Pith. sign in

REVIEW 4 major objections 5 minor 54 references

The paper claims that a reasoning model's answer is correct exactly when its uncertainty about the final answer and the 'momentum' of its latent thought trajectory collapse at the same moment, and it builds a training-free gate that truncat

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 18:44 UTC pith:5D3MM5GK

load-bearing objection A genuinely new combination of entropy and latent-geometry signals for truncating reasoning, with solid empirical promise, but the headline efficiency claim omits the proxy verifier's compute and one key hyperparameter is never set. the 4 major comments →

arxiv 2607.17188 v1 pith:5D3MM5GK submitted 2026-07-19 cs.AI

Is Your Model Thinking or Just Stagnating? PUMA: Diagnosing Reasoning Pathology via Phase-Momentum Alignment

Cheng Yan , Guangyang Ye , Wuyang Zhang , Fan Xu , Zhijun Fan , Xiang Xia , Yanyong Zhang This is my paper
classification cs.AI
keywords Large Reasoning ModelsTest-time ScalingEfficient ReasoningLatent Representation AnalysisPhase-Momentum AlignmentCognitive-Energy ModelOverthinkingEntropy-based FSM
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper sets out to prove that the correctness of a chain-of-thought trace is readable online from the synchronization of two internal signals: an entropy-based phase (how uncertain the model is about the final answer) and a geometric momentum (how much the latent representation is actually moving). It formalizes the Phase-Momentum Alignment Hypothesis, builds the Cognitive-Energy Model on two orthogonal metrics, and operationalizes it in PUMA, a training-free framework that monitors entropy with a lightweight finite-state machine and computes the expensive geometric quantity only when a phase transition is suspected. The payoff is a plug-in intervention that truncates generation exactly when the model has reached the answer and, separately, detects 'stagnation' — high uncertainty with dead momentum — so that hallucination can be cut early. A sympathetic reader would care because it targets the overthinking paradox without retraining or extra generation passes, and the paper reports consistent efficiency gains and accuracy improvements across models from 1.5B to 32B on math, science, and general reasoning benchmarks.

Core claim

The central claim is the Phase-Momentum Alignment Hypothesis: a reasoning trajectory is correct precisely when the drop in entropic uncertainty coincides with the collapse of Net Reasoning Momentum (semantic velocity plus intensity curvature). When this synchronization occurs, the model has reached a fixed point in its semantic manifold and the chain-of-thought can be truncated; when momentum dies while uncertainty stays high, the model has entered passive stagnation, and continuing to generate only produces hallucinated text. PUMA realizes this by coupling an entropy-based finite state machine (tracking Divergence, Convergence, and Undetermined phases with a sliding window) with an event-tr

What carries the argument

The load-bearing object is the pair (U_t, Φ_t): the Proxy-based Lookahead Entropy U_t (Eq. 9), computed by a small 1.5B proxy verifier that assigns probabilities to candidate answers given the current context, and the Net Reasoning Momentum Φ_t (Eq. 8), defined as the semantic velocity G_t between consecutive reasoning steps plus λ times the second difference (curvature) of the Global Thinking Intensity I_t — a tortuosity ratio of cumulative layer-wise transformation to net displacement, measured with angular distance. The phase (entropy) tells the system whether the model is still exploring or has converged; the momentum tells it whether the latent trajectory is still doing constructive wor

Load-bearing premise

Everything rests on the lookahead entropy from the 1.5B proxy verifier being a faithful real-time monitor of the target model's epistemic uncertainty: if the proxy's candidate set is unrepresentative or the proxy is confidently wrong, the entropy-based phase detection misclassifies states, the geometric gate never triggers properly, and PUMA either truncates prematurely or never intervenes.

