pith. sign in

arxiv: 2605.21883 · v1 · pith:UWFYZCPYnew · submitted 2026-05-21 · 💻 cs.CL

Token-weighted Direct Preference Optimization with Attention

Chengyu Huang , Zhuohang Li , Sheng-Yen Chou , Claire Cardie This is my paper

Pith reviewed 2026-05-22 07:00 UTC · model grok-4.3

classification 💻 cs.CL
keywords Direct Preference Optimizationtoken weightingattention mechanismsLLM alignmentpreference optimizationpairwise judgment
0
0 comments X

The pith

AttentionPO derives token weights from the LLM's own attention when it judges response pairs, improving Direct Preference Optimization.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Standard Direct Preference Optimization treats every token in a response as equally important during alignment. The paper introduces a general token-weighted DPO objective and then AttentionPO as a concrete method that pulls token weights directly from the model's attention. It obtains those weights by prompting the LLM to act as a pairwise judge between chosen and rejected responses and reading the attention scores from the two resulting forward passes. The resulting weights are content-dependent and require no separate model or training step. Experiments on AlpacaEval, MT-Bench, and ArenaHard show that this yields stronger performance than prior preference optimization techniques.

Core claim

AttentionPO instantiates token-weighted DPO by prompting the language model to serve as a pairwise judge, extracting attention scores from the two forward passes required for that judgment, and using those scores to scale each token's contribution inside the DPO loss function.

What carries the argument

Attention scores extracted when the LLM is prompted to compare chosen and rejected responses, used to reweight tokens inside the token-weighted DPO objective.

If this is right

  • AttentionPO surpasses standard DPO and existing token-level preference optimization methods on AlpacaEval, MT-Bench, and ArenaHard.
  • The method requires only two additional forward passes per training example.
  • Token weights adjust automatically according to the content of the specific responses being compared.
  • The weighting scheme follows directly from a token-weighted reinforcement learning formulation of the DPO objective.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same attention-based weighting could be inserted into other preference optimization losses that operate on log-probability differences.
  • Internal model states such as attention may serve as lightweight proxies for token importance across a wider set of alignment algorithms.
  • Varying the exact prompt used to elicit the pairwise judgment might produce still sharper token weights for particular domains.

Load-bearing premise

The attention scores produced when the model judges response pairs reliably indicate which tokens matter most for the preference signal.

What would settle it

A controlled run in which uniform or random token weights replace the attention-derived weights and performance on AlpacaEval, MT-Bench, and ArenaHard remains equal or higher would falsify the contribution of the attention mechanism.

Figures

Figures reproduced from arXiv: 2605.21883 by Chengyu Huang, Claire Cardie, Sheng-Yen Chou, Zhuohang Li.

Figure 1
Figure 1. Figure 1: AttentionPO weighs each token by attention, [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Workflow of AttentionPO. First, we prompt [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Results for different layers. x-axis: index of [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: An example for LLaMA-3-8B-Base-SFT. Prompt: Given the task definition and input, reply with output. Given a text, write a compressed version of it in a single sentence. This little museum celebrates the ingenuity and courage of those who sought to escape to the West, and commemorates those who died trying to do so. Exhibits include the shopping cart in which a mother smuggled her infant son across the bord… view at source ↗
Figure 5
Figure 5. Figure 5: An example for LLaMA-3-8B-Instruct. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Pairwise Judge Prompt Verbalized Self Judge System Prompt You are an expert in linguistics. Given a question and a response, split the response into parts and rate the importance of each part relative to the user's question on a scale of 1 (filler) to 5 (critical content). ## Guidelines * Split the response into parts that are as fine-grained as possible. The split needs to be on phrase-level or word-level… view at source ↗
Figure 7
Figure 7. Figure 7: Verbalized Self-judge System Prompt Verbalized Self Judge User Prompt Question: {prompt} Response: {tokens} [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Verbalized Self-judge User Prompt 19 [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗
read the original abstract

