pith. sign in

arxiv: 2605.26172 · v1 · pith:7LA3PIOHnew · submitted 2026-05-25 · 💻 cs.LG

ARBITER: Reasoning Trajectory Basins and Majority Vote Failures in Test-Time Sampling

Meng Cai , Lars Kulik , Farhana Choudhury This is my paper

Pith reviewed 2026-06-29 22:19 UTC · model grok-4.3

classification 💻 cs.LG
keywords test-time samplingreasoning trajectoriesmajority votereasoning basinslanguage modelsmath benchmarksARBITER
0
0 comments X

The pith

Language model test-time sampling trajectories form reasoning basins that cause majority vote to select the most popular answer rather than the correct one, and ARBITER recovers part of the gap using only the model's own samples and hidden

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

When language models generate multiple reasoning trajectories for a question at test time, the trajectories cluster into a small number of reasoning basins, each defined by a shared normalized final answer and the solution paths that reach it. Majority vote then favors the basin with the largest number of trajectories instead of the accurate one, producing errors where the correct answer is present but outnumbered. ARBITER counters this by adding conservative evidence drawn from the base model's sampled outputs and hidden states on top of the majority consensus, rather than attempting direct correction. The method shows accuracy gains on math benchmarks across multiple model families, with no cases of net loss, and recovers a measurable share of the headroom visible to an oracle that can pick the best trajectory from the same pool.

Core claim

Reasoning trajectories concentrate into a small number of clusters, or reasoning basins, each defined by a normalized final answer and the solutions that reach it. A majority vote therefore selects the most stable basin rather than the most accurate one, which creates wrong-majority failures where the correct answer is present but outvoted. ARBITER models interactions between basins using only the base model's own sampled outputs, hidden states, and derived evidence. It applies conservative additive evidence on top of consensus rather than direct correction; in its parameter-free form it adds same-model evidence to the majority prior, while an extended form augments this with bounded residua

What carries the argument

Reasoning basins, clusters of trajectories sharing a normalized final answer and the solution paths leading to it; the mechanism explains why majority vote fails and supplies the structure that ARBITER uses to add conservative evidence.

If this is right

  • ARBITER produces consistent accuracy gains on GSM8K, MMLU-HS-Math and related benchmarks across three model families.
  • On Llama-3.1-8B MMLU-HS-Math the method lifts accuracy from the mid-78 percent range to the mid-82 percent range.
  • The gains recover roughly 22 percent of the headroom visible to a same-pool top-2 oracle.
  • Both the parameter-free ARBITER-Δ and the hidden-state-augmented ARBITER-Enc versions operate without external information or additional training.
  • No net-negative accuracy cases appear across the reported model-benchmark combinations.

Where Pith is reading between the lines

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

  • If reasoning basins remain stable when sampling parameters such as temperature change, ARBITER could be combined with temperature sweeps to enlarge the recoverable headroom.
  • The same basin structure may appear in non-mathematical reasoning tasks, allowing the conservative-evidence approach to transfer without task-specific redesign.
  • Hidden-state residuals that survive the bounded encoding step suggest that intermediate activations contain additional recoverable signals about trajectory quality beyond the final answer token.
  • Scaling the sample pool size while keeping the same conservative addition rule would likely increase the fraction of oracle headroom that can be captured internally.

Load-bearing premise

Interactions between basins can be modeled using only the base model's sampled outputs, hidden states, and derived evidence, and conservative additive evidence on top of consensus will produce gains without net-negative effects on any benchmark.

What would settle it

A controlled experiment on one of the tested model families and math benchmarks in which ARBITER applied to the same sample pool produces lower accuracy than plain majority vote.

Figures

Figures reproduced from arXiv: 2605.26172 by Farhana Choudhury, Lars Kulik, Meng Cai.