What would settle it

Construct or collect reasoning traces where the final answer is wrong but, at some step, both the entropy-based phase has dropped (Eq. 11) and the momentum has collapsed (Φ_t < ε); if PUMA truncates and accepts a nontrivial fraction of these wrong answers, the Phase-Momentum Alignment Hypothesis is falsified. A simpler quantitative test: measure the precision of the PUMA gate at the moment it fires across a diverse set of traces — if the gate's precision is not high (e.g., below 90%), the alignment claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the hypothesis holds, test-time compute can be cut by roughly 20-50% on math, science, and general reasoning benchmarks, with no accuracy loss and often a gain, because the gate prevents both premature truncation and overthinking.
  • The dual-signal gate eliminates 'deceptive convergence': low entropy alone no longer triggers acceptance, so confident hallucinations are filtered out by the momentum check.
  • Stagnation detection (TRUNCATEfail) can be paired with corrective interventions such as retrieval-augmented restart, which on a geometry case from a math benchmark turns a wrong answer into the correct one.
  • The framework is training-free and model-agnostic, applying to any LRM that exposes hidden states and has access to a small proxy, and it generalizes across scales from 1.5B to 32B and across domains.

Where Pith is reading between the lines

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

  • The hypothesis can be verified directly as a calibration claim: the set of traces that simultaneously satisfy 'entropy below α·U_max' and 'Φ_t below ε' should have near-100% correctness; any systematic wrong subpopulation that passes the gate would invalidate the threshold pair (α, ε) as scale-agnostic.
  • The proxy-based entropy is a channel whose own confidence matters: the paper's finding that a 1.5B proxy outperforms larger proxies suggests the entropy signal is best read as a relative, context-anchored signal, and that choosing the proxy per query could further sharpen the gate.
  • The Cognitive-Energy Model suggests a diagnostic for silent failure in long-horizon planning: vanishing momentum with high entropy means the model is no longer exploring; one could use the same Φ_t signal to trigger memory retrieval or a reasoning restart, not just truncation.
  • A physical analogy suggests a refined trigger: reasoning behaves like a damped harmonic system, so the ideal intervention is the first zero-crossing of momentum after entropy collapse; a Kalman-filtered estimate of Φ_t might make the gate more robust than hard thresholds.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes PUMA, a training-free framework for diagnosing and intervening in the chain-of-thought reasoning of large reasoning models. It posits a Phase-Momentum Alignment Hypothesis, under which correct reasoning is characterized by temporal synchronization between an uncertainty signal (entropic phase) and a geometric latent-trajectory signal (cognitive momentum). The Cognitive-Energy Model defines two metrics: Geometric Cognitive Effort, via Global Thinking Intensity (tortuosity) and Local Semantic Gain (velocity), and Entropic Cognitive Uncertainty, via a proxy-model lookahead entropy. These are combined into a Net Reasoning Momentum Φ_t (Eq. 8) and used to classify reasoning states (Eq. 12). PUMA couples a lightweight entropy-based finite-state machine with event-triggered geometric analysis, then applies adaptive truncation or a circuit-breaking restart. Experiments on DeepSeek-R1-Distill 1.5B/7B/14B and QwQ-32B across five benchmarks report accuracy improvements and target-token savings relative to vanilla decoding and several baselines, with ablations and sensitivity analyses.

Significance. If the efficiency claim holds end-to-end, PUMA would be a practically useful, training-free method for reducing overthinking in LRMs, and the dual-signal design is a reasonable response to the known failure of entropy-only truncation. The paper has notable strengths: it evaluates across four model scales and five benchmarks, includes ablation variants isolating the entropy and latent signals, provides sensitivity analyses for key thresholds, and includes a detailed case study. These are the right kinds of empirical checks for an inference-time intervention method. However, the theoretical framing is mostly definitional rather than derivational, and the central efficiency claim is currently stated only in terms of target-model tokens, omitting the proxy verifier's inference cost. The contribution is therefore promising but not yet fully established.