Direct Preference Optimization (DPO) aligns Large Language Models with human preferences without the need for a separate reward model. However, DPO treats all tokens in responses equally, neglecting the differing importance of individual tokens. Existing token-level PO methods compute the token weights using either token-position-based heuristic functions or probability estimates given by a separately trained model, which lacks robustness and incurs extra training cost. In contrast, we propose Token-weighted DPO (TwDPO) -- a novel training objective grounded on token-weighted RL -- and AttentionPO -- an instantiation of TwDPO that uses attention from the LLM itself to estimate token weights. AttentionPO prompts the LLM to serve as a pairwise judge and check where the model attends when comparing the responses. This design makes AttentionPO content-aware, adjusting weights based on response content, and efficient, incurring only two extra forward passes per example. Experiment results show that AttentionPO significantly improves performance on AlpacaEval, MT-Bench, and ArenaHard, surpassing existing Preference Optimization methods.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes Token-weighted Direct Preference Optimization (TwDPO), a generalization of DPO that incorporates token-level weights into the loss. AttentionPO is presented as an efficient instantiation where token weights are estimated from the attention patterns of the base LLM when it is prompted to serve as a pairwise judge between chosen and rejected responses. The authors assert that this yields content-aware weighting with minimal overhead (two extra forward passes) and leads to improved performance over standard DPO and prior token-weighted variants on AlpacaEval, MT-Bench, and ArenaHard.

Significance. If the performance gains are robust and attributable to the attention-based weighting, this work would represent a meaningful advance in preference optimization by providing a lightweight, model-internal mechanism for token importance without requiring additional trained components or hand-crafted heuristics. It could influence how future alignment methods handle variable token contributions in responses.

major comments (2)
  1. [Method] The core of AttentionPO relies on extracting token weights from attention maps during the pairwise judgment prompt. The manuscript does not report any validation of these weights against external preference signals or human annotations, leaving open the possibility that they capture model artifacts rather than preference-relevant importance. This assumption is load-bearing for the claim that the method is 'content-aware' and superior.
  2. [Experiments] While improvements on AlpacaEval, MT-Bench, and ArenaHard are claimed, the results section lacks ablations that isolate the effect of the attention weights (e.g., uniform weights, position-based, or independent judge model). Without these, it is unclear whether the reported gains over DPO and other methods arise from the proposed weighting or from the two extra forward passes. This directly affects the strength of the central empirical claim.
minor comments (2)
  1. [Abstract] The abstract states 'significantly improves performance' but does not include any specific metrics or effect sizes; including at least one key result would strengthen the summary.
  2. [Notation] Ensure consistent use of symbols for the token weight function across equations and text.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments. We address each major comment below, indicating planned revisions to the manuscript where appropriate.

read point-by-point responses

  1. Referee: [Method] The core of AttentionPO relies on extracting token weights from attention maps during the pairwise judgment prompt. The manuscript does not report any validation of these weights against external preference signals or human annotations, leaving open the possibility that they capture model artifacts rather than preference-relevant importance. This assumption is load-bearing for the claim that the method is 'content-aware' and superior.

    Authors: We acknowledge that explicit validation against human annotations would provide additional support. However, the weights in AttentionPO are obtained directly from the base LLM's self-attention when it is prompted to act as a pairwise judge between the chosen and rejected responses. This makes the weighting inherently content-aware, as the attention patterns depend on the specific semantic content of the pair being compared rather than fixed heuristics. We will add qualitative examples in the revised manuscript to illustrate how the attention focuses on tokens that differentiate the responses. revision: partial

  2. Referee: [Experiments] While improvements on AlpacaEval, MT-Bench, and ArenaHard are claimed, the results section lacks ablations that isolate the effect of the attention weights (e.g., uniform weights, position-based, or independent judge model). Without these, it is unclear whether the reported gains over DPO and other methods arise from the proposed weighting or from the two extra forward passes. This directly affects the strength of the central empirical claim.

    Authors: We agree that further ablations would strengthen the empirical claims. Our current comparisons are against DPO and existing token-weighted PO methods, but we will add controls using uniform token weights and position-based weights in the revised results section. These ablations will help isolate the contribution of the attention-derived weights. We will also clarify that the two additional forward passes are required to obtain the content-dependent weights and are not an incidental source of improvement. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical method with independent benchmark evaluation

full rationale

The paper proposes TwDPO as a token-weighted extension of DPO and AttentionPO as an instantiation that extracts token weights via attention maps obtained from two forward passes of the same LLM acting as a pairwise judge. No equations, derivations, or first-principles results are presented that reduce the claimed performance gains to a fitted parameter, self-referential quantity, or self-citation chain. The central claims rest on empirical results across AlpacaEval, MT-Bench, and ArenaHard rather than any closed-form prediction that is definitionally equivalent to its inputs. The attention-weighting step is a methodological choice whose validity is tested externally via benchmark improvements, not by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that LLM attention can serve as a reliable proxy for token importance in preference judgments. No free parameters or invented entities are described in the abstract.

