CoRE: Enhancing Metacognition with Label-free Self-evaluation in LRMs

Haoxi Li; Jie Zhang; Sikai Bai; Song Guo

REVIEW 4 major objections 5 minor 1 cited by

CoRE: Enhancing Metacognition with Label-free Self-evaluation in LRMs

T0 review · 4 major / 5 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 The paper claims that quasi-periodic fluctuations in a reasoning model's hidden-state trajectory reveal when it is stuck in a loop, and that early exit at those moments both shortens the chain of thought and improves accuracy.

desk verdict Promising but unproven early-exit method: the cycle-detection signal isn't isolated from the benefit of simply truncating long CoT. read the letter →

arxiv 2507.06087 v1 pith:ZZJPKYFU submitted 2025-07-08 cs.LG

Haoxi Li , Sikai Bai , Jie Zhang , Song Guo This is my paper

classification cs.LG

keywords label-freeself-evaluationchain-of-reasoningembeddingoverthinkingearlyexitlargereasoningmodelslatenttrajectorymetacognitioncyclicredundancydetection

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

The paper tries to establish that a large reasoning model can judge, without any external label or fine-tuning, whether its own intermediate reasoning has become redundant, by watching the geometry of its hidden states across reasoning steps. It introduces CoRE, the Chain-of-Reasoning Embedding, the sequence of last-token hidden states of each step, and shows that quasi-periodic oscillations of magnitude and angle in this trajectory correspond to repetitive 'overthinking' loops. On that basis CoRE-Eval detects the loop with a sliding-window correlation and triggers an early exit. The reported result is that chain-of-thought length drops by 13.7% to 33.2% while accuracy rises, including a 10-point gain to 70% on AIME 2024 with a 32B model.

What carries the argument

The load-bearing object is CoRE, the Chain-of-Reasoning Embedding: for a chain-of-thought $S_1,\dots,S_T$, each step is embedded by taking the last token's last-layer hidden state $h_t = \mathrm{Encoder}^{(L)}(S_t)[-1]$, with each step encoded independently using its own positional encoding. From adjacent embeddings the method computes the magnitude change $\delta^{\mathrm{mag}}_t = \|h_{t+1}-h_t\|_2$ and the cosine similarity $c^{\mathrm{ang}}_t$, combined into the composite signal $z_t = \delta^{\mathrm{mag}}_t(1-c^{\mathrm{ang}}_t)$. A sliding window of size $W=32$ compares $z$ against its lagged versions for candidate periods up to $P_{\max}=8$, and a hysteresis controller requires the correlation $\rho_t$ to exceed $\rho^*=0.7$ for $M$ consecutive steps before declaring a cycle.

What would settle it

A direct test would be to rerun CoRE-Eval on the same reasoning traces but replace the independent step embeddings with context-aware hidden states (for example, the last-token representation of the full prefix ending at each step) and compare detected cycles against human-annotated redundant steps. If the quasi-periodic signal largely disappears or stops correlating with genuinely redundant reasoning, the central claim fails; a cheaper check is to shuffle the step order and re-encode, in which case persisting cycles would indicate artifacts of the independent encoding.

Watch

Extended reading notes

Core claim

The central claim is that redundant reasoning in large reasoning models produces a characteristic latent-space signature: the step-embedding trajectory oscillates quasi-periodically, with magnitude changes and angular reversals anticorrelated, while efficient reasoning produces a pruned, more directional trajectory. CoRE-Eval operationalizes this by computing a composite signal $z_t = \delta^{\mathrm{mag}}_t (1 - c^{\mathrm{ang}}_t)$ and detecting sustained self-similarity at a candidate period with a sliding-window Pearson correlation; only when the same period is confirmed for $M$ consecutive steps does a finite-state hysteresis controller terminate reasoning and force a final answer. The paper asserts this is the first training-free, label-free self-evaluation mechanism for reasoning models, and that it consistently reduces token length while improving accuracy across GSM8K, MATH-500, AIME 2024, and additional math and code benchmarks.

Load-bearing premise

The method's load-bearing premise is that encoding each reasoning step alone, with its own positional encoding and no context of prior steps or the prompt, produces an embedding that faithfully represents what the model is doing at that step; if those isolated step embeddings are artifacts of the encoding rather than genuine reasoning-state representations, the detected cycles and the whole self-evaluation signal rest on an invalid foundation.

Editorial extensions

If this is right

If this is right, a training-free signal can shorten long chain-of-thought at inference time, reducing compute without degrading answer quality and in most settings improving it.
The geometric loop detector could serve as a general metacognitive monitor for any white-box reasoning model, flagging when the model begins to repeat itself.
The reported scaling behavior suggests the benefit grows with model capability: the 32B model gains the most on AIME, suggesting stronger hidden-state geometry in more capable models.
Because it is label-free and prompt-agnostic, the method can be layered onto existing white-box LRMs without dataset construction or fine-tuning.
The signal transfers beyond mathematics to competition math and code generation, so the loop pattern is not an artifact of one task distribution.

Reading between the lines

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

Editorially, the same trajectory geometry could be used not only to exit early but to decide when to intervene, re-prompt, or backtrack, turning loop detection into a general control signal for reasoning.
One testable extension, not attempted here, is to use the quasi-periodicity signal as an auxiliary reward or regularizer in reinforcement learning to discourage overthinking during training.
The independent step embedding is a strong simplifying choice; a context-aware variant that encodes each step with the full prefix might yield cleaner cycles and better detection, though the paper does not compare them.
A natural stress test is to apply the detector to non-mathematical long-form reasoning, such as planning or open-ended generation, where repetitive loops may look different and the threshold may need recalibration.

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.

Desk Editor's Note

read the letter

Here's the short version: this is a promising test-time early-exit method for white-box reasoning models, built on a new geometric signal (magnitude times angle complement) with a sliding-window periodicity detector and hysteresis control. The idea is real, and the empirical study is reasonably thorough—three benchmarks, three model sizes, plus two extra tasks in the appendix. The authors also clearly state the limitations (white-box only, latency overhead). But the central claim—that quasi-periodic fluctuations identify useful stopping points—is not actually tested, because there is no length-matched baseline that truncates at the same token count without the cycle detector. Without that control, the reported gains could just be the effect of cutting off a rambling CoT tail, not evidence that the detector itself carries information.

The strongest parts: the composite signal is novel as far as the cited prior work goes, and the authors are honest about tuning Pmax and window size from qualitative inspection. The paper also correctly notes the difficulty of label-free evaluation in this setting.

The main soft spots, in order: (1) the missing truncation control is load-bearing; I'd want to see random or fixed-length early exit at the same average budget before believing the cycle detector adds value. (2) Hyperparameters ρ*=0.7, M=8, Pmax=8, W=32 are tuned on the same benchmarks used for evaluation, which makes the reported results partially circular. (3) The AIME result is 30 problems; 70% vs 60% is three answers, so that headline should be read with caution. (4) The abstract says "improving answer accuracy by around 10%" while the average improvement in Table 1 is about 3.6%—that inconsistency needs fixing. (5) The "first label-free" claim is undercut by the paper's own citations, e.g., [36] and [43] do unsupervised or output-free internal-state evaluation; the authors should narrow the claim. The independent-step encoding in Eq. (1) is a more minor concern, since the method's viability depends less on that than on the missing baseline.

Who is this for? Researchers working on efficient reasoning and early exit in LLMs. It's a useful read, and the method is plausible enough to warrant a serious referee—but only if the authors add a length-matched control and temper the claims. I'd send it to review with the expectation of major revision.

Referee Report

4 major / 5 minor

Summary. The paper proposes CoRE-Eval, a training-free and label-free early-exit mechanism for large reasoning models (LRMs). It defines a Chain-of-Reasoning Embedding (CoRE) as the sequence of step-level hidden states h_t = Encoder(L)(S_t)[-1], then extracts geometric signals (L2 magnitude change and cosine angle change) to form a composite signal z_t = delta_mag_t (1 - cos_t). Quasi-periodic fluctuations in z_t are detected via a sliding-window Pearson correlation over candidate periods, with a hysteresis controller (rho*, M, Pmax, W) that triggers early exit by appending a 'Final Answer:' prompt. Experiments on GSM8K, MATH-500, and AIME 2024 with DeepSeek-R1-Distill 7B/14B/32B report CoT length reductions of 13.7-33.2% and accuracy improvements, including a 10-point absolute gain on AIME with the 32B model. Additional results on AIMO Validation AMC and HumanEval with the 7B model are reported. The paper also includes trajectory visualizations and a case study showing an early-exit example.

Significance. If the central claim holds, CoRE-Eval would be a practically valuable contribution: a white-box, training-free, label-free signal that identifies when an LRM begins to loop and can terminate generation productively, saving inference cost without sacrificing accuracy. The paper is commendably precise in its algorithmic description, provides illustrative visualizations of the purported cyclic patterns, and includes an honest limitations section acknowledging white-box requirements and latency overhead. However, the empirical evidence as presented does not yet isolate the value of the quasi-periodicity signal from the effect of truncation itself, and the key representation choice at Eq. (1) is not validated. The contribution is promising but not yet established.

major comments (4)

[4.3, Table 1] The reported accuracy and length gains are confounded with the effect of truncation itself. CoRE-Eval always exits at a detector-selected position; without a control that truncates the same model at the same token budgets using a position-only or random rule (e.g., appending 'Final Answer:' at fixed token counts), Table 1 cannot establish that quasi-periodicity carries information about good stopping points. Please add such a length-matched truncation baseline for each model and benchmark.
[3, Eq. (1)] The step embedding h_t = Encoder(L)(S_t)[-1] encodes each reasoning step independently with its own positional encoding, ignoring the prior steps and the full prompt context. Unless this is demonstrated to faithfully represent the model's reasoning state, the magnitude and angle dynamics underlying Findings 1-3 may be artifacts of the isolated encoding rather than genuine properties of the reasoning trajectory. Please validate against context-aware hidden states (e.g., hidden states from processing the full prefix with causal attention) or provide evidence that independent-step embeddings preserve the trajectory dynamics.
[5.1, 5.3, 6.3] Key hyperparameters (rho*=0.7, M=8, Pmax=8, W=32) are chosen after qualitative observation of CoRE trajectories on the same benchmarks used for evaluation (MATH-500 and GSM8K) and tuned via ablations on MATH-500 (Fig. 6). Evaluating on those same benchmarks is therefore partially circular for the claim of label-free self-evaluation. Please report hyperparameter sensitivity on held-out tasks, use a separate validation split, or clarify how 'label-free' is maintained when benchmark labels guide hyperparameter selection.
[5.2, Table 1] The AIME result is based on only 30 problems; the headline +10% accuracy on the 32B model corresponds to three additional correct answers, and no confidence intervals or significance tests are reported for any accuracy numbers. The claim of consistent accuracy improvement should be statistically qualified, especially for small-sample benchmarks.

minor comments (5)

[Abstract / 5.2] The abstract says 'improving answer accuracy by around 10%', while Section 5.2 states 'enhances accuracy by an average of 3.58%'. Please reconcile these numbers and clarify which comparison each refers to.
[Footnote 2] The footnote says 'We use selected samples from MATH-500 and GSM8K for qualitative analysis. The evaluation set contains 1,269 and 480 problems, respectively.' The order appears reversed; it should be GSM8K (1,269) and MATH-500 (480) if the described exclusions are applied.
[Figure 6] The submitted text for Figure 6 is corrupted with escape sequences and lacks clear axis labels and legend. Please replot the ablation figure so that the effects of rho* and M are legible.
[Table 1] In the 14B MATH-500 row for DEER, the accuracy change is written as '+0.2s', which appears to be a typo for '+0.2' or '+0.20'.
[5.1] The evaluation uses greedy decoding with a single sample; reporting variance across multiple decoding runs or providing per-problem error bars would strengthen the robustness claims.

Circularity Check

2 steps flagged · score 6.0 of 10

CoRE's quasi-periodicity signal is partly tautological under the independent-step embedding of Eq. (1), and the detector's key thresholds are tuned on the same benchmarks used to report the headline gains.

self definitional [Sec. 3, Eq. (1); Sec. 4.1, Finding 2]
"ht = f (St) =Encoder(L)(St)[−1] ∈ Rd (1) ... St is processed as an input segment with its own positional encoding. ... Finding 2: Quasi-periodic fluctuations in CoRE trajectories reliably signal reasoning loops and verbal redundancy, reflecting inefficient cognitive processing."

By Eq. (1), each CoRE point ht is computed by encoding the step text St alone, 'with its own positional encoding' and without conditioning on previous steps. Consequently, any verbatim repetition in the reasoning trace is mapped to identical (or near-identical) embedding values, and the resulting z-sequence is periodic exactly where the text repeats. The 'finding' that CoRE quasi-periodicity signals reasoning loops is therefore true by construction for the loop examples shown (e.g., repeated 'Wait, let me compute...' steps in Fig. 2); it is a re-encoding of lexical repetition, not an independent discovery about the model's latent reasoning state. The paper presents this definitional equivalence as empirical evidence for label-free metacognition.
fitted input called prediction [Sec. 5.3 Ablation Study; Sec. 5.1 Implementation Details; Sec. 6.3 Hyperparameter Settings]
"Through qualitative observations of CoRE trajectories on samples from MATH-500 and GSM8K, we found that the dominant periodic patterns typically occur within cycles of length 2 to 6, with no cycles exceeding 8 steps. Based on this empirical insight, we set the maximum period Pmax = 8. ... The best result is observed at ρ∗ = 0.7, yielding an accuracy of 88.20% with an average token length of 1906. ... The best trade-off is achieved at M = 8, with an accuracy of 88.20% and a reasoning length of 1906 tokens."

The detector hyperparameters are selected by optimizing accuracy and length on the same benchmarks that later certify the method. The ablation's 1906-token operating point corresponds to the MATH-500 CoRE-Eval row in Table 1 (1906.0 tokens), so ρ*=0.7 and M=8 are effectively chosen on the MATH-500 evaluation set, and Pmax/W are fixed from qualitative observations on MATH-500 and GSM8K rather than an independent validation split. The reported accuracy/length gains are therefore in-sample results of a threshold search, not label-free predictions on unseen data.

full rationale

CoRE-Eval is a training-free heuristic, so the absence of a theoretical first-principles derivation is not itself circularity. Two load-bearing steps do exhibit circularity, however. First, Eq. (1) encodes each reasoning step independently, so a loop of repeated sentences is mapped, by construction, to a quasi-periodic CoRE trajectory; the paper's Finding 2 is then a restatement of lexical repetition in embedding coordinates rather than an independent latent-state discovery. Second, the main detector parameters (ρ*, M, W, Pmax) are calibrated on MATH-500 and GSM8K material, and the ablation reports the best operating point in terms of accuracy and token length on those same datasets; Table 1's improvements are thus partly the result of in-sample tuning. The separate missing control—a length-matched truncation baseline without any trajectory signal—is a serious attribution confound for the claim that quasi-periodicity carries the information, but it is an experimental weakness rather than a definitional circularity. Because the empirical accuracy/efficiency outcome does not collapse entirely into the construction, the appropriate score is 6: partial circularity, not total.

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

The central method rests on several free hyperparameters that are tuned or fitted on the same benchmarks used for reporting, plus the untested assumption that per-step independent hidden states faithfully represent the reasoning state. No new physical or conceptual entities are introduced.

free parameters (4)

rho* (confidence threshold) = 0.7
Tuned via ablation on MATH-500 (Sec. 5.3) to balance accuracy and length; no held-out validation set is used.
M (stability duration) = 8
Tuned via ablation (Sec. 5.3); best at M=8, with larger values reducing both accuracy and efficiency.
Pmax (maximum period) = 8
Set based on qualitative observations of CoRE trajectories on 50 GSM8K and 20 MATH-500 samples (Sec. 6.3), the same benchmarks used for evaluation.
W (window size) = 32
Derived as 4*Pmax (Sec. 6.3), so it inherits the tuning of Pmax and is not independently validated.

assumptions (4)

domain assumption Step-level hidden states computed by encoding each step separately capture the state of reasoning
Eq. (1) encodes each step S_t independently with its own positional encoding, not conditioned on previous steps; the paper does not validate this against context-aware embeddings.
ad hoc to paper Quasi-periodic fluctuation of the composite signal z_t reliably indicates unproductive reasoning loops
Finding 2 is based on visual inspection of a small number of examples; there is no quantitative validation that periodic z_t correlates with reasoning errors or redundancy.
domain assumption Early termination at a detected cycle preserves or improves answer correctness
The paper assumes the model has already reached a satisfactory state when the cycle is detected; no oracle analysis is provided comparing early exit to full-length reasoning.
standard math Pearson correlation on normalized segments detects periodicity relevant to reasoning loops
Sliding-window correlation is a standard technique, but its suitability for non-stationary reasoning trajectories is assumed rather than demonstrated.

how reviews work

0 comments

Cite this review

Pith. "Pith review of CoRE: Enhancing Metacognition with Label-free Self-evaluation in LRMs." pith.science (2026). https://pith.science/paper/ZZJPKYFU

@misc{pith2026250706087,
  author       = {Pith},
  title        = {Pith review of: CoRE: Enhancing Metacognition with Label-free Self-evaluation in LRMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZZJPKYFU}},
  note         = {Machine review of arXiv:2507.06087}
}

read the original abstract

Large reasoning models (LRMs) have demonstrated impressive capabilities in domains like mathematics and program synthesis. Despite their strong performance, LRMs often exhibit overthinking -- excessive and redundant reasoning steps that introduce inefficiencies during inference. This phenomenon raises an important question for LRM self-evaluation: How can a model autonomously assess the correctness of its own reasoning trajectory without external labels? To address this, we propose Chain-of-Reasoning Embedding (CoRE), a series of hidden states in latent space to enable label-free self-evaluation on intermediate reasoning steps of LRMs, so as to enhance metacognition abilities for improved reasoning efficiency. By analyzing the geometric properties of the CoRE trajectories, we reveal that redundant reasoning usually presents cyclical fluctuations, which correspond to repetitive and unconscious reflection/exploration. Leveraging this insight, we further introduce a training-free, label-free self-evaluation framework, CoRE-Eval, to detect such patterns and dynamically determine whether to terminate reasoning early. Extensive experiments on mathematical reasoning benchmarks (GSM8K, MATH-500, and AIME) and across model sizes from 7B to 32B demonstrate that CoRE-Eval reduces chain-of-thought length by 13.7% to 33.2% while improving answer accuracy by around 10%, achieving 70.0% accuracy on the challenging AIME benchmark with the 32B model.

Figures

Figures reproduced from arXiv: 2507.06087 by the authors.

**Figure 2.** CoRE Visualization. Blue and Red Long CoT trajectories represent changes in the angle and magnitude dimensions, respectively. Shaded regions highlight notable reasoning patterns. {87.4, 51.0, 48.7, 46.2, 88.1, 51.3}, {48.7, 46.2, 88.1, 51.3, 48.3, 46.1}. {87.4, 51.0, 48.7, 46.2, 88.1}, {46.2, 88.1, 51.3, 48.3, 46.1}. {87.4, 51.0, 48.7, 46.2}, {88.1, 51.3, 48.3, 46.1}. No Answer Step 185: "Wait, let me compute (1 +2i… view at source ↗

**Figure 3.** An overview of the Chain-of-Embedding Reasoning Self-evaluation ( [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (10 more)

**Figure 4.** Figure 4: Relative accuracy vs. token length reduction [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 5.** Figure 5: CoRE-Eval vs. Original (i.e., on AIME). In addition, the breakdown in [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

**Figure 6.** Figure 6: Ablation Study. Impact of ρ ∗ . The threshold ρ ∗ controls the sensitivity of cycle detection by specifying the minimum correlation strength required to recognize a cyclic pattern. As shown in [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

**Figure 7.** Figure 7: Prompt Template for CoRE-Eval [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]

**Figure 8.** Figure 8: shows the prompt that enables the LRMs to terminate the reasoning process once cyclic redundancy is detected by CoRE-Eval. Prompt For Early Exit in CoRE-Eval “…Okay, I have finished thinking.\n</think>\n\n. Final Answer:" [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]

**Figure 9.** Figure 9: Prompt in D-Prompt [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]

**Figure 10.** Figure 10: is the setting of baseline NoThinking. Prompt in NoThinking "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n Please reason step by step, and put your final answer within \\boxed{}.\n" + f"{mess['content']}<|im_end|>\n<|im_start|>assistan…

**Figure 11.** Figure 11: CoRE-Eval Visualization Example in MATH-500. Question "If $n \\equiv 2 \\pmod{7}$, then find the remainder when $(n + 2)(n + 4)(n + 6)$ is divided by 7. Cosine Value Magnitude Step [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]

**Figure 13.** Figure 13: CoRE-Eval Visualization Example in GSM8K 16 [PITH_FULL_IMAGE:figures/full_fig_p016_13.png]

**Figure 14.** Figure 14: An example demonstrating CoRE-Eval Metacognition capability and concise response. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_14.png]

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

Is Your Model Thinking or Just Stagnating? PUMA: Diagnosing Reasoning Pathology via Phase-Momentum Alignment
cs.AI 2026-07 conditional novelty 5.0 of 10

PUMA detects when a reasoning model's entropy drop aligns with hidden-state momentum, truncates at that point, and reports improved accuracy-efficiency on 1.5B-32B reasoning models.

Reference graph

Works this paper leans on

51 extracted references · 16 canonical work pages · cited by 1 Pith paper

[36]

Unsuper- vised real-time hallucination detection based on the internal states of large language models

Weihang Su, Changyue Wang, Qingyao Ai, Yiran Hu, Zhijing Wu, Yujia Zhou, and Yiqun Liu. Unsuper- vised real-time hallucination detection based on the internal states of large language models. In Findings of the 62nd Annual Meeting of the Association for Computational Linguistics, ACL 2024, pages 14379–14391. Association for Computational Linguistics (ACL), 2024

work page 2024
[43]

Latent space chain-of-embedding enables output-free llm self-evaluation

Yiming Wang, Pei Zhang, Baosong Yang, Derek F Wong, and Rui Wang. Latent space chain-of-embedding enables output-free llm self-evaluation. arXiv preprint arXiv:2410.13640, 2024

arXiv 2024
[1]

L1: Controlling how long a reasoning model thinks with reinforcement learning

Pranjal Aggarwal and Sean Welleck. L1: Controlling how long a reasoning model thinks with reinforcement learning. arXiv preprint arXiv:2503.04697, 2025

arXiv 2025
[2]

Sketch-of-thought: Efficient llm reasoning with adaptive cognitive-inspired sketching

Simon A Aytes, Jinheon Baek, and Sung Ju Hwang. Sketch-of-thought: Efficient llm reasoning with adaptive cognitive-inspired sketching. arXiv preprint arXiv:2503.05179, 2025

arXiv 2025
[3]

Discovering latent knowledge in language models without supervision

Collin Burns, Haotian Ye, Dan Klein, and Jacob Steinhardt. Discovering latent knowledge in language models without supervision. In The Eleventh International Conference on Learning Representations, 2023

2023
[4]

The dynamics of cortical activity during sensorimotor behavior

Rishidev Chaudhuri, Boris Ger¸ sgorin, Zhixin Wang, Mark M Churchland, and Surya Ganguli. The dynamics of cortical activity during sensorimotor behavior. Neuron, 102(4):768–781, 2019

work page 2019
[5]

Towards reasoning era: A survey of long chain-of-thought for reasoning large language models

Qiguang Chen, Libo Qin, Jinhao Liu, Dengyun Peng, Jiannan Guan, Peng Wang, Mengkang Hu, Yuhang Zhou, Te Gao, and Wanxiang Che. Towards reasoning era: A survey of long chain-of-thought for reasoning large language models. arXiv preprint arXiv:2503.09567, 2025

arXiv 2025
[6]

Unlocking the capabilities of thought: A reasoning boundary framework to quantify and optimize chain-of-thought

Qiguang Chen, Libo Qin, Jiaqi Wang, Jingxuan Zhou, and Wanxiang Che. Unlocking the capabilities of thought: A reasoning boundary framework to quantify and optimize chain-of-thought. Advances in Neural Information Processing Systems, 37:54872–54904, 2024

work page 2024

Show all 51 references

[7]

Do not think that much for 2+ 3=? on the overthinking of o1-like llms

Xingyu Chen, Jiahao Xu, Tian Liang, Zhiwei He, Jianhui Pang, Dian Yu, Linfeng Song, Qiuzhi Liu, Mengfei Zhou, Zhuosheng Zhang, et al. Do not think that much for 2+ 3=? on the overthinking of o1-like llms. arXiv preprint arXiv:2412.21187, 2024

2024 arXiv
[8]

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, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021

arXiv
[9]

Stepwise perplexity-guided refinement for efficient chain-of-thought reasoning in large language models

Yingqian Cui, Pengfei He, Jingying Zeng, Hui Liu, Xianfeng Tang, Zhenwei Dai, Yan Han, Chen Luo, Jing Huang, Zhen Li, et al. Stepwise perplexity-guided refinement for efficient chain-of-thought reasoning in large language models. arXiv preprint arXiv:2502.13260, 2025

2025 arXiv
[10]

The effect of cognitive load on economic decision making: A survey and new experiments

Cary Deck and Salar Jahedi. The effect of cognitive load on economic decision making: A survey and new experiments. European Economic Review, 78:97–119, 2015

2015
[11]

Dynamic parallel tree search for efficient llm reasoning

Yifu Ding, Wentao Jiang, Shunyu Liu, Yongcheng Jing, Jinyang Guo, Yingjie Wang, Jing Zhang, Zengmao Wang, Ziwei Liu, Bo Du, et al. Dynamic parallel tree search for efficient llm reasoning. arXiv preprint arXiv:2502.16235, 2025

2025 arXiv
[12]

Do llms know about hallucination? an empirical investigation of llm’s hidden states

Hanyu Duan, Yi Yang, and Kar Yan Tam. Do llms know about hallucination? an empirical investigation of llm’s hidden states. arXiv preprint arXiv:2402.09733, 2024

2024 arXiv
[13]

Metacognition and reasoning

Logan Fletcher and Peter Carruthers. Metacognition and reasoning. Philosophical Transactions of the Royal Society B, 367(1594):1366–1378, 2012

2012
[14]

Understanding neural dynamics: A generative modeling perspective

Hongyu Gao, Omri Barak, and John P Cunningham. Understanding neural dynamics: A generative modeling perspective. In Advances in Neural Information Processing Systems (NeurIPS) , volume 35, pages 3593–3606, 2022

2022
[15]

Spuq: Perturbation-based uncertainty quantification for large language models

Xiang Gao, Jiaxin Zhang, Lalla Mouatadid, and Kamalika Das. Spuq: Perturbation-based uncertainty quantification for large language models. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pa...

2024
[16]

Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025

2025 arXiv
[17]

Trajectory analysis: concepts and applications

W Helland-Hansen and GJ Hampson. Trajectory analysis: concepts and applications. Basin Research, 21(5):454–483, 2009

2009
[18]

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. arXiv preprint arXiv:2103.03874, 2021. 10

2021 arXiv
[19]

A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions

Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Informatio...

2025
[20]

Openai o1 system card

Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Hel- yar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card. arXiv preprint arXiv:2412.16720, 2024

2024 arXiv
[21]

Language models (mostly) know what they know

Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield-Dodds, Nova DasSarma, Eli Tran-Johnson, et al. Language models (mostly) know what they know. arXiv preprint arXiv:2207.05221, 2022

2022 arXiv
[22]

C3ot: Generating shorter chain-of-thought without compromising effectiveness

Yu Kang, Xianghui Sun, Liangyu Chen, and Wei Zou. C3ot: Generating shorter chain-of-thought without compromising effectiveness. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 24312–24320, 2025

2025
[23]

How well do llms compress their own chain-of-thought? a token complexity approach

Ayeong Lee, Ethan Che, and Tianyi Peng. How well do llms compress their own chain-of-thought? a token complexity approach. arXiv preprint arXiv:2503.01141, 2025

2025 arXiv
[24]

Inference-time intervention: Eliciting truthful answers from a language model.Advances in Neural Information Processing Systems, 36:41451–41530, 2023

Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. Inference-time intervention: Eliciting truthful answers from a language model.Advances in Neural Information Processing Systems, 36:41451–41530, 2023

2023
[25]

Reward-guided speculative decoding for efficient llm reasoning

Baohao Liao, Yuhui Xu, Hanze Dong, Junnan Li, Christof Monz, Silvio Savarese, Doyen Sahoo, and Caiming Xiong. Reward-guided speculative decoding for efficient llm reasoning. arXiv preprint arXiv:2501.19324, 2025

2025 arXiv
[26]

O1-pruner: Length-harmonizing fine-tuning for o1-like reasoning pruning

Haotian Luo, Li Shen, Haiying He, Yibo Wang, Shiwei Liu, Wei Li, Naiqiang Tan, Xiaochun Cao, and Dacheng Tao. O1-pruner: Length-harmonizing fine-tuning for o1-like reasoning pruning. arXiv preprint arXiv:2501.12570, 2025

2025 arXiv
[27]

Reasoning models can be effective without thinking

Wenjie Ma, Jingxuan He, Charlie Snell, Tyler Griggs, Sewon Min, and Matei Zaharia. Reasoning models can be effective without thinking. arXiv preprint arXiv:2504.09858, 2025

2025 arXiv
[28]

Cot-valve: Length- compressible chain-of-thought tuning

Xinyin Ma, Guangnian Wan, Runpeng Yu, Gongfan Fang, and Xinchao Wang. Cot-valve: Length- compressible chain-of-thought tuning. arXiv preprint arXiv:2502.09601, 2025

2025 arXiv
[29]

s1: Simple test-time scaling

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

2025 arXiv
[30]

Self-training elicits concise reasoning in large language models

Tergel Munkhbat, Namgyu Ho, Seo Hyun Kim, Yongjin Yang, Yujin Kim, and Se-Young Yun. Self-training elicits concise reasoning in large language models. arXiv preprint arXiv:2502.20122, 2025

2025 arXiv
[31]

Karl Pearson. Vii. note on regression and inheritance in the case of two parents. proceedings of the royal society of London, 58(347-352):240–242, 1895
[32]

Trajectory analysis via a geometric feature space approach

Mark D Rintoul and Andrew T Wilson. Trajectory analysis via a geometric feature space approach. Statistical Analysis and Data Mining: The ASA Data Science Journal, 8(5-6):287–301, 2015

2015
[33]

Dast: Difficulty-adaptive slow-thinking for large reasoning models

Yi Shen, Jian Zhang, Jieyun Huang, Shuming Shi, Wenjing Zhang, Jiangze Yan, Ning Wang, Kai Wang, and Shiguo Lian. Dast: Difficulty-adaptive slow-thinking for large reasoning models. arXiv preprint arXiv:2503.04472, 2025

2025
[34]

Reflexion: Language agents with verbal reinforcement learning

Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems, 36:8634–8652, 2023

2023
[35]

Herbert A. Simon. Rational choice and the structure of the environment. Psychological Review, 63(2):129– 138, 1956

1956
[37]

Stop overthinking: A survey on efficient reasoning for large language models

Yang Sui, Yu-Neng Chuang, Guanchu Wang, Jiamu Zhang, Tianyi Zhang, Jiayi Yuan, Hongyi Liu, Andrew Wen, Shaochen Zhong, Hanjie Chen, et al. Stop overthinking: A survey on efficient reasoning for large language models. arXiv preprint arXiv:2503.16419, 2025. 11

2025 arXiv
[38]

A survey of reasoning with foundation models

Jiankai Sun, Chuanyang Zheng, Enze Xie, Zhengying Liu, Ruihang Chu, Jianing Qiu, Jiaqi Xu, Mingyu Ding, Hongyang Li, Mengzhe Geng, et al. A survey of reasoning with foundation models. arXiv preprint arXiv:2312.11562, 2023

2023 arXiv
[39]

Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, Cheng Chen, Cheng Li, Chenjun Xiao, Chenzhuang Du, Chonghua Liao, et al. Kimi k1. 5: Scaling reinforcement learning with llms. arXiv preprint arXiv:2501.12599, 2025

2025 arXiv
[40]

Just ask for calibration: Strategies for eliciting calibrated confidence scores from language models fine-tuned with human feedback

Katherine Tian, Eric Mitchell, Allan Zhou, Archit Sharma, Rafael Rafailov, Huaxiu Yao, Chelsea Finn, and Christopher D Manning. Just ask for calibration: Strategies for eliciting calibrated confidence scores from language models fine-tuned with human feedback. In Proceedings o...

2023
[41]

Harnessing the reasoning economy: A survey of efficient reasoning for large language models

Rui Wang, Hongru Wang, Boyang Xue, Jianhui Pang, Shudong Liu, Yi Chen, Jiahao Qiu, Derek Fai Wong, Heng Ji, and Kam-Fai Wong. Harnessing the reasoning economy: A survey of efficient reasoning for large language models. arXiv preprint arXiv:2503.24377, 2025

2025 arXiv
[42]

Self-consistency improves chain of thought reasoning in language models

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

2022 arXiv
[44]

Chain-of-thought prompting elicits reasoning in large language models

Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837, 2022

2022
[45]

When more is less: Understanding chain-of-thought length in llms

Yuyang Wu, Yifei Wang, Tianqi Du, Stefanie Jegelka, and Yisen Wang. When more is less: Understanding chain-of-thought length in llms. arXiv preprint arXiv:2502.07266, 2025

2025 arXiv
[46]

Tokenskip: Controllable chain-of-thought compression in llms

Heming Xia, Yongqi Li, Chak Tou Leong, Wenjie Wang, and Wenjie Li. Tokenskip: Controllable chain-of-thought compression in llms. arXiv preprint arXiv:2502.12067, 2025

2025
[47]

Chain of draft: Thinking faster by writing less

Silei Xu, Wenhao Xie, Lingxiao Zhao, and Pengcheng He. Chain of draft: Thinking faster by writing less. arXiv preprint arXiv:2502.18600, 2025

2025 arXiv
[48]

Dynamic early exit in reasoning models

Chenxu Yang, Qingyi Si, Yongjie Duan, Zheliang Zhu, Chenyu Zhu, Zheng Lin, Li Cao, and Weiping Wang. Dynamic early exit in reasoning models. arXiv preprint arXiv:2504.15895, 2025

2025
[49]

Natural language reasoning, a survey.ACM Computing Surveys, 56(12):1–39, 2024

Fei Yu, Hongbo Zhang, Prayag Tiwari, and Benyou Wang. Natural language reasoning, a survey.ACM Computing Surveys, 56(12):1–39, 2024

2024
[50]

Reasoning models know when they’re right: Probing hidden states for self-verification

Anqi Zhang, Yulin Chen, Jane Pan, Chen Zhao, Aurojit Panda, Jinyang Li, and He He. Reasoning models know when they’re right: Probing hidden states for self-verification. arXiv preprint arXiv:2504.05419, 2025

2025 arXiv
[51]

<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n Please reason step by step, and put your final answer within \\boxed{}.\n

Tianyi Zhang, Tao Yu, Tatsunori Hashimoto, Mike Lewis, Wen-tau Yih, Daniel Fried, and Sida Wang. Coder reviewer reranking for code generation. In International Conference on Machine Learning, pages 41832–41846. PMLR, 2023. 12 Appendix 6.1 Limitations While CoRE-Eval offers an ...

2023

Pith tools

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

[1] [36]

Unsuper- vised real-time hallucination detection based on the internal states of large language models

Weihang Su, Changyue Wang, Qingyao Ai, Yiran Hu, Zhijing Wu, Yujia Zhou, and Yiqun Liu. Unsuper- vised real-time hallucination detection based on the internal states of large language models. In Findings of the 62nd Annual Meeting of the Association for Computational Linguistics, ACL 2024, pages 14379–14391. Association for Computational Linguistics (ACL), 2024

work page 2024

[2] [43]

Latent space chain-of-embedding enables output-free llm self-evaluation

Yiming Wang, Pei Zhang, Baosong Yang, Derek F Wong, and Rui Wang. Latent space chain-of-embedding enables output-free llm self-evaluation. arXiv preprint arXiv:2410.13640, 2024

arXiv 2024

[3] [1]

L1: Controlling how long a reasoning model thinks with reinforcement learning

Pranjal Aggarwal and Sean Welleck. L1: Controlling how long a reasoning model thinks with reinforcement learning. arXiv preprint arXiv:2503.04697, 2025

arXiv 2025

[4] [2]

Sketch-of-thought: Efficient llm reasoning with adaptive cognitive-inspired sketching

Simon A Aytes, Jinheon Baek, and Sung Ju Hwang. Sketch-of-thought: Efficient llm reasoning with adaptive cognitive-inspired sketching. arXiv preprint arXiv:2503.05179, 2025

arXiv 2025

[5] [3]

Discovering latent knowledge in language models without supervision

Collin Burns, Haotian Ye, Dan Klein, and Jacob Steinhardt. Discovering latent knowledge in language models without supervision. In The Eleventh International Conference on Learning Representations, 2023

2023

[6] [4]

The dynamics of cortical activity during sensorimotor behavior

Rishidev Chaudhuri, Boris Ger¸ sgorin, Zhixin Wang, Mark M Churchland, and Surya Ganguli. The dynamics of cortical activity during sensorimotor behavior. Neuron, 102(4):768–781, 2019

work page 2019

[7] [5]

Towards reasoning era: A survey of long chain-of-thought for reasoning large language models

Qiguang Chen, Libo Qin, Jinhao Liu, Dengyun Peng, Jiannan Guan, Peng Wang, Mengkang Hu, Yuhang Zhou, Te Gao, and Wanxiang Che. Towards reasoning era: A survey of long chain-of-thought for reasoning large language models. arXiv preprint arXiv:2503.09567, 2025

arXiv 2025

[8] [6]

Unlocking the capabilities of thought: A reasoning boundary framework to quantify and optimize chain-of-thought

Qiguang Chen, Libo Qin, Jiaqi Wang, Jingxuan Zhou, and Wanxiang Che. Unlocking the capabilities of thought: A reasoning boundary framework to quantify and optimize chain-of-thought. Advances in Neural Information Processing Systems, 37:54872–54904, 2024

work page 2024