Figure 1
Figure 1. Figure 1: One wrong-majority case for ARBITER-∆. (a) Basin Story Graph: dominant basin B1=A (n=18), challenger B2=B (n=4), three singleton basins. Edges from each evidence source (§4.2) to the top-2 are weighted by the per-basin counts (b, f, p, g); the dashed red edge marks the conflicting frames between B1 and B2. The sampled pool favors A (b1=18, br=4), but the compact framed+guided main policy already favors B (… view at source ↗
Figure 2
Figure 2. Figure 2: Token-wise hidden-state kNN graph for the same example. Each panel shows a [PITH_FULL_IMAGE:figures/full_fig_p032_2.png] view at source ↗
read the original abstract

When language models use test-time sampling, they generate multiple reasoning trajectories and select an answer by majority vote. We show that these trajectories are not independent: for a given question, they concentrate into a small number of clusters, or reasoning basins, each defined by a normalized final answer and the solutions that reach it. A majority vote therefore selects the most stable basin rather than the most accurate one, which creates wrong-majority failures where the correct answer is present but outvoted. We introduce ARBITER, a model-agnostic approach that models interactions between basins using only the base model's own sampled outputs, hidden states, and derived evidence. Most direct correction strategies fail; ARBITER instead uses conservative additive evidence on top of consensus. In its simplest parameter-free form, ARBITER-{\Delta} adds same-model evidence to the majority prior, while ARBITER-Enc augments this with bounded residual signals from hidden states over complete solutions. On GSM8K with Qwen3-4B, consensus over K=24 samples achieves around the mid-94% range, while a same-pool top-2 oracle reaches around the mid-96% range. ARBITER recovers a subset of these cases using zero external information. Across three model families and three math benchmarks, it yields consistent gains with no net-negative cases; for example, on Llama-3.1-8B MMLU-HS-Math, it improves accuracy from the mid-78% range to the mid-82% range, recovering about 22% of the available oracle headroom, indicating that this headroom can be partially recovered from the sample pool itself.

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 paper claims that test-time sampling trajectories in LLMs cluster into a small number of reasoning basins (defined by normalized final answers), causing majority vote to select the most populous rather than most accurate basin and producing wrong-majority failures. ARBITER is proposed as a model-agnostic correction that models basin interactions using only the base model's sampled outputs, hidden states, and derived evidence; its simplest form (ARBITER-Δ) is parameter-free and adds conservative same-model evidence to the majority prior, while ARBITER-Enc augments this with bounded residual signals from hidden states. Experiments across three model families and three math benchmarks report consistent accuracy gains with no net-negative cases, recovering a portion (e.g., ~22%) of same-pool oracle headroom (e.g., GSM8K Qwen3-4B consensus ~mid-94% to higher with ARBITER).

Significance. If the empirical results hold, the work is significant because it identifies a structural failure mode in the widely used majority-vote aggregation for test-time scaling and supplies a practical, zero-external-data remedy that operates entirely within the sample pool. The emphasis on conservative additive evidence and the parameter-free ARBITER-Δ variant are notable strengths that could translate to reproducible improvements in reasoning benchmarks without additional training or external verifiers.

major comments (2)
  1. [Abstract] Abstract: the reported ranges (mid-94% consensus, mid-96% oracle on GSM8K; mid-78% to mid-82% on Llama-3.1-8B MMLU-HS-Math) are given without exact accuracies, standard errors, or run counts, so the statistical reliability of the 'consistent gains with no net-negative cases' claim cannot be assessed from the provided summary.
  2. [Method] Method description (implied in abstract): the claim that ARBITER-Δ is strictly parameter-free and that ARBITER-Enc uses only 'bounded residual signals' requires an explicit statement of the basin-clustering procedure, the exact additive evidence formula, and confirmation that no test-set tuning occurs, because any hidden dependence on the same sample pool would undermine the 'recovers headroom from the sample pool itself' interpretation.
minor comments (2)
  1. [Abstract] The abstract uses approximate ranges ('mid-94% range') rather than precise percentages; replace with exact values and confidence intervals in the results section.
  2. [Method] Clarify the precise definition of 'normalized final answer' used to delineate basins and how hidden-state residuals are extracted and bounded.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive evaluation, recognition of the structural failure mode in majority voting, and recommendation for minor revision. We address each major comment below and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the reported ranges (mid-94% consensus, mid-96% oracle on GSM8K; mid-78% to mid-82% on Llama-3.1-8B MMLU-HS-Math) are given without exact accuracies, standard errors, or run counts, so the statistical reliability of the 'consistent gains with no net-negative cases' claim cannot be assessed from the provided summary.