axioms (1)
  • domain assumption Attention weights from the LLM itself, obtained via pairwise judgment prompting, provide content-aware and robust estimates of token importance for preference optimization.
    This premise is invoked to justify why AttentionPO is both efficient and superior to heuristic or separately trained weighting methods.

pith-pipeline@v0.9.0 · 5704 in / 1223 out tokens · 47327 ms · 2026-05-22T07:00:37.027062+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

51 extracted references · 51 canonical work pages · 7 internal anchors

  1. [1]

    Aho and Jeffrey D

    Alfred V. Aho and Jeffrey D. Ullman , title =. 1972

    work page 1972

  2. [2]

    Publications Manual , year = "1983", publisher =

    work page 1983

  3. [3]

    Chandra and Dexter C

    Ashok K. Chandra and Dexter C. Kozen and Larry J. Stockmeyer , year = "1981", title =. doi:10.1145/322234.322243

    work page doi:10.1145/322234.322243 1981

  4. [4]

    Scalable training of

    Andrew, Galen and Gao, Jianfeng , booktitle=. Scalable training of

    work page

  5. [5]

    Dan Gusfield , title =. 1997

    work page 1997

  6. [6]

    Tetreault , title =

    Mohammad Sadegh Rasooli and Joel R. Tetreault , title =. Computing Research Repository , volume =. 2015 , url =

    work page 2015

  7. [7]

    A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =

    Ando, Rie Kubota and Zhang, Tong , Issn =. A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =. Journal of Machine Learning Research , Month = dec, Numpages =

    work page

  8. [8]

    Proceedings of International Conference on Learning Representations , year =

    Ning Yang and Hai Lin and Yibo Liu and Baoliang Tian and Guoqing Liu and Haijun Zhang , title =. Proceedings of International Conference on Learning Representations , year =

    work page

  9. [9]

    Manning and Chelsea Finn , title =

    Rafael Rafailov and Archit Sharma and Eric Mitchell and Stefano Ermon and Christopher D. Manning and Chelsea Finn , title =. Advances in Neural Information Processing Systems , year =

    work page

  10. [10]

    Yu and Meng Cao , title =

    Aiwei Liu and Haoping Bai and Zhiyun Lu and Yanchao Sun and Xiang Kong and Simon Wang and Jiulong Shan and Albin Madappally Jose and Xiaojiang Liu and Lijie Wen and Philip S. Yu and Meng Cao , title =. Proceedings of International Conference on Learning Representations , year =

    work page

  11. [11]

    Proceedings of Empirical Methods in Natural Language Processing , year =

    Kailai Yang and Zhiwei Liu and Qianqian Xie and Jimin Huang and Erxue Min and Sophia Ananiadou , title =. Proceedings of Empirical Methods in Natural Language Processing , year =

    work page

  12. [12]

    Proceedings of International Conference on Machine Learning , year =

    Yongcheng Zeng and Guoqing Liu and Weiyu Ma and Ning Yang and Haifeng Zhang and Jun Wang , title =. Proceedings of International Conference on Machine Learning , year =

    work page

  13. [13]

    Proceedings of International Conference on Learning Representations , year =

    Ruichen Shao and Bei Li and Gangao Liu and Yang Chen and Xiang Zhou and Jingang Wang and Xunliang Cai and Peng Li , title =. Proceedings of International Conference on Learning Representations , year =

    work page

  14. [14]

    Proceedings of Empirical Methods in Natural Language Processing , year =

    Ruichen Shao and Bei Li and Gangao Liu and Yang Chen and Xiang Zhou and Jingang Wang and Xunliang Cai and Peng Li , title =. Proceedings of Empirical Methods in Natural Language Processing , year =

    work page

  15. [15]

    Proceedings of International Conference on Machine Learning , year =

    Kawin Ethayarajh and Winnie Xu and Niklas Muennighoff and Dan Jurafsky and Douwe Kiela , title =. Proceedings of International Conference on Machine Learning , year =

    work page

  16. [16]

    Proceedings of the 27th International Conference on Artificial Intelligence and Statistics , year =

    Mohammad Gheshlaghi Azar and Mark Rowland and Bilal Piot and Daniel Guo and Daniele Calandriello and Michal Valko and Rémi Munos , title =. Proceedings of the 27th International Conference on Artificial Intelligence and Statistics , year =

    work page

  17. [17]

    Proceedings of Findings of the Association for Computational Linguistics , year =

    Ryan Park and Rafael Rafailov and Stefano Ermon and Chelsea Finn , title =. Proceedings of Findings of the Association for Computational Linguistics , year =

    work page

  18. [18]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , year =

    Jiwoo Hong and Noah Lee and James Thorne , title =. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , year =

    work page 2024

  19. [19]

