Pith. sign in

REVIEW 4 major objections 5 minor 19 references

Entropy-based pruning of chain-of-thought traces gives no advantage over random selection; apparent token-level gains on math come from numeric tokens, and activation patching shows reasoning content is distributed across the trace.

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

2026-08-03 00:35 UTC pith:T77B2G6R

load-bearing objection Useful negative result about entropy pruning, but the authors need to calibrate their entropy proxy and soften their claims before I'd trust the headline. the 4 major comments →

arxiv 2607.28707 v1 pith:T77B2G6R submitted 2026-07-30 cs.CL

Demystifying Entropy-based Selection for Chain-of-Thought Compression in Large Reasoning Models

Sara Candussio , Daniel Scalena , Luca Bortolussi , Elisabetta Fersini , Malvina Nissim , Gabriele Sarti This is my paper
classification cs.CL
keywords chain-of-thought compressionentropy-based pruningtoken entropysentence entropyactivation patchingrandom baselinereasoning tracenumeric token confound
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 tests whether token- and sentence-level entropy reliably identify the parts of a reasoning trace that matter for the answer. Across six models and math, logic, and science benchmarks, entropy-based pruning never beats randomly keeping the same fraction of the trace at sentence level. At token level, low-entropy selection appears to help on math, but the paper shows this is a side effect of numeric tokens having low entropy while also carrying content. When numbers are excluded, the advantage vanishes. