Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Large Reasoning Models are not thinking straight: on the unreliability of thinking trajectories

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Large reasoning models often disregard correct solutions injected directly into their own chain of thought, and continue reasoning to a wrong answer.

desk verdict A genuinely intriguing observation about reasoning models ignoring injected correct answers, but the protocol doesn't yet prove it; deserves a careful referee. read the letter →

arxiv 2507.00711 v1 pith:PQJ73JLM submitted 2025-07-01 cs.LG

Jhouben Cuesta-Ramirez , Samuel Beaussant , Mehdi Mounsif This is my paper
classification cs.LG
keywords largereasoningmodelschain-of-thoughtoverthinkingground-truthinjectionAIME2024reinforcementlearningreliability
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Large reasoning models -- models trained to produce a visible chain of thought before answering -- are supposed to reason transparently. This paper tests what happens when a correct, fully worked solution is inserted into the middle of that chain, prefixed so it reads like the model's own next thought. Across three models of very different sizes and training recipes, on the AIME2024 mathematics benchmark, the injected ground truth is usually ignored: the model keeps generating new reasoning steps, often doubting or discarding the correct solution, and frequently ends on a wrong answer. The authors take this as evidence that current reasoning models lack a reliable mechanism for integrating external corrective signals, and that chain-of-thought traces cannot be trusted as transparent, faithful records of the reasoning that produced the answer.

What carries the argument

The load-bearing mechanism is the injection prompt $P_t = R[:t] \oplus V \oplus Y$: take the model's own trace up to the $t$-th thought, concatenate a naturalistic transition phrase $V$, then concatenate the ground-truth solution $Y$. A thought-separator heuristic based on keywords such as "Alternatively" and "Wait" splits the trace into thoughts, defines the injection points, and lets the authors count "overthinking tokens" generated after the injection. This construction turns a static benchmark problem into a counterfactual experiment: if the model genuinely reasons from its premises, it should recognise $Y$ as a valid resolution and stop; if it instead continues and errs, the trace is shown to be decoupled from correctness. The same setup supplies the paper's quantitative measures, including the completion ratio and the number of overthinking tokens produced before a correct or incorrect answer.

What would settle it

On a problem where the model rejects the injected ground truth, rewrite the solution using the model's own baseline phrasing, token choices, and punctuation, then re-inject it; if the model accepts the rewritten solution but not the original, the reported failure is an artifact of the injection format rather than a general inability to integrate correct information.

Watch

Extended reading notes

Core claim

The central claim is that large reasoning models frequently fail to accept an explicitly provided correct solution even when it is placed inside their own ongoing reasoning trace. The authors construct prompts $P_t = R[:t] \oplus V \oplus Y$, where $R[:t]$ is the model's own generated reasoning up to thought $t$, $Y$ is the ground-truth solution (reasoning plus final answer), and $V$ is a short prefix ("$\langle$think$\rangle$ Okay, so" at $t=0$, "Alternatively," otherwise) meant to make $Y$ look like the model's own continuation. Rather than adopting $Y$, the models typically produce hundreds or thousands of additional overthinking tokens, question the injected solution, and often commit to an incorrect final answer. This behavior appears across all three tested models (LLaMA 70B, Qwen 7B, DeepScalR 1.5B) and across AIME2024 problems, although the amount of overthinking is strongly problem-dependent. In the rare positive cases, the model accepted the solution only after many thoughts, sometimes explicitly out of resignation or time pressure rather than recognition of correctness.

Load-bearing premise

The load-bearing premise is that inserting the ground-truth solution with the prefix V makes it read to the model as one of its own prior thoughts, so ignoring it counts as a failure to integrate a correct answer rather than a reaction to foreign text.