    Proceedings of International Conference on Machine Learning , year =

    Haoran Xu and Amr Sharaf and Yunmo Chen and Weiting Tan and Lingfeng Shen and Benjamin Van Durme and Kenton Murray and Young Jin Kim , title =. Proceedings of International Conference on Machine Learning , year =

    work page

  20. [20]

    Advances in Neural Information Processing Systems , year =

    Zheng Yuan and Hongyi Yuan and Chuanqi Tan and Wei Wang and Songfang Huang and Fei Huang , title =. Advances in Neural Information Processing Systems , year =

    work page

  21. [21]

    Advances in Neural Information Processing Systems , year =

    Yu Meng and Mengzhou Xia and Danqi Chen , title =. Advances in Neural Information Processing Systems , year =

    work page

  22. [22]

    Slic-hf: Sequence likelihood calibration with human feedback

    Yao Zhao and Rishabh Joshi and Tianqi Liu and Misha Khalman and Mohammad Saleh and Peter J. Liu , title =. arXiv preprint arXiv:2305.10425 , year =

    work page arXiv

  23. [23]

    Proximal Policy Optimization Algorithms

    John Schulman and Filip Wolski and Prafulla Dhariwal and Alec Radford and Oleg Klimov , title =. arXiv preprint arXiv:1707.06347 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  24. [24]

    Long Ouyang and Jeff Wu and Xu Jiang and Diogo Almeida and Carroll L. Wainwright and Pamela Mishkin and Chong Zhang and Sandhini Agarwal and Katarina Slama and Alex Ray and John Schulman and Jacob Hilton and Fraser Kelton and Luke Miller and Maddie Simens and Amanda Askell and Peter Welinder and Paul Christiano and Jan Leike and Ryan Lowe , title =. arXiv...

    work page internal anchor Pith review Pith/arXiv arXiv

  25. [25]

    Proceedings of International Conference on Machine Learning , year=

    From Crowdsourced Data to High-Quality Benchmarks: Arena-Hard and BenchBuilder Pipeline , author=. Proceedings of International Conference on Machine Learning , year=

    work page

  26. [26]

    Gonzalez and Ion Stoica , month =

    Tianle Li and Wei-Lin Chiang and Evan Frick and Lisa Dunlap and Banghua Zhu and Joseph E. Gonzalez and Ion Stoica , month =. From Live Data to High-Quality Benchmarks: The Arena-Hard Pipeline , url =

    work page

  27. [27]

    Hashimoto , title =

    Xuechen Li and Tianyi Zhang and Yann Dubois and Rohan Taori and Ishaan Gulrajani and Carlos Guestrin and Percy Liang and Tatsunori B. Hashimoto , title =. GitHub repository , howpublished =. 2023 , month =

    work page 2023

  28. [28]

    Proceedings of Conference on Language Modeling , year=

    Length-Controlled AlpacaEval: A Simple Way to Debias Automatic Evaluators , author=. Proceedings of Conference on Language Modeling , year=

    work page

  29. [29]

    Xing and Hao Zhang and Joseph E

    Lianmin Zheng and Wei-Lin Chiang and Ying Sheng and Siyuan Zhuang and Zhanghao Wu and Yonghao Zhuang and Zi Lin and Zhuohan Li and Dacheng Li and Eric P. Xing and Hao Zhang and Joseph E. Gonzalez and Ion Stoica , title =. Advances in Neural Information Processing Systems, Datasets and Benchmarks Track , year =

    work page

  30. [30]

    The Llama 3 Herd of Models

    Aaron Grattafiori and Abhimanyu Dubey and Abhinav Jauhri and Abhinav Pandey and Abhishek Kadian and Ahmad Al-Dahle and Aiesha Letman and Akhil Mathur and Alan Schelten and Alex Vaughan et al. , title =. arXiv preprint arXiv:2407.21783 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  31. [31]

    GPT-4 Technical Report

    Aaron Hurst and Adam Lerer and Adam P. Goucher and Adam Perelman and Aditya Ramesh and Aidan Clark and AJ Ostrow and Akila Welihinda and Alan Hayes and Alec Radford et al. , title =. arXiv preprint arXiv:2303.08774 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  32. [32]