    Authors: We agree that the abstract's use of ranges without exact figures, standard errors, or run counts limits assessment of statistical reliability. In the revised version we will replace the range phrasing with precise accuracy values (including standard errors where computed), state the number of runs, and retain the 'no net-negative cases' claim only with supporting per-benchmark details moved from the main text. revision: yes

  2. Referee: [Method] Method description (implied in abstract): the claim that ARBITER-Δ is strictly parameter-free and that ARBITER-Enc uses only 'bounded residual signals' requires an explicit statement of the basin-clustering procedure, the exact additive evidence formula, and confirmation that no test-set tuning occurs, because any hidden dependence on the same sample pool would undermine the 'recovers headroom from the sample pool itself' interpretation.

    Authors: We will add an explicit subsection detailing the basin-clustering procedure (normalized final-answer grouping), the exact additive-evidence formula used by ARBITER-Δ, and the bounded-residual construction for ARBITER-Enc. We will also insert a clear statement confirming that all signals are computed from the per-question sample pool alone, with no test-set tuning or external data, thereby preserving the internal-to-the-pool interpretation. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper presents ARBITER as a parameter-free empirical method that operates exclusively on quantities already present in the base model's sample pool (outputs, hidden states, derived evidence). The central claims consist of direct experimental observations of accuracy gains across three model families and three benchmarks, with no net-negative cases and partial recovery of oracle headroom from the same pool. No equations, derivation steps, or self-citations appear in the abstract or described method that could reduce any prediction or result to a fitted input or self-definition by construction. The approach is explicitly described as adding conservative additive evidence on top of consensus rather than fitting parameters to the target metric. The derivation chain is therefore self-contained and does not exhibit any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no explicit free parameters, axioms, or invented entities; all claims rest on empirical observation of basins and internal signals.

pith-pipeline@v0.9.1-grok · 5840 in / 1042 out tokens · 36071 ms · 2026-06-29T22:19:28.326234+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

31 extracted references · 28 canonical work pages · 14 internal anchors

  1. [1]

    Phi-4 Technical Report

    Marah Abdin, Jyoti Aneja, Harkirat Behl, S \'e bastien Bubeck, Ronen Eldan, Suriya Gunasekar, Michael Harrison, Russell J. Hewett, Mojan Javaheripi, Piero Kauffmann, James R. Lee, Yin Tat Lee, Yuanzhi Li, Weishung Liu, Caio C. T. Mendes, Anh Nguyen, Eric Price, Gustavo de Rosa, Olli Saarikivi, Adil Salim, Shital Shah, Xin Wang, Rachel Ward, Yue Wu, Dingli...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  2. [2]

    ``I may not have articulated myself clearly'': Diagnosing dynamic instability in LLM reasoning at inference time

    Jinkun Chen, Fengxiang Cheng, Sijia Han, and Vlado Keselj. ``I may not have articulated myself clearly'': Diagnosing dynamic instability in LLM reasoning at inference time. arXiv preprint arXiv:2602.02863, 2026. URL https://arxiv.org/abs/2602.02863

    work page arXiv 2026

  3. [3]

    Universal self-consistency for large language model generation

    Xinyun Chen, Renat Aksitov, Uri Alon, Jie Ren, Kefan Xiao, Pengcheng Yin, Sushant Prakash, Charles Sutton, Xuezhi Wang, and Denny Zhou. Universal self-consistency for large language model generation. arXiv preprint arXiv:2311.17311, 2023. URL https://arxiv.org/abs/2311.17311

    work page arXiv 2023

  4. [4]

    Training Verifiers to Solve Math Word Problems

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021. URL https://arxiv.org/abs/2110.14168

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  5. [5]

    Can LLMs predict their own failures? self-awareness via internal circuits

    Amirhosein Ghasemabadi and Di Niu. Can LLMs predict their own failures? self-awareness via internal circuits. arXiv preprint arXiv:2512.20578, 2025. URL https://arxiv.org/abs/2512.20578

    work page arXiv 2025

  6. [6]