major comments (4)
  1. [Section 5.1, Eq. (9)] The reported token savings are not end-to-end. U_t is defined by generating a candidate set Y from the proxy M_phi at the current context C_t, and the FSM in Section 5.1 needs U_t at every monitored step (or sliding window) to detect phase transitions. Tables 1 and 2 count only target-model tokens; the paper does not report k, candidate length, the number of U_t evaluations, or any proxy inference cost. For QwQ-32B on MATH-500, PUMA saves ~27% of target tokens, but if each U_t evaluation costs a few hundred proxy-generated tokens and is repeated over hundreds of steps, the proxy overhead can exceed the savings. Please report total compute (target + proxy) or a bound on the overhead; otherwise the central accuracy-efficiency claim is not supported.
  2. [Section 4.3, Eq. (12)] The central 'discovery' is partly built into the definitions. SHealthy is defined as PConv ∧ Φ_t < ε, and SStagnant as PDiv ∧ Φ_t < ε. Thus the statement that healthy reasoning corresponds to synchronized low entropy and low momentum is true by construction, not an empirical finding. This does not invalidate the intervention results, but it means the paper overstates what the experiments establish. Please either reframe the Cognitive-Energy Model as an engineering heuristic or test the alignment hypothesis independently—for example, by comparing the distribution of alignment events on correct vs. incorrect trajectories before any intervention is applied.
  3. [Section 4.2, Eq. (9); Section 6.4] The proxy entropy channel is asserted but not validated. U_t is generated by a 1.5B proxy model; the paper provides no direct evidence that this lookahead entropy tracks the target model's epistemic uncertainty or that the candidate set Y is representative. The proxy-scale result in Section 6.4 (larger proxies worsen Accuracy Loss) further suggests the entropy signal is sensitive in ways the theory does not explain. Please provide a calibration analysis—e.g., correlation of U_t with target-model semantic entropy or with final outcome correctness on a held-out set—and report k, candidate sampling strategy, and candidate length. Without this, the risk of premature or missed truncation due to an unfaithful entropy signal is not quantified.
  4. [Table 2, AIME 2025 row] There is an apparent internal inconsistency: PUMA entropy and full PUMA report exactly the same token cost (7581.37) on AIME 2025 with the 1.5B model, yet accuracy differs substantially (13.33% vs. 23.33%). Under the described mechanism, the geometric check should prevent PUMA from truncating at the same point as the entropy-only variant when entropy is deceptively low, and any corrective restart would change token consumption. Please explain how these identical costs arise, or correct the table. This issue casts doubt on the reliability of the reported numbers.
minor comments (5)
  1. [Section 1] In the contributions list, 'formal interpretation interpretation' is duplicated.
  2. [Eq. (8)] The formula for Net Reasoning Momentum contains a stray double vertical bar in the intensity-fluctuation term; please clean up the notation.
  3. [Section 6.4] The proxy-model-scale analysis is described as illustrated in Figure 3(a), but Figure 3(a) is the Reasoning Quadrant Distribution; the reference should likely point to a different panel or a new figure.
  4. [Tables 1 and 2] No error bars or statistical significance tests are reported despite 'averaged over three independent runs'. Please include variance or confidence intervals, especially for the accuracy-difference claims.
  5. [Eq. (8) and Section 6.1] The hyperparameter λ in the momentum definition is never assigned a value and is not included in the sensitivity analysis. Please state its value and, if possible, its sensitivity, since it directly controls the contribution of the intensity-curvature term.

Circularity Check

0 steps flagged

No significant circularity: PUMA's diagnostic states are operationalized from the stated hypothesis and validated against external accuracy labels, with no load-bearing self-citations or fitted predictions.

full rationale

The paper's central move is to define the Phase-Momentum Alignment Hypothesis, operationalize it via Eqs. (5)-(12), and then test the resulting intervention policy on standard benchmarks against ground-truth labels. The diagnostic state SHealthy is indeed defined as low entropic uncertainty plus low momentum (Eq. 12), so the framework's behavior encodes the hypothesis; but the correctness of that encoding is not assumed. Accuracy is measured independently from the definitions, and the paper reports both gains and losses across ablations (e.g., PUMA_entropy and PUMA_latent) and models, which would be impossible if the result were true by construction. The thresholds alpha, epsilon and window w are stated hyperparameters with sensitivity analyses, not parameters fitted to the target labels. There are no self-citations used as load-bearing evidence, no imported uniqueness theorem, and no ansatz smuggled in via citation. The critic's point that proxy-verifier compute is omitted from token savings is a validity/efficiency-accounting concern, not a circularity of the claimed derivation. Therefore the derivation is self-contained with respect to circularity.

Axiom & Free-Parameter Ledger

8 free parameters · 4 axioms · 0 invented entities

The framework depends on a handful of design choices and assumptions: proxy reliability, the layer-depth assumption, and the posited alignment hypothesis. The free parameters are mostly fixed by hand and only partially swept; λ is never given.

free parameters (8)
  • λ (Intensity Curvature weight)
    Appears in Eq. 8 defining Net Reasoning Momentum Φ_t = G_t + λ·|Δ²I_t|; no value is given in the implementation details, leaving the metric underspecified.
  • α (phase sensitivity) = 0.05
    Threshold in Eq. 11 for convergence/divergence detection; chosen in §6.1, swept in §6.4.
  • ε (momentum relaxation threshold) = 0.1
    Threshold in Eq. 12 for low momentum; chosen in §6.1, swept in §6.4.
  • w (sliding window) = 3
    Window size for phase stability in Eq. 11; fixed in §6.1.
  • m (final layer fraction) = 25% of layers
    Number of deepest layers used for geometric metrics; set as 'final 25% of layers' in §6.1.
  • k (initial uncertainty window) = k≤5
    Length of initial window for U_init in §4.2; not precisely specified.
  • Step granularity = 3 newline separators
    Discretization of reasoning steps in §6.1; affects all time-series metrics but is not ablated.
  • Proxy model scale = 1.5B
    Default proxy M_φ; the paper argues smaller proxies give sharper signals, but the overhead is not counted.
axioms (4)
  • domain assumption The deepest transformer layers encode the most abstract semantics relevant to reasoning.
    Used to justify computing geometric metrics only on the final 25% of layers (§4.1).
  • domain assumption The 1.5B proxy model's conditional entropy P_φ(y|C_t) is a reliable surrogate for the target model's epistemic uncertainty.
    Lookahead entropy in Eq. 9 assumes proxy rankings align with target correctness; supported only indirectly by ablation and sensitivity analysis.
  • domain assumption Reasoning steps can be reliably segmented by every three newline separators.
    Default in §6.1; all time-series metrics are sensitive to the discretization choice.
  • ad hoc to paper The Phase-Momentum Alignment Hypothesis (correctness correlates with temporal synchronization of entropy and momentum) holds across models and domains.
    Posited in §1 and operationalized in Eq. 12; no derivation is provided, only empirical validation.

pith-pipeline@v1.3.0-alltime-deepseek · 18265 in / 12765 out tokens · 105093 ms · 2026-08-01T18:44:38.998630+00:00 · methodology

0 comments
read the original abstract

Test-time scaling empowers Large Reasoning Models (LRMs) to tackle complex tasks via extensive Chain-of-Thought (CoT). However, this often induces the "overthinking" paradox, where redundant reasoning increases computational overhead without guaranteeing accuracy. Existing test-time efficiency optimization methods primarily fall into two categories: information-theoretic approaches, which are prone to "deceptive convergence" where low uncertainty masks hallucinations, and latent representation analyses, which are often post-hoc, lacking the real-time sensitivity for dynamic reasoning. To bridge this gap, we first posit the Phase-Momentum Alignment Hypothesis, asserting that reasoning correctness hinges on the temporal synchronization between geometric momentum and uncertainty resolution. We then theoretically formulate the Cognitive-Energy Model to characterize these dynamics through two orthogonal dimensions: Geometric Cognitive Effort, quantified by latent velocity and tortuosity, and Entropic Cognitive Uncertainty. To operationalize this, we introduce PUMA (Phase-Uncertainty Momentum Alignment), a training-free framework employing a tiered diagnostic architecture. By coupling lightweight phase monitoring with event-triggered geometric analysis, PUMA effectively distinguishes active exploration from passive stagnation, enabling precise interventions through adaptive truncation or corrective measures. Extensive experiments on LRMs spanning 1.5B to 32B demonstrate that PUMA consistently outperforms state-of-the-art baselines across diverse benchmarks, achieving a superior accuracy-efficiency trade-off and robust cross-domain generalization.

