22a.html . [32] Yuval Pinter and Michael Elhadad. “Emptying the Ocean with a Spoon: Should We Edit Mod- els?” In: Findings of the Ass... | https://arxiv.org/abs/2505.22156v1 |
Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence, IAAI 2024, Fourteenth Symposium on Educational Advances in Artificial Intelligence, EAAI 2014, February 20-27, 2024, Vancouver, Canada . Ed. by Michael J. Wooldridge, Jennifer G. Dy, and Sriraam Natarajan. AAAI Press, 2024, pp. 19449–19457. ... | https://arxiv.org/abs/2505.22156v1 |
work. B Training details InComeS is trained on around 1.5 billion tokens, which mainly come from summarization and QA datasets. Specifically, for summerization datasets, we select 4.5e6instances from S2ORC [22], 1.15e6instances from AG News Corpus9; and for QA datasets, we use squad [35], a modified version 10of the na... | https://arxiv.org/abs/2505.22156v1 |
/ 24.44 12.73 / 10.91 24.12 / 10.12 KN [3] 20.60 / - 17.16 / - 19.46 / - 16.01 / - 06.70 / - 21.23 / - IKE [46] 61.70 / - 45.55 / - 48.80 / - 59.15 / - 57.39 / - 40.21 / - SERAC [28] 89.56 / 78.32 60.56 / 46.45 45.67 / 39.36 92.69 / 89.61 66.59 / 63.60 48.96 / 41.32 ICL 93.31 / 82.95 65.81 / 49.75 62.23 / 41.49 68.86 /... | https://arxiv.org/abs/2505.22156v1 |
receives a lower prediction score than accurate facts. It constructs out-of-scope data by substituting the subject entity with a comparable description that has the same predicate. C.2 Evaluation metrics This section explains the evaluation metrics used in the extended ZsRE[41, 44] and Wiki counterfact [2, 44]. General... | https://arxiv.org/abs/2505.22156v1 |
location of the codebook layer 13 and 18 for Llama-3.2-1B and Qwen2.5-7B, respectively. Surprisingly, the ϵvalue used in the original paper (1-3) seems insufficient for the complex editing experiments in this paper. Therefore, we increase it to 50. The number of optimization steps for the value vector is set to 100. IK... | https://arxiv.org/abs/2505.22156v1 |
Stratified Selective Sampling for Instruction Tuning with Dedicated Scoring Strategy Paramita Mirza1, Lucas Weber1, Fabian Küch1 1Fraunhofer IIS {first.last}@iis.fraunhofer.de Abstract Recent work shows that post-training datasets for LLMs can be substantially downsampled without noticeably deteriorating performance. H... | https://arxiv.org/abs/2505.22157v1 |
& response quality) Generate Brainstorm Fact QA Reason Extract Math Code Quality scoring …Extraction: 0.025 Factual QA: 0.025 Brainstorming: 0.15 Reasoning: 0.1 Coding: 0.2 Math: 0.25 Generation: 0.25 15% 25% 2.5% 10% 2.5% 25% 20% Code Fact QA Brainstorm Generate Extract Reason Math Take inspiration from here Code …Clu... | https://arxiv.org/abs/2505.22157v1 |
require a preprocessing step to bucket samples by task type. While some works mention this as part of data anal- ysis (Ouyang et al., 2022a; Conover et al., 2023) or filtering pipelines (Grattafiori et al., 2024b), con- crete methods and applications remain underex- plored. Furthermore, as we detail in Section 2, only ... | https://arxiv.org/abs/2505.22157v1 |
closely follows CaR (Ge et al., 2024), which clusters data points and samples one representative from each cluster. IT data categorization. Previous work has pro- posed various task types (e.g., open QA ,brain- storming ,creative writing ), to guide IT data collec- tion from human annotators (Ouyang et al., 2022a; Cono... | https://arxiv.org/abs/2505.22157v1 |
fi) and a general-purpose difficulty scorer (yielding qi); the results are combined into an overall preference score pi. 3.Clustering + Ranking . Ultimately, we select the samples with the highest piwhile using a clustering approach to maintain diversity and minimise redundancies in D′. We detail each step in the remai... | https://arxiv.org/abs/2505.22157v1 |
Appendix A.2.1; Liu et al., 2024b). Our goal is to train a general and robust difficulty scorer for our data selection pipeline that predicts how likely it is for an average model to solve a data point incorrectly, independent of its category l. To source the training set Ddifffor such a scorer, we collect 20k instruct... | https://arxiv.org/abs/2505.22157v1 |
(ii)pro- duce a revised version that improves or fixes the original code (see Figure 13, Appendix A.3). The resulting score, qcode, is based on the normalized Levenshtein similarity between lines of the orig- 1https://huggingface.co/hkust-nlp/ deita-quality-scorer 4 inal ( lo0, ..., lo n) and revised ( lr0, ..., lr m) ... | https://arxiv.org/abs/2505.22157v1 |
score pi=fi·qi, where fiandqiare the difficulty and quality scores, respectively. Each of them is normalized using min- max scaling, with the 1st and 99th percentiles as the minimum and maximum values across all samples inD. The quality scores are normalized per scorer (qmath,qcode,qifandqdeita) as they have differ- in... | https://arxiv.org/abs/2505.22157v1 |
each base model using theSFTTrainer from the Transformer Reinforce- 4γis a hyperparameter and set to 75. 5Without augmented problems. 6Randomly sampled. random (baseline) longest (baseline) deita (baseline) difficulty quality combination combination++ all data16.1 12.8 7.1 13.9 51.1 24.0 25.0 16.035.0 43.7 31.7 31.5 37... | https://arxiv.org/abs/2505.22157v1 |
consistent gains of all tested con- ditions for all five bases and is outperformed by full-data training in only two cases, confirming that the proposed sampling approach transfers well across models. A benchmark -wise analysis (Fig- ure 3c) shows similarly stable improvements, with combination ++delivering consistent ... | https://arxiv.org/abs/2505.22157v1 |
k(see Figure 6a bottom, for the resulting category distribution). FromDskewed , we sample m= 25 kitems with the same strategies as previously and fine -tune Mistral -7B-Base on each D′ skewed . We expected diversity-aware sampling strategies like Deita and combination ++to mitigate bias by enforcing ei- ther semantic s... | https://arxiv.org/abs/2505.22157v1 |
used evaluation frameworks. While we try to mitigate this issue as much as possible, we can- not guarantee that difficulty scores exactly reflect a model’s capacity to solve a given data point. Quality scorer. We did not develop nor have dedicated quality scorers for samples belonging to Reasoning ,Factual QA andExtrac... | https://arxiv.org/abs/2505.22157v1 |
Srini- vasan, Tianyi Zhou, Heng Huang, and Hongxia Jin. 2024. Alpagasus: Training a better alpaca with fewer data. In The Twelfth International Conference on Learning Representations . Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. 2018. Think you have sol... | https://arxiv.org/abs/2505.22157v1 |
Models. arXiv preprint . ArXiv:2407.21783 [cs]. Aaron Grattafiori and 1 others. 2024b. The Llama 3 Herd of Models. arXiv preprint . ArXiv:2407.21783 [cs]. Yexiao He, Ziyao Wang, Zheyu Shen, Guoheng Sun, Yucong Dai, Yongkai Wu, Hongyi Wang, and Ang Li. 2024. SHED: Shapley-based automated dataset refinement for instructi... | https://arxiv.org/abs/2505.22157v1 |
Schwenk. 2019. Mlqa: Evaluating cross-lingual extractive question answer- ing. arXiv preprint arXiv:1910.07475 , arXiv: 1910.07475. Ming Li, Yong Zhang, Zhitao Li, Jiuhai Chen, Lichang Chen, Ning Cheng, Jianzong Wang, Tianyi Zhou, and Jing Xiao. 2024a. From quantity to quality: Boosting LLM performance with self-guided... | https://arxiv.org/abs/2505.22157v1 |
AI. 2025. Mistral small 3. https://mistral. ai/news/mistral-small-3 . Apache 2.0 License. Terufumi Morishita, Gaku Morio, Atsuki Yamaguchi, and Yasuhiro Sogawa. 2023. Learning deductive rea- soning from synthetic corpus based on formal logic. 11 InProceedings of the 40th International Conference on Machine Learning , v... | https://arxiv.org/abs/2505.22157v1 |
R Bowman. 2024. Gpqa: A graduate-level google-proof q&a benchmark. In First Conference on Language Modeling . Melanie Sclar, Yejin Choi, Yulia Tsvetkov, and Alane Suhr. 2024. Quantifying language models’ sensitiv- ity to spurious features in prompt design or: How i learned to start worrying about prompt formatting. InT... | https://arxiv.org/abs/2505.22157v1 |
Hupkes. 2023a. The icl consistency test. arXiv preprint arXiv:2312.04945 . Lucas Weber, Elia Bruni, and Dieuwke Hupkes. 2023b. Mind the instructions: a holistic evaluation of con- sistency and interactions in prompt-based learning. InProceedings of the 27th Conference on Computa- tional Natural Language Learning (CoNLL... | https://arxiv.org/abs/2505.22157v1 |
Learning , ICML’24. JMLR.org. Yingxiu Zhao, Bowen Yu, Binyuan Hui, Haiyang Yu, Minghao Li, Fei Huang, Nevin L Zhang, and Yongbin Li. 2024b. Tree-instruct: A preliminary study of the intrinsic relationship between complexity and align- ment. In Proceedings of the 2024 Joint International Conference on Computational Ling... | https://arxiv.org/abs/2505.22157v1 |
subsequent Ta- ble 4 shows the respective data source from which we obtained the data, as well as the type of evalu- ation metric that we use to evaluate the 18 LLMs from Table 5. During the data collection for the difficulty scorer, we collect training sets from different bench- marks as well as data points from OpenA... | https://arxiv.org/abs/2505.22157v1 |
interview train 303 pass@1 competition train 289 pass@1 CONALA (Yin et al., 2018) Default train 97 Bleu Table 4: Datasets used in difficulty scorer training. A.2.3 Training the difficulty scorer – details We equip regular CausalLLMs with a regression head by pooling the final dense layers and adding a linear projection... | https://arxiv.org/abs/2505.22157v1 |
instruction- following scorer. dataset collection of evaluation details.8Table 9 presents the evaluation results of various LLMs as the annotator/judge. ifeval_like_5k alpaca_gpt4 open_math_instruct_2 flan_v2_90k ultrainteract_coding wizardlm_evol_instruct sharegpt_en 200k_agentinst_random5.2 99.7 6.4 23.24.0 2.26.6 9.... | https://arxiv.org/abs/2505.22157v1 |
25k 50k 100k 700k0.02 0.000.020.04Score gainGPQA all data random (baseline) longest (baseline) deita (baseline) combination++ (e) GPQA (Rein et al., 2024) 1k 5k 10k 25k 50k 100k 700k0.00.10.20.30.4Score gainIFeval all data random (baseline) longest (baseline) deita (baseline) combination++ (f) IFeval (Zhou et al., 2023... | https://arxiv.org/abs/2505.22157v1 |
- keyword_frequency, e.g., five hashtags, 'but'two times, letter 'r'at least 3 times - language, e.g., english, two languages - length, e.g., number of words, number of sentences, number of paragraphs - punctuation, e.g., no commas, quotation - start_and_ending, e.g., start with 'Hello ', end with 'Thank you! ' - writi... | https://arxiv.org/abs/2505.22157v1 |
arXiv:2505.22169v1 [cs.CL] 28 May 2025RELIABLE EVAL: A Recipe for Stochastic LLM Evaluation via Method of Moments Gili Lior1Eliya Habba1Shahar Levy1Avi Caciularu2Gabriel Stanovsky1 1The Hebrew University of Jerusalem2Google Research gili.lior@mail.huji.ac.il Abstract LLMs are highly sensitive to prompt phrasing, yet st... | https://arxiv.org/abs/2505.22169v1 |
benchmarks. Our findings, shown in Figure 1, reveal the statis- tical differences between models, highlighting the need for stochastic evaluation. Moreover, we show that the number of resamplings required to reliably estimate model performance varies depending on both the model and the dataset being evaluated. We hope ... | https://arxiv.org/abs/2505.22169v1 |
ϵand confidence level δ, letSDbe the space of all meaning-preserving prompt perturbations of dataset D, and let S′⊂SDbe a random subset of sizen. Then, we say that nsamples yield a reliable evaluation if for every moment µi(expected value and variance), it holds that: P S′⊂SD |S′|=n µi(M, S′)−µi(M, S D) > ϵ < δ (4) 2... | https://arxiv.org/abs/2505.22169v1 |
ˆMand compute its empirical mo- ments over large Nas proxy for true moments. In the following section, we will show that choosing a relatively cheap model gives empirically good estimates, which hold across models. For each can- didate sample size n= 1,2, . . . , N , compute the set of deviations between the empirical ... | https://arxiv.org/abs/2505.22169v1 |
Figure 2b shows that the smaller model provides a valid up- per bound on convergence behavior. This suggests that smaller models can serve as effective proxies for estimating the number of prompt resamplings required for reliable stochastic evaluation of larger models. This is shown also for the GPQA-Diamond and Simple... | https://arxiv.org/abs/2505.22169v1 |
herd of models. arXiv preprint arXiv:2407.21783 . Alex Gu, Wen-Ding Li, Naman Jain, Theo Olausson, Ce- line Lee, Koushik Sen, and Armando Solar-Lezama. 2024a. The counterfeit conundrum: Can code lan- guage models grasp the nuances of their incorrectgenerations? In Findings of the Association for Com- putational Linguis... | https://arxiv.org/abs/2505.22169v1 |
features in prompt design or: How i learned to start worrying about prompt formatting. InThe Twelfth International Conference on Learning Representations . Charlotte Siska, Katerina Marazopoulou, Melissa Ailem, and James Bono. 2024. Examining the ro- bustness of LLM evaluation to the distributional as- sumptions of ben... | https://arxiv.org/abs/2505.22169v1 |
to determine alignment between predictions and gold answers. We adopt the judging prompt from the official SimpleQA repository.2Our judge model is Atla Selene Mini (Alexandru et al., 2025), which currently ranks highest among open-source models on the Judge Arena Leaderboard.3 2https://github.com/openai/simple-evals 3h... | https://arxiv.org/abs/2505.22169v1 |
Reverse Preference Optimization for Complex Instruction Following Xiang Huang1,2, Ting-En Lin1, Feiteng Fang1, Yuchuan Wu1, Hangyu Li1, Yuzhong Qu2*, Fei Huang1, Yongbin Li1* 1Tongyi Lab 2State Key Laboratory for Novel Software Technology, Nanjing University, China {chengjun.hx, ting-en.lte, shengxiu.wyc, shuide.lyb}@a... | https://arxiv.org/abs/2505.22172v1 |
for alignment and distinguish chosen and rejected responses based on their total scores (the number of constraints that were successfully fol- lowed). In a multi-preference alignment scenario, this method presents two main drawbacks: Firstly, using the total score difference as the gap between two responses may not acc... | https://arxiv.org/abs/2505.22172v1 |
low instructions with certain constraints, such as “write a story about Ne Zha with less than 200 words andend it with an emoji ”. Zhou et al. (2023) first propose this challenge and construct IFEV AL dataset with verifiable constraints that can be judged with simple rules. Subsequent bench- marks such as FollowBench (... | https://arxiv.org/abs/2505.22172v1 |
reward model into the Bradley-Terry (BT) rank- ing objective (Bradley and Terry, 1952) p(yw≻ yl|x) = σ(r(x, yw)−r(x, yl)), DPO aims to maximize the probability of the preferred output ywand minimize that of the undesirable output yl. The optimization objective is formulated as: LDPO=−E(x,yw,yl)∼D" logσ βlogπθ(yw|x) πre... | https://arxiv.org/abs/2505.22172v1 |
AABCDEF①Same Score Different Score C# : Mention the word 'AI' at least once … Valid preference pairABCDEF ABCDEF ABCDEF ABCDEFABCDEF ABCDEF ABCDEF ABCDEFABCDEF ABCDEF ABCDEF ABCDEF ABCDEF ABCDEF DEF ABC[[ {{"constraint""constraint" : : "The response must be fewer than 200 words""The response must be fewer than 200 word... | https://arxiv.org/abs/2505.22172v1 |
more diverse data in a multi- turn setting. We performed role modeling for users and the system, assigning them with unique profiles. By leveraging these profiles as a context-specific conversation background, we introduce variations during generation, thereby preventing the repeated generation of similar instructions.... | https://arxiv.org/abs/2505.22172v1 |
used to train the SFT model and construct a unified dialogue history for a fair comparison of different alignment methods. 5.4 Quality Evaluation Method To evaluate the quality of the sampled responses, we use LLMs to determine whether the response adheres to each constraint one by one. Consider- ing the autoregressive... | https://arxiv.org/abs/2505.22172v1 |
70.08 78.35 Table 1: Main result of two multi-turn complex instructions following dataset. For SysBench, we report CSR, ISR, and SSR. For Multi-IF, we report the average accuracy in three steps. The accuracy are the average of four metrics: Prompt-level strict-accuracy, Inst-level strict-accuracy, Prompt-level loose-ac... | https://arxiv.org/abs/2505.22172v1 |
works efficiently on the 70B model and surpasses GPT-4o on most metrics. Besides, experiments on Qwen-2.5 model exhibit similar superiority to other baselines. RPO can distinguish chosen and rejected exam- ples more efficiently. As the pair-wise alignment method aims to maximize the probability between chosen and rejec... | https://arxiv.org/abs/2505.22172v1 |
(chosen not worse than rejected on any aspect), and Perfect preference pair (chosen is perfect). We further analyze the Perfect rate when there are fewer than five constraints to follow and no less than five constraints. Direct setting con- structs preference pairs on these five responses. Refine setting additionally p... | https://arxiv.org/abs/2505.22172v1 |
to incur a smaller alignment tax. For the 70B models, both DPO and RPO in- troduce bonuses over the Instruct model on some tasks, such as RPO on AlignBench and GSM8K. 7 Conclusion In this paper, we explore the noise issue of aligning with multiple preferences. We introduce RPO, a simple and efficient method that dynami... | https://arxiv.org/abs/2505.22172v1 |
Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian, Clemens Winter, Philippe Tillet, Felipe Petroski Such, Dave Cum- mings, Matthias Plappert, Fotios Chantzis, Eliza- beth Barnes, Ariel Herbert-V oss, William Hebgen Guss, Alex Nichol, Alex Paino, Nikolas Tezak, Jie Tang, Igor ... | https://arxiv.org/abs/2505.22172v1 |
Lucile Saulnier, Marie- Anne Lachaux, Pierre Stock, Sandeep Subramanian, Sophia Yang, Szymon Antoniak, Teven Le Scao, Théophile Gervet, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. 2024a. Mixtral of experts. Preprint , arXiv:2401.04088. Yuxin Jiang, Yufei Wang, Xingshan Zeng, Wanjun Zhong, Liang... | https://arxiv.org/abs/2505.22172v1 |
Ke, Xiaotao Gu, Lindong Wu, Hao Huang, Jinfeng Zhou, Wenchuang Li, Binxin Hu, Wendy Gao, Jiaxin Xu, et al. 2024. Bench- marking complex instruction-following with mul- tiple constraints composition. arXiv preprint arXiv:2407.03978 . Can Xu, Qingfeng Sun, Kai Zheng, Xiubo Geng, Pu Zhao, Jiazhan Feng, Chongyang Tao, and ... | https://arxiv.org/abs/2505.22172v1 |
While CRAB’s method can potentially produce constraints of varying quality, the simplicity of RPO’s reversal mechanism minimizes the likelihood of errors. Furthermore, CRAB is focused on natural language generation (NLG) tasks, requiring large language models (LLMs) to fully understand the constraint system to generate... | https://arxiv.org/abs/2505.22172v1 |
It is worth noting that, after introducing characters, we can still generate the atomic constraints mentioned above. Therefore, in our observation, the introduction of character-driven data collection primarily brings benefits. C Discussion about Constraints that Can not be Reversed We observed that at the current stag... | https://arxiv.org/abs/2505.22172v1 |
of the SFT model was inferior to that of the Instruct model in some cases. We believe this might be due to the fact that the Instruct model has already undergone instruction tuning, especially for latest models (such as LLaMA-3.1 andQwen-2.5 ) which have been optimized for instruction-following tasks. Consequently, the... | https://arxiv.org/abs/2505.22172v1 |
with an exclamation mark. When the user mentions ’chal- lenge’, they must be encouraged to keep going.When the user mentions ’challenge’, they must be discouraged from continuing.When the user mentions ’challenge’, there is no need to encourage them to keep going. Regardless of the topic, each re- ply must end with a q... | https://arxiv.org/abs/2505.22172v1 |
words. Reversed constraint(Predicted): In any case, the response length must not be less than 200 words. Original constraint: In any feedback request, provide at least three reasons for the star rating. Reversed constraint(Predicted): In any feedback request, provide at most three reasons for the star rating. Original ... | https://arxiv.org/abs/2505.22172v1 |
must not be mentioned in responses‘: For example, prohibiting certain keywords, prohibiting the use of emojis, commas, Arabic numerals, prohibiting quoting poems, idioms, prohibiting discussion on certain topics. - ‘Starting content‘ (used in a minority of cases, less than 15%): For example, starting with "Hello," a ca... | https://arxiv.org/abs/2505.22172v1 |
Constrain ontology, used in system prompt generation. G Detail of Profile Generation You are a master of character modeling and constraint generation. Here is the basic information about the character (a chatbot) you need to model: <Start of Character Reference Information> {system_inspired_corpus} <End of Character Re... | https://arxiv.org/abs/2505.22172v1 |
used only for immediate, single-round, pure text dialogue scenarios. Absolutely avoid generating triggers like "if the user is silent for more than 30 seconds," which are unrelated to pure text dialogue; absolutely avoid generating constraints like "when the user expresses interest in xx for three consecutive rounds," ... | https://arxiv.org/abs/2505.22172v1 |
childhood. Every weekend, he would visit her, listening to stories about Italy and learning to make traditional Italian dishes. However, as time went by, Marco became engulfed in his busy work schedule and modern life, and his connection with his grandmother dwindled. After her passing, he felt a deep sense of loss, re... | https://arxiv.org/abs/2505.22172v1 |
in the response for added persuasive effect. 6. When the user expresses obvious frustration or stress, first offer emotional support, such as ’I understand how you’re feeling right now, it’s really tough,’ and then provide specific advice. 7. If the user asks about handling relationships with colleagues, the response m... | https://arxiv.org/abs/2505.22172v1 |
provided information. The user queries must meet the following requirements (these requirements are very important and must be followed): ### In terms of expression: 1. Natural and concise expression: The expression must not be stiff and rigid like robot speech. Use more colloquial expressions, fewer overly formal expr... | https://arxiv.org/abs/2505.22172v1 |
query: Table 20: Prompt for query generation. I Prompt for Fine-grained Evaluation You are now an expert in evaluating the results of large models. Below, you will face an evaluation task regarding a large model’s ability to adhere to a system prompt. I will provide you with the current round’s query, the current round... | https://arxiv.org/abs/2505.22172v1 |
arXiv:2505.22176v1 [cs.CL] 28 May 2025TabXEval : Why this is a Bad Table? An eXhaustive Rubric for Table Evaluation ∗Vihang Pancholi1 ∗Jainit Bafna2 ∗Tejas Anvekar1 Manish Shrivastava2 †Vivek Gupta1 1Arizona State University2IIIT Hyderabad {vpancho1,tanvekar,vgupta140}@asu.edu {jainit.bafna,m.shrivastava}@research.iiit... | https://arxiv.org/abs/2505.22176v1 |
LLM- based table evaluation method that aligns ref- erence tables structurally and compares them semantically and syntactically via our rubric. •We construct TabXBench , a diverse bench- markderivedfrommulti-domaindatasets,val- idating evaluation metricsthrough structured perturbations and human assessments. •We analyz... | https://arxiv.org/abs/2505.22176v1 |
exact string matching to establish aprecisebaselinealignment. Next,werefinethis alignment using an LLM to account for abbrevi- ations, synonyms, and structural transformations (e.g., merged columns, row/column transpositions). Purely exact matching can be overly strict, missing semanticallyequivalentbutsyntacticallydif... | https://arxiv.org/abs/2505.22176v1 |
range of potentialtableperturbations. Thisdiversityensures thatTabXBench captures common pitfalls such asmissingrows/columns,reorderedheaders,unit mismatches,numericdiscrepancies,andcomplex structural variations (e.g., row/column transposi- tion). TabXBench consists of 50 handpicked “clean” (reference) tables from six ... | https://arxiv.org/abs/2505.22176v1 |
LLM based baseline correlates, only 30.6%and 40.6% Pearson’s ρcorrelation for the Rubric Structure DescriptorandCellLevelDescriptorrespectively. 5 Revealing that it fails to understand the rubric, and quantify the structural and contextual challenges in table evaluation, and hence is unable to correctly align with huma... | https://arxiv.org/abs/2505.22176v1 |
ble outputs, capturing both coarse and fine-grained discrepancies thatare critical in real-worldscenar- ios. 4.3 What Sets TabXEval Apart? Akeycriteriaofanyevaluationmetricistoachieve a balance between specificity (i.e., avoiding false positives) and sensitivity (i.e., avoiding false neg- atives). InFigure3,wevisualize... | https://arxiv.org/abs/2505.22176v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.