Editorial extensions

If this is right

  • Chain-of-thought traces cannot be taken at face value as the model's reasoning: a model can write a long, plausible-looking trace while ignoring a correct solution it was just given.
  • External correction or human feedback injected at test time is unlikely to steer current reasoning models reliably, since they often reject the correct information rather than integrate it.
  • Benchmark gains from spending more inference compute may partly reflect spending more tokens on the same fragile heuristics rather than better understanding.
  • Evaluation metrics based on completion length or number of thoughts will mislead, because the same models show both overthinking and underthinking depending on the problem.
  • New training or prompting techniques are needed to make reasoning models receptive to corrective signals, which the paper identifies as the main direction for future work.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct test of the paper's interpretation would be to rewrite the injected ground-truth solution in the model's own discovered phrasing, tokens, and punctuation from its baseline trace; if the model then accepts the solution, the original rejection was an artifact of the injection format rather than an inability to use correct information.
  • The same disregard for externally provided correct information may extend beyond mathematics to code repair, document analysis, or any task where a correct answer is present in context while the model persists with its initial plan; the paper only tests AIME2024.
  • The behavior is consistent with reward hacking during reinforcement learning: if long traces and correct final answers are correlated in the training signal, the model may learn to keep generating tokens regardless of content, but that causal story is not established by this study.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper reports an empirical study of three reasoning models (Llama-70B, Qwen-7b, DeepScalR-1.5B) on the AIME2024 benchmark. The authors generate baseline chains of thought, segment them into 'thoughts' using a keyword heuristic, and inject the official ground-truth solution Y into the trajectory at position t via P_t = R[:t] ⊕ V ⊕ Y, with V chosen to mimic the model's writing style. They observe that the models often continue generating tokens after the injection, sometimes discarding the injected correct solution and ending on a wrong answer. They interpret this as evidence that large reasoning models lack robust mechanisms to integrate corrective signals, thereby questioning the reliability of chain-of-thought as a transparent indicator of reasoning. The paper is explicitly preliminary and ends with a limitations section.

Significance. If the effect is robust, this is a meaningful contribution to the overthinking and underthinking literature: it would challenge the view that CoT tokens are causally relevant for correctness and would have implications for human-in-the-loop correction and reward modeling. The experimental setup is simple and reproducible in principle, and it uses an external benchmark with known ground-truth answers. However, the current evidence is largely anecdotal and qualitative: there are no control conditions, no precise metric definition, no confidence intervals, and no per-model aggregate success rates after injection. The significance is therefore conditional on validation of the injection protocol, which is the load-bearing assumption of the paper.

major comments (4)
  1. [§3.2, Eq. (1)] The load-bearing assumption is that V = '<think> Okay, so ' or 'Alternatively, ' makes the appended official solution Y appear as the model's own prior thought. The paper provides no evidence for this. Official AIME solutions are human-written and can lie outside the model's RL-shaped continuation distribution, and the prefix 'Alternatively, ' may in fact signal a candidate that is later rejected, since such candidates are often abandoned in these models' trajectories. The observed disregard could therefore be a reaction to foreign or negatively cued text rather than an inability to integrate correct information. Please (i) report the model's log-probability of Y under its continuation distribution, (ii) include control injections of neutral filler text and of a plausible but incorrect solution, (iii) ablate the prefix V, and (iv) ask the model to explicitly verify Y. Without these, Eq. (1) does not establish the central claim.
  2. [§3.1, Figures 1–3] The central quantitative claim that models 'often' disregard correct solutions is not backed by aggregate statistics. The term 'overthinking tokens' is never defined precisely, and the figures provide no error bars, no number of runs or seeds, no sampling temperature, and no per-model accuracy numbers after injection. Please define the metric, report how many of the 30 AIME2024 problems led to final incorrect answers after injection for each model and each injection position t, and include confidence intervals or a statistical test to support the word 'often.'
  3. [§3.2, no control condition] There is no baseline in which the model is allowed to continue from R[:t] without any injection, and no comparison with injecting a non-solution text. Such controls are necessary to distinguish 'the model rejects correct information' from 'the model ignores any appended text,' which would be a weaker and less surprising claim. A minimal control set would compare final accuracy and token counts after injection of Y, after injection of a random wrong answer, and after no injection.
  4. [Limitations] The limitations paragraph concedes that the injection heuristic 'might not fully align with the models' internal reasoning dynamics.' This heuristic is exactly the premise on which the paper's main conclusion rests, so the concession should be treated as a validation requirement rather than a peripheral caveat. The manuscript should report experiments that directly test this alignment, such as the probability checks and control injections listed above.
minor comments (6)
  1. [§1] The phrase 'are not apart from this issue' should be 'are not exempt from this issue.'
  2. [Figure 1 caption] The caption refers to 'all four models,' but the paper states that three models were tested; please reconcile this mismatch.
  3. [§3.2, Figure 2] Please define 'completion ratio' and clarify the coloring scheme in the caption; the current description is difficult to parse.
  4. [Figure 2 caption] The typo 'wether' should be corrected to 'whether.'
  5. [References] The references list contains Ma et al. 2025a and Ma et al. 2025b, which appear to be the same preprint; please merge or distinguish them.
  6. [§3.2, quoted trace] The quoted statement 'But perhaps given the time I have...' is not accompanied by the full generated trace; please include trace excerpts in an appendix for verification.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the central claim is an externally benchmarked empirical observation, with protocol-fairness concerns that are validity issues, not circular reductions.