Figures

Figures reproduced from arXiv: 2607.17188 by Cheng Yan, Fan Xu, Guangyang Ye, Wuyang Zhang, Xiang Xia, Yanyong Zhang, Zhijun Fan.

Figure 1
Figure 1. Figure 1: Motivation: Active Thinking vs. Stagnation. Visu [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the PUMA Framework. (1) Real-time Monitoring via an entropy-based FSM to track phase transitions; [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comprehensive Performance Analysis of PUMA. (a) Reasoning Quadrant Distribution of QwQ-32B on MATH-500, [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Hyperparameter Robustness Analysis. Performance metrics (averaged across five benchmarks) for 1.5B, 7B, and 32B [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison on a MATH500 geometry task. QwQ-32B generates a lengthy but flawed reasoning chain based on [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗

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

54 extracted references · 29 linked inside Pith

  1. [1]

    Pranjal Aggarwal and Sean Welleck. 2025. L1: Controlling how long a reason- ing model thinks with reinforcement learning.arXiv preprint arXiv:2503.04697 (2025)

    Pith/arXiv arXiv 2025

  2. [2]

    Riccardo Ali, Francesco Caso, Christopher Irwin, and Pietro Liò. 2025. Entropy- lens: The information signature of transformer computations.arXiv preprint arXiv:2502.16570(2025)

    arXiv 2025

  3. [3]

    Daman Arora and Andrea Zanette. 2025. Training language models to reason efficiently.arXiv preprint arXiv:2502.04463(2025)

    arXiv 2025

  4. [4]

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

    arXiv 2025

  5. [5]

    Qiguang Chen, Libo Qin, Jiaqi Wang, Jingxuan Zhou, and Wanxiang Che. 2024. Unlocking the capabilities of thought: A reasoning boundary framework to quan- tify and optimize chain-of-thought.Advances in Neural Information Processing Systems37 (2024), 54872–54904

    2024

  6. [6]

    Xin Chen, Hanxian Huang, Yanjun Gao, Yi Wang, Jishen Zhao, and Ke Ding

  7. [7]

    Zhuo Chen, Zhuotao Jin, Di Luo, Marin Soljačić, et al . 2025. L 2M: Mutual Information Scaling Law for Long-Context Language Modeling.arXiv preprint arXiv:2503.04725(2025)

    arXiv 2025

  8. [8]

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. 2021. Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168(2021)

    Pith/arXiv arXiv 2021

  9. [9]

    1999.Elements of information theory

    Thomas M Cover. 1999.Elements of information theory. John Wiley & Sons

    1999

  10. [10]

    Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Yarin Gal. 2024. Detecting hallucinations in large language models using semantic entropy.Nature630, 8017 (2024), 625–630

    2024

  11. [11]

    Sicheng Feng, Gongfan Fang, Xinyin Ma, and Xinchao Wang. 2025. Efficient reasoning models: A survey.arXiv preprint arXiv:2504.10903(2025)

    arXiv 2025

  12. [12]

    Zeyu Gan, Yun Liao, and Yong Liu. 2025. Rethinking external slow-thinking: From snowball errors to probability of correct reasoning.arXiv preprint arXiv:2501.15602(2025)

    Pith/arXiv arXiv 2025

  13. [13]

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

    Pith/arXiv arXiv 2025

  14. [14]

    Tingxu Han, Zhenting Wang, Chunrong Fang, Shiyu Zhao, Shiqing Ma, and Zhenyu Chen. 2025. Token-budget-aware llm reasoning. InFindings of the Association for Computational Linguistics: ACL 2025. 24842–24855

    2025

  15. [15]

    Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. 2021. Measuring mathematical problem solving with the math dataset.arXiv preprint arXiv:2103.03874(2021)

    Pith/arXiv arXiv 2021

  16. [16]

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

    Pith/arXiv arXiv 2024

  17. [17]

    2025.Information-theoretic foundations for machine learning

    Hong Jun Jeon. 2025.Information-theoretic foundations for machine learning. Ph. D. Dissertation. Stanford University

    2025

  18. [18]

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

    2025

  19. [19]

    Jannik Kossen, Jiatong Han, Muhammed Razzak, Lisa Schut, Shreshth Malik, and Yarin Gal. 2024. Semantic entropy probes: Robust and cheap hallucination detection in llms.arXiv preprint arXiv:2406.15927(2024)

    Pith/arXiv arXiv 2024

  20. [20]

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

    Pith/arXiv arXiv 2025

  21. [21]

    Haoxi Li, Sikai Bai, Jie Zhang, and Song Guo. 2025. CoRE: Enhancing Metacog- nition with Label-free Self-evaluation in LRMs.arXiv preprint arXiv:2507.06087 (2025)

    Pith/arXiv arXiv 2025

  22. [22]

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

    Pith/arXiv arXiv 2025

  23. [23]

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

    Pith/arXiv arXiv 2025

  24. [24]

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

    Pith/arXiv arXiv 2025

  25. [25]

    Xinyin Ma, Guangnian Wan, Runpeng Yu, Gongfan Fang, and Xinchao Wang

  26. [26]

    Xiaoye Qu, Yafu Li, Zhaochen Su, Weigao Sun, Jianhao Yan, Dongrui Liu, Ganqu Cui, Daizong Liu, Shuxian Liang, Junxian He, et al. 2025. A survey of efficient reasoning for large reasoning models: Language, multimodality, and beyond. arXiv preprint arXiv:2503.21614(2025)

    arXiv 2025

  27. [27]

    David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R Bowman. 2024. Gpqa: A graduate-level google-proof q&a benchmark. InFirst Conference on Language Modeling

    2024

  28. [28]

    Ruifeng Ren and Yong Liu. 2025. Revisiting transformers through the lens of low entropy and dynamic sparsity.arXiv preprint arXiv:2504.18929(2025)

    Pith/arXiv arXiv 2025

  29. [29]

    Xuan Shen, Yizhou Wang, Xiangxi Shi, Yanzhi Wang, Pu Zhao, and Jiuxiang Gu

  30. [30]

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

    arXiv 2025

  31. [31]

    Zhenyi Shen, Hanqi Yan, Linhai Zhang, Zhanghao Hu, Yali Du, and Yulan He. 2025. Codi: Compressing chain-of-thought into continuous space via self- distillation.arXiv preprint arXiv:2502.21074(2025)

    Pith/arXiv arXiv 2025

  32. [32]

    Efficient reasoning with hidden thinking.arXiv preprint arXiv:2501.19201 (2025)

    Pith/arXiv arXiv 2025

  33. [33]

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

    Pith/arXiv arXiv 2025

  34. [34]

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

    Pith/arXiv arXiv 2025

  35. [35]

    DiJia Su, Hanlin Zhu, Yingchen Xu, Jiantao Jiao, Yuandong Tian, and Qinqing Zheng. 2025. Token assorted: Mixing latent and text tokens for improved lan- guage model reasoning.arXiv preprint arXiv:2502.03275(2025)

    Pith/arXiv arXiv 2025

  36. [36]

    Jean-Francois Ton, Muhammad Faaiz Taufiq, and Yang Liu. 2024. Under- standing chain-of-thought in llms through information theory.arXiv preprint arXiv:2411.11984(2024)

    Pith/arXiv arXiv 2024

  37. [37]