    GPT-4o System Card

    Josh Achiam and Steven Adler and Sandhini Agarwal and Lama Ahmad and Ilge Akkaya and Florencia Leoni Aleman and Diogo Almeida and Janko Altenschmidt and Sam Altman and Shyamal Anadkat et al. , title =. arXiv preprint arXiv:2410.21276 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  33. [33]

    Gomez and Lukasz Kaiser and Illia Polosukhin , title=

    Ashish Vaswani and Noam Shazeer and Niki Parmar and Jakob Uszkoreit and Llion Jones and Aidan N. Gomez and Lukasz Kaiser and Illia Polosukhin , title=. Advances in Neural Information Processing Systems , year=

    work page

  34. [34]

    Manning , title=

    Kevin Clark and Urvashi Khandelwal and Omer Levy and Christopher D. Manning , title=. Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP , year=

    work page 2019

  35. [35]

    Jain and Wallace , title=. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) , year=

    work page 2019

  36. [36]

    Smith , title=

    Sofia Serrano and Noah A. Smith , title=. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , year=

    work page

  37. [37]

    Sarah Wiegreffe and Yuval Pinter , title=. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) , year=

    work page 2019

  38. [38]

    Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: System Demonstrations , year=

    Jesse Vig , title=. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: System Demonstrations , year=

    work page

  39. [39]

    Thomas McCoy and Ellie Pavlick and Tal Linzen , title=

    R. Thomas McCoy and Ellie Pavlick and Tal Linzen , title=. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , year=

    work page

  40. [40]

    Conference on Natural Language Processing , year=

    Gregor Wiedemann and Steffen Remus and Avi Chawla and Chris Biemann , title=. Conference on Natural Language Processing , year=

    work page

  41. [41]

    Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , year =

    Samira Abnar and Willem Zuidema , title =. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , year =

    work page

  42. [42]

    R. A. Bradley and M. E. Terry , title =. Biometrika , volume =. 1952 , doi =

    work page 1952

  43. [43]

    Proceedings of International Conference on Machine Learning , year =

    Guangxuan Xiao and Yuandong Tian and Beidi Chen and Song Han and Mike Lewis , title =. Proceedings of International Conference on Machine Learning , year =

    work page

  44. [44]

    Findings of the Association for Computational Linguistics: EMNLP 2025 , year =

    Chengyu Huang and Tanya Goyal , title =. Findings of the Association for Computational Linguistics: EMNLP 2025 , year =

    work page 2025

  45. [45]

    Proceedings of International Conference on Machine Learning , year =

    Ganqu Cui and Lifan Yuan and Ning Ding and Guanming Yao and Bingxiang He and Wei Zhu and Yuan Ni and Guotong Xie and Ruobing Xie and Yankai Lin and Zhiyuan Liu and Maosong Sun , title =. Proceedings of International Conference on Machine Learning , year =

    work page

  46. [46]

    Proceedings of International Conference on Learning Representations , year =

    Ilya Loshchilov and Frank Hutter , title =. Proceedings of International Conference on Learning Representations , year =

    work page

  47. [47]

    Zephyr: Direct Distillation of LM Alignment

    Lewis Tunstall and Edward Beeching and Nathan Lambert and Nazneen Rajani and Kashif Rasul and Younes Belkada and Shengyi Huang and Leandro von Werra and Clémentine Fourrier and Nathan Habib, et al , title =. arXiv preprint arXiv:2310.16944 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  48. [48]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , year =

    Ning Ding and Yulin Chen and Bokai Xu and Yujia Qin and Shengding Hu and Zhiyuan Liu and Maosong Sun and Bowen Zhou , title =. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , year =

    work page 2023

  49. [49]

    Albert Q. Jiang and Alexandre Sablayrolles and Arthur Mensch and Chris Bamford and Devendra Singh Chaplot and Diego de las Casas and Florian Bressand and Gianna Lengyel and Guillaume Lample and Lucile Saulnier and Lélio Renard Lavaud and Marie-Anne Lachaux and Pierre Stock and Teven Le Scao and Thibaut Lavril and Thomas Wang and Timothée Lacroix and Willi...

    work page internal anchor Pith review Pith/arXiv arXiv

  50. [50]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year =

    Dongfu Jiang and Xiang Ren and Bill Yuchen Lin , title =. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year =

    work page

  51. [51]

    arXiv preprint arXiv:2411.19943 , year =

    Zicheng Lin and Tian Liang and Jiahao Xu and Qiuzhi Lin and Xing Wang and Ruilin Luo and Chufan Shi and Siheng Li and Yujiu Yang and Zhaopeng Tu , title =. arXiv preprint arXiv:2411.19943 , year =

    work page arXiv