full rationale

The paper contains no derivation that reduces a predicted quantity to a fitted input. Its core evidence is a controlled prompting experiment: Eq. (1) defines Pt = R[:t] ⊕ V ⊕ Y, and the model's continuation is compared against the known AIME 2024 answer, an external benchmark. The reader's circularity score of 0 is correct. The only load-bearing assumption is that the V prefix makes Y appear as the model's own thought; Section 3.2 asserts this ('as if it was generated by the model itself') rather than proving it, and the Limitations section concedes that the injection 'might not fully align with the models' internal reasoning dynamics.' That is a threat to construct or external validity—the experiment may probe reactions to out-of-distribution text rather than an inability to integrate corrections—but it is not circularity: the observed disregard is not equivalent to the prompt by definition, and no fitted parameter is relabeled as a prediction. The paper also does not rely on a load-bearing self-citation chain; references to related work (Wang et al. 2025; Chen et al. 2025) motivate but do not constitute the result. Hence score 0.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The paper relies on methodological assumptions about the AIME2024 ground truth, thought segmentation, injection style matching, and model representativeness. There are no fitted parameters or invented entities.

assumptions (4)
  • domain assumption AIME2024 ground-truth solutions are correct and constitute valid corrective information.
    The paper uses these solutions as ground truth and expects models to accept them; if the solutions were incorrect or malformed, the experimental setup would be invalid.
  • ad hoc to paper The keyword-based heuristic using 'Alternatively' and 'Wait' accurately segments reasoning into meaningful thought units.
    This splitting rule is defined in Section 3.1 without validation against human annotation or alternative segmentations, and the analysis of overthinking depends on this segmentation.
  • ad hoc to paper The prefix V ('<think> Okay, so ' or 'Alternatively, ') makes the injected solution appear as the model's own generated thought.
    This assumption in Section 3.2 is load-bearing: the conclusion that models reject corrective information depends on the model treating the injected text as self-generated.
  • domain assumption The three tested models are representative of large reasoning models across training paradigms.
    The paper generalizes from LLaMA 70B, Qwen 7B, and DeepScalR 1.5B to claims about 'current LRMs' without evidence of representativeness.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Large Reasoning Models are not thinking straight: on the unreliability of thinking trajectories." pith.science (2026). https://pith.science/paper/PQJ73JLM

@misc{pith2026250700711,
  author       = {Pith},
  title        = {Pith review of: Large Reasoning Models are not thinking straight: on the unreliability of thinking trajectories},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PQJ73JLM}},
  note         = {Machine review of arXiv:2507.00711}
}
read the original abstract

Large Language Models (LLMs) trained via Reinforcement Learning (RL) have recently achieved impressive results on reasoning benchmarks. Yet, growing evidence shows that these models often generate longer but ineffective chains of thought (CoTs), calling into question whether benchmark gains reflect real reasoning improvements. We present new evidence of overthinking, where models disregard correct solutions even when explicitly provided, instead continuing to generate unnecessary reasoning steps that often lead to incorrect conclusions. Experiments on three state-of-the-art models using the AIME2024 math benchmark reveal critical limitations in these models ability to integrate corrective information, posing new challenges for achieving robust and interpretable reasoning.

Figures

Figures reproduced from arXiv: 2507.00711 by the authors.

Figure 1
Figure 1. Averaged number of overthinking tokens generated for each problem from AIME2024. Amount of overthinking tokens is problem-dependent but all four models overthink. The correct/incorrect coloring is based on whether the problem was successfully solved on the first try. nomenon emerges in all models regardless of their size and post-training method. The amount of overthinking tokens however is largely problem￾dependent… view at source ↗
Figure 2
Figure 2. Illustration of the overthinking issue on two problems from AIME2024 with DeepScalR1.5B. Pt is pictured in orange (assistant tokens) while "LRM response" represents the completion tokens with the coloring indicating wether the model predicted the correct answer or not. The completion ratio indicates the amount of completion tokens generated in proportion to the total amount of tokens [PITH_FULL_IMAGE:figures/full_f… view at source ↗
Figure 3
Figure 3. Averaged number of overthinking tokens gen￾erated for each problem from AIME2024 with model LLaMA 70B. The average is conditioned to the model providing a correct answer(green) or not(purple) at the end of its reasoning. This behavior likely stems from overfitting to specific patterns and spurious correlations that do not generalize. As shown in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Cognitive Demand Steering for Adaptive Meta-Reasoning in Large Language Models

    cs.AI 2026-08 conditional novelty 6.0 of 10

    A training-free meta-reasoning loop that tracks remaining cognitive demand and steers an LLM's next action improves average accuracy by about 9 percent over chain-of-thought across three models and six benchmarks.

Reference graph

Works this paper leans on

39 extracted references · 7 canonical work pages · cited by 1 Pith paper

  1. [1]

    Emmanuel Ameisen, Jack Lindsey, Adam Pearce, Wes Gurnee, Nicholas L. Turner, Brian Chen, Craig Citro, David Abrahams, Shan Carter, Basil Hosmer, Jonathan Marcus, Michael Sklar, Adly Templeton, Trenton Bricken, Callum McDougall, Hoagy Cunningham, Thomas Henighan, Adam Jermyn, Andy Jones, and 8 others. 2025. https://transformer-circuits.pub/2025/attribution...

    2025

  2. [2]

    Maksym Andriushchenko, Francesco Croce, and Nicolas Flammarion. 2025. https://arxiv.org/abs/2404.02151 Jailbreaking Leading Safety-Aligned LLMs with Simple Adaptive Attacks . Preprint, arXiv:2404.02151

    arXiv 2025

  3. [3]

    Chang, and Prithviraj Ammanabrolu

    Zachary Ankner, Mansheej Paul Brandon Cui, Jonathan D. Chang, and Prithviraj Ammanabrolu. 2023. https://arxiv.org/pdf/2408.11791 Critique-out-loud reward models . arXiv preprint arXiv:2408.11791

    arXiv 2023

  4. [4]

    Daman Arora and Himanshu Gaurav Singh. 2023. https://arxiv.org/pdf/2305.15074 Have llms advanced enough? a challenging problem solving benchmark for large language models . arXiv preprint arXiv:2305.15074

    arXiv 2023

  5. [5]

    Ashok, and Shashank Shet

    Ramakrishna Bairi, Atharv Sonwane, Aditya Kanade, Vageesh D C, Arun Iyer, Suresh Parthasarathy, Sriram Rajamani, B. Ashok, and Shashank Shet. 2023. https://arxiv.org/abs/2309.12499 CodePlan: Repository-level Coding using LLMs and Planning . Preprint, arXiv:2309.12499

    arXiv 2023

  6. [6]

    Bradley Brown, Jordan Juravsky, Ryan Saul Ehrlich, Ronald Clark, Quoc V Le, Christopher Re, and Azalia Mirhoseini. 2025. https://openreview.net/forum?id=0xUEBQV54B Large Language Monkeys: Scaling Inference Compute with Repeated Sampling

    work page 2025

  7. [7]

    Xingyu Chen, Jiahao Xu, Tian Liang, Zhiwei He, Jianhui Pang, Dian Yu, Linfeng Song, Qiuzhi Liu, Mengfei Zhou, Zhuosheng Zhang, Rui Wang, Zhaopeng Tu, Haitao Mi, and Dong Yu. 2025. https://arxiv.org/abs/2412.21187 Do not think that much for 2+3=? on the overthinking of o1-like llms . Preprint, arXiv:2412.21187

    arXiv 2025

  8. [8]

    Kewei Cheng, Jingfeng Yang, Haoming Jiang, Zhengyang Wang, Binxuan Huang, Ruirui Li, Shiyang Li, Zheng Li, Yifan Gao, Xian Li, Bing Yin, and Yizhou Sun. 2024. https://arxiv.org/abs/2408.00114 Inductive or Deductive? Rethinking the Fundamental Reasoning Abilities of LLMs . Preprint, arXiv:2408.00114

    arXiv 2024

Show all 39 references
  1. [9]

    Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert Huben, and Lee Sharkey. 2023. https://arxiv.org/abs/2309.08600 Sparse autoencoders find highly interpretable features in language models . Preprint, arXiv:2309.08600

    2023 arXiv

  2. [10]

    DeepSeek-AI. 2025. https://arxiv.org/abs/2501.12948 Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning . Preprint, arXiv:2501.12948

    2025 arXiv

  3. [11]

    Jacob Dunefsky, Philippe Chlenski, and Neel Nanda. 2024. https://arxiv.org/abs/2406.11944 Transcoders find interpretable llm feature circuits . Preprint, arXiv:2406.11944

    2024 arXiv

  4. [12]

    Zitian Gao, Boye Niu, Xuzheng He, Haotian Xu, Hongzhang Liu, Aiwei Liu, Xuming Hu, and Lijie Wen. 2023. https://arxiv.org/pdf/2410.01707 Interpretable contrastive monte carlo tree search reasoning . arXiv preprint arXiv:2410.01707

    2023 arXiv

  5. [13]

    Zhenyu Hou, Xin Lv, Rui Lu, Jiajie Zhang, Yujiang Li, Zijun Yao, Juanzi Li, Jie Tang, and Yuxiao Dong. 2025. https://arxiv.org/pdf/2501.11651 Advancing language model reasoning through reinforcement learning and inference scaling . arXiv preprint arXiv:2501.11651

    2025 arXiv

  6. [14]

    Subbarao Kambhampati, Karthik Valmeekam, Lin Guan, Mudit Verma, Kaya Stechly, Siddhant Bhambri, Lucas Saldyt, and Anil Murthy. 2024. https://arxiv.org/abs/2402.01817 Llms can't plan, but can help planning in llm-modulo frameworks . Preprint, arXiv:2402.01817

    2024 arXiv

  7. [15]

    Miranda, Alisa Liu, Nouha Dziri, Shane Lyu, Yuling Gu, Saumya Malik, Victoria Graf, Jena D

    Nathan Lambert, Jacob Morrison, Valentina Pyatkin, Shengyi Huang, Hamish Ivison, Faeze Brahman, Lester James V. Miranda, Alisa Liu, Nouha Dziri, Shane Lyu, Yuling Gu, Saumya Malik, Victoria Graf, Jena D. Hwang, Jiangjiang Yang, Ronan Le Bras, Oyvind Tafjord, Chris Wilhelm, Luc...

    2025 arXiv

  8. [16]

    Tom Lieberum, Senthooran Rajamanoharan, Arthur Conmy, Lewis Smith, Nicolas Sonnerat, Vikrant Varma, János Kramár, Anca Dragan, Rohin Shah, and Neel Nanda. 2024. https://arxiv.org/abs/2408.05147 Gemma scope: Open sparse autoencoders everywhere all at once on gemma 2 . Preprint,...

    2024 arXiv

  9. [17]

    Jack Lindsey, Wes Gurnee, Emmanuel Ameisen, Brian Chen, Adam Pearce, Nicholas L. Turner, Craig Citro, David Abrahams, Shan Carter, Basil Hosmer, Jonathan Marcus, Michael Sklar, Adly Templeton, Trenton Bricken, Callum McDougall, Hoagy Cunningham, Thomas Henighan, Adam Jermyn, A...

    2025

  10. [18]

    Tang, Manan Roongta, Colin Cai, Jeffrey Luo, Li Erran Li, Raluca Ada Popa, and Ion Stoica

    Michael Luo, Sijun Tan, Justin Wong, Xiaoxiang Shi, William Y. Tang, Manan Roongta, Colin Cai, Jeffrey Luo, Li Erran Li, Raluca Ada Popa, and Ion Stoica. 2025. Deepscaler: Surpassing o1-preview with a 1.5b model by scaling rl. https://pretty-radio-b75.notion.site/DeepScaleR-Su...

    2025

  11. [20]

    Wenjie Ma, Jingxuan He, Charlie Snell, Tyler Griggs, Sewon Min, and Matei Zaharia. 2025 b . https://arxiv.org/abs/2504.09858 Reasoning models can be effective without thinking . Preprint, arXiv:2504.09858

    2025 arXiv

  12. [21]

    Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès, and Tatsunori Hashimoto. 2025. https://arxiv.org/abs/2501.19393 s1: Simple test-time scaling . Preprint, arXiv:2501.19393

    2025 arXiv

  13. [22]

    OpenAI. 2024. https://cdn.openai.com/o1-system-card-20241205.pdf Openai o1 system card

    2024

  14. [23]

    OpenAI. 2025. https://cdn.openai.com/pdf/2221c875-02dc-4789-800b-e7758f3722c1/o3-and-o4-mini-system-card.pdf Openai o3 and o4-mini system card

    2025

  15. [24]

    Qwen-Team. 2024. https://qwenlm.github.io/blog/qwq-32b-preview/ Qwq: Reflect deeply on the boundaries of the unknown

    2024

  16. [25]

    Shyam Sundhar Ramesh, Yifan Hu, Iason Chaimalas, Viraj Mehta, Pier Giuseppe Sessa, Haitham Bou Ammar, and Ilija Bogunovic. 2024. https://arxiv.org/abs/2405.20304 Group robust preference optimization in reward-free rlhf . Preprint, arXiv:2405.20304

    2024 arXiv

  17. [26]

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. http://dblp.uni-trier.de/db/journals/corr/corr1707.html#SchulmanWDRK17 Proximal policy optimization algorithms. CoRR, abs/1707.06347

    2017 arXiv

  18. [27]

    Yunhao Tang, Kunhao Zheng, Gabriel Synnaeve, and Rémi Munos. 2025. https://arxiv.org/abs/2503.19595 Optimizing language models for inference time objectives using reinforcement learning . Preprint, arXiv:2503.19595

    2025 arXiv

  19. [28]

    Ye Tian, Baolin Peng, Linfeng Song, Lifeng Jin, Dian Yu, Haitao Mi, and Dong Yu. 2024. https://arxiv.org/abs/2404.12253 Toward Self-Improvement of LLMs via Imagination, Searching, and Criticizing . Preprint, arXiv:2404.12253

    2024 arXiv

  20. [29]

    Karthik Valmeekam, Kaya Stechly, and Subbarao Kambhampati. 2024. https://arxiv.org/abs/2409.13373 Llms still can't plan; can lrms? a preliminary evaluation of openai's o1 on planbench . Preprint, arXiv:2409.13373

    2024 arXiv

  21. [30]

    Junlin Wang, Jue Wang, Ben Athiwaratkun, Ce Zhang, and James Zou. 2024. https://arxiv.org/abs/2406.04692 Mixture-of-Agents Enhances Large Language Model Capabilities . Preprint, arXiv:2406.04692

    2024 arXiv

  22. [31]

    Lei Wang, Wanyu Xu, Yihuai Lan, Zhiqiang Hu, Yunshi Lan, Roy Ka-Wei Lee, and Ee-Peng Lim. 2023. https://arxiv.org/pdf/2305.04091 Plan-and-solve prompting: Improving zero-shot chain-of-thought reasoning by large language models . arXiv preprint arXiv:2305.04091

    2023 arXiv

  23. [32]

    Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed H. Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2022. https://arxiv.org/pdf/2203.11171 Self-consistency improves chain of thought reasoning in language models . arXiv preprint arXiv:2203.11171

    2022 arXiv

  24. [33]

    Yue Wang, Qiuzhi Liu, Jiahao Xu, Tian Liang, Xingyu Chen, Zhiwei He, Linfeng Song, Dian Yu, Juntao Li, Zhuosheng Zhang, Rui Wang, Zhaopeng Tu, Haitao Mi, and Dong Yu. 2025. https://arxiv.org/abs/2501.18585 Thoughts are all over the place: On the underthinking of o1-like llms ....

    2025 arXiv

  25. [34]

    Chi, Quoc V Le, and Denny Zhou

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian ichter, Fei Xia, Ed H. Chi, Quoc V Le, and Denny Zhou. 2022. https://openreview.net/forum?id=_VjQlMeSB_J Chain of Thought Prompting Elicits Reasoning in Large Language Models . In Advances in Neural Information Proc...

    2022

  26. [35]

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. 2023. https://proceedings.neurips.cc/paper_files/paper/2023/file/271db9922b8d1f4dd7aaef84ed5ac703-Paper-Conference.pdf Tree of Thoughts: Deliberate Problem Solving with Large Lan...

    2023

  27. [36]

    Edward Yeo, Yuxuan Tong, Morry Niu, Graham Neubig, and Xiang Yue. 2025. https://arxiv.org/abs/2502.03373 Demystifying long chain-of-thought reasoning in llms . Preprint, arXiv:2502.03373

    2025 arXiv

  28. [37]

    Yang Yue, Zhiqi Chen, Rui Lu, Andrew Zhao, Zhaokai Wang, Shiji Song, and Gao Huang. 2025. Does reinforcement learning really incentivize reasoning capacity in llms beyond the base model? arXiv preprint arXiv:2504.13837

    2025 arXiv

  29. [38]

    Zhi Zheng, Zhuoliang Xie, Zhenkun Wang, and Bryan Hooi. 2025. https://arxiv.org/abs/2501.08603 Monte Carlo Tree Search for Comprehensive Exploration in LLM-Based Automatic Heuristic Design . Preprint, arXiv:2501.08603

    2025 arXiv

  30. [39]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  31. [40]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.