    Alexander Matt Turner, Lisa Thiergart, Gavin Leech, David Udell, Juan J Vazquez, Ulisse Mini, and Monte MacDiarmid. 2023. Steering language models with activation engineering.arXiv preprint arXiv:2308.10248(2023)

    Pith/arXiv arXiv 2023

  38. [38]

    Qwen Team. 2025. Qwq-32b: Embracing the power of reinforcement learning

    2025

  39. [39]

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

    Pith/arXiv arXiv 2025

  40. [40]

    Yubo Wang, Xueguang Ma, Ge Zhang, Yuansheng Ni, Abhranil Chandra, Shiguang Guo, Weiming Ren, Aaran Arulraj, Xuan He, Ziyan Jiang, et al. 2024. Mmlu-pro: A more robust and challenging multi-task language understand- ing benchmark.Advances in Neural Information Processing Systems37 (2024), 95266–95290

    2024

  41. [41]

    Martina G Vilas, Safoora Yousefi, Besmira Nushi, Eric Horvitz, and Vidhisha Balachandran. 2025. Tracing the Traces: Latent Temporal Signals for Efficient and Accurate Reasoning.arXiv preprint arXiv:2510.10494(2025)

    arXiv 2025

  42. [42]

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

    2022

  43. [43]

    Yuyang Wu, Yifei Wang, Ziyu Ye, Tianqi Du, Stefanie Jegelka, and Yisen Wang

  44. [44]

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

    Pith/arXiv arXiv 2024

  45. [45]

    Yige Xu, Xu Guo, Zhiwei Zeng, and Chunyan Miao. 2025. Softcot: Soft chain- of-thought for efficient reasoning with llms.arXiv preprint arXiv:2502.12134 (2025)

    Pith/arXiv arXiv 2025

  46. [46]

    Shiming Yang, Yuxuan Tong, Xinyao Niu, Graham Neubig, and Xiang Yue. [n. d.]. Demystifying Long Chain-of-Thought Reasoning. InForty-second International Conference on Machine Learning

  47. [47]

    When more is less: Understanding chain-of-thought length in llms.arXiv preprint arXiv:2502.07266(2025)

    Pith/arXiv arXiv 2025

  48. [48]

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

    arXiv 2025

  49. [49]

    stagnation

    Jiajie Zhang, Nianyi Lin, Lei Hou, Ling Feng, and Juanzi Li. 2025. Adaptthink: Reasoning models can learn when to think.arXiv preprint arXiv:2505.13417 (2025). Is Your Model Thinking or Just Stagnating? PUMA: Diagnosing Reasoning Pathology via Phase-Momentum Alignment A Extended Experimental Main Results A.1 Scalability on Lightweight Models Superior Pare...

    Pith/arXiv arXiv 2025

  50. [51]

    Xixian Yong, Xiao Zhou, Yingying Zhang, Jinlin Li, Yefeng Zheng, and Xian Wu

  51. [52]

    Think or Not? Exploring Thinking Efficiency in Large Reasoning Models via an Information-Theoretic Lens.arXiv preprint arXiv:2505.18237(2025)

    Pith/arXiv arXiv 2025

  52. [53]

    Mert Yuksekgonul, Varun Chandrasekaran, Erik Jones, Suriya Gunasekar, Ranjita Naik, Hamid Palangi, Ece Kamar, and Besmira Nushi. 2023. Attention satisfies: A constraint-satisfaction lens on factual errors of language models.arXiv preprint arXiv:2309.15098(2023)

    Pith/arXiv arXiv 2023

  53. [2024]

    arXiv preprint arXiv:2403.03348(2024)

    Learning to maximize mutual information for chain-of-thought distillation. arXiv preprint arXiv:2403.03348(2024)

    Pith/arXiv arXiv 2024

  54. [2025]

    Cot-valve: Length-compressible chain-of-thought tuning.arXiv preprint arXiv:2502.09601(2025)

    Pith/arXiv arXiv 2025