    The Llama 3 Herd of Models

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, et al. The Llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. URL https://arxiv.org/abs/2407.21783

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  7. [7]

    A stitch in time saves nine: Proactive self-refinement for language models

    Jinyi Han, Xinyi Wang, Haiquan Zhao, Tingyun Li, Zishang Jiang, Sihang Jiang, Jiaqing Liang, Xin Lin, Weikang Zhou, Zeye Sun, Fei Yu, and Yanghua Xiao. A stitch in time saves nine: Proactive self-refinement for language models. arXiv preprint arXiv:2508.12903, 2025. URL https://arxiv.org/abs/2508.12903

    work page arXiv 2025

  8. [8]

    Measuring Massive Multitask Language Understanding

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. In International Conference on Learning Representations, 2021 a . URL https://arxiv.org/abs/2009.03300

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  9. [9]

    Measuring Mathematical Problem Solving With the MATH Dataset

    Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the MATH dataset. In Advances in Neural Information Processing Systems, 2021 b . URL https://arxiv.org/abs/2103.03874

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  10. [10]

    Large Language Models Cannot Self-Correct Reasoning Yet

    Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Wei Yu, Xinying Song, and Denny Zhou. Large language models cannot self-correct reasoning yet. In International Conference on Learning Representations, 2024. URL https://arxiv.org/abs/2310.01798

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  11. [11]

    Semantic self-consistency: Enhancing language model reasoning via semantic weighting, 2024

    Tim Knappe, Ryan Li, Ayush Chauhan, Kaylee Chhua, Kevin Zhu, and Sean O'Brien. Semantic self-consistency: Enhancing language model reasoning via semantic weighting, 2024. URL https://arxiv.org/abs/2410.07839

    work page arXiv 2024

  12. [12]

    Efficient Memory Management for Large Language Model Serving with PagedAttention

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with PagedAttention . In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles, 2023. URL https://arxiv.org/abs/2309.06180

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  13. [13]

    Rush, and Thomas Wolf

    Quentin Lhoest, Albert Villanova del Moral, Yacine Jernite, Abhishek Thakur, Patrick von Platen, Suraj Patil, Julien Chaumond, Mariama Drame, Julien Plu, Lewis Tunstall, Joe Davison, Mario S a s ko, Gunjan Chhablani, Bhavitvya Malik, Simon Brandeis, Teven Le Scao, Victor Sanh, Canwen Xu, Nicolas Patry, Angelina McMillan-Major, Philipp Schmid, Sylvain Gugg...

    work page doi:10.18653/v1/2021.emnlp-demo.21 2021

  14. [14]

    GSM -plus: A comprehensive benchmark for evaluating the robustness of LLM s as mathematical problem solvers

    Qintong Li, Leyang Cui, Xueliang Zhao, Lingpeng Kong, and Wei Bi. GSM -plus: A comprehensive benchmark for evaluating the robustness of LLM s as mathematical problem solvers. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2961--2984, Bangkok, Thailand, 2024. Association for Computa...

    work page doi:10.18653/v1/2024.acl-long.163 2024

  15. [15]

    Entropy-gated branching for efficient test-time reasoning

    Xianzhi Li, Ethan Callanan, Abdellah Ghassel, and Xiaodan Zhu. Entropy-gated branching for efficient test-time reasoning. arXiv preprint arXiv:2503.21961, 2025. URL https://arxiv.org/abs/2503.21961

    work page arXiv 2025

  16. [16]

    CLUE : Non-parametric verification from experience via hidden-state clustering, 2025

    Zhenwen Liang, Ruosen Li, Yujun Zhou, Linfeng Song, Dian Yu, Xinya Du, Haitao Mi, and Dong Yu. CLUE : Non-parametric verification from experience via hidden-state clustering, 2025. URL https://arxiv.org/abs/2510.01591

    work page arXiv 2025

  17. [17]

    Hidden States as Early Signals: Step-level Trace Evaluation and Pruning for Efficient Test-Time Scaling

    Zhixiang Liang, Beichen Huang, Zheng Wang, and Minjia Zhang. Hidden states as early signals: Step-level trace evaluation and pruning for efficient test-time scaling. arXiv preprint arXiv:2601.09093, 2026. URL https://arxiv.org/abs/2601.09093

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  18. [18]

    Let's Verify Step by Step

    Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let's verify step by step. In International Conference on Learning Representations, 2024. URL https://arxiv.org/abs/2305.20050

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  19. [19]

    Self-refine: Iterative refinement with self-feedback

    Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, Shashank Gupta, Bodhisattwa Prasad Majumder, Katherine Hermann, Sean Welleck, Amir Yazdanbakhsh, and Peter Clark. Self-refine: Iterative refinement with self-feedback. In Advances in Neural Information Processing Sy...

    2023

  20. [20]

    GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in Large Language Models

    Iman Mirzadeh, Keivan Alizadeh, Hooman Shahrokhi, Oncel Tuzel, Samy Bengio, and Mehrdad Farajtabar. GSM-Symbolic : Understanding the limitations of mathematical reasoning in large language models. arXiv preprint arXiv:2410.05229, 2024. URL https://arxiv.org/abs/2410.05229

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  21. [21]

    Latent self-consistency for reliable majority-set selection in short- and long-answer reasoning, 2025

    Jungsuk Oh and Jay-Yoon Lee. Latent self-consistency for reliable majority-set selection in short- and long-answer reasoning, 2025. URL https://arxiv.org/abs/2508.18395

    work page arXiv 2025

  22. [22]

    PyTorch: An Imperative Style, High-Performance Deep Learning Library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas K\"opf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. PyTorch : An imperative style, high-pe...

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  23. [23]

    Internalizing LLM reasoning via discovery and replay of latent actions

    Zhenning Shi, Yijia Zhu, Junhan Shi, Xun Zhang, Lei Wang, and Congcong Miao. Internalizing LLM reasoning via discovery and replay of latent actions. arXiv preprint arXiv:2602.04925, 2026. URL https://arxiv.org/abs/2602.04925

    work page arXiv 2026

  24. [24]

    Narasimhan, and Shunyu Yao

    Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik R. Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning. In Advances in Neural Information Processing Systems, 2023. URL https://proceedings.neurips.cc/paper_files/paper/2023/hash/1b44b878bb782e6954cd888628510e90-Abstract-Conference.html

    2023

  25. [25]

    Scaling LLM test-time compute optimally can be more effective than scaling parameters for reasoning

    Charlie Victor Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. Scaling LLM test-time compute optimally can be more effective than scaling parameters for reasoning. In International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=4FWAwZtd2n

    2025

  26. [26]

    Accurate failure prediction in agents does not imply effective failure prevention

    Rakshith Vasudev, Melisa Russak, Dan Bikel, and Waseem Alshikh. Accurate failure prediction in agents does not imply effective failure prevention. arXiv preprint arXiv:2602.03338, 2026. URL https://arxiv.org/abs/2602.03338

    work page arXiv 2026

  27. [27]

    Self-Consistency Improves Chain of Thought Reasoning in Language Models

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V. Le, Ed H. Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. In International Conference on Learning Representations, 2023. URL https://arxiv.org/abs/2203.11171

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  28. [28]

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems, 2022. URL https://arxiv.org/abs/2201.11903

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  29. [29]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, R \'e mi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. Transformers: State-of-the...

    work page doi:10.18653/v1/2020.emnlp-demos.6 2020

  30. [30]

    Qwen3 Technical Report

    An Yang, Anfeng Li, Baosong Yang, et al. Qwen3 technical report. arXiv preprint arXiv:2505.09388, 2025. URL https://arxiv.org/abs/2505.09388

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  31. [31]

    Small language models need strong verifiers to self-correct reasoning

    Yunxiang Zhang, Muhammad Khalifa, Lajanugen Logeswaran, Jaekyeom Kim, Moontae Lee, Honglak Lee, and Lu Wang. Small language models need strong verifiers to self-correct reasoning. In Findings of the Association for Computational Linguistics: ACL 2024, pages 15637--15653, Bangkok, Thailand, 2024. Association for Computational Linguistics. doi:10.18653/v1/2...

    work page doi:10.18653/v1/2024.findings-acl.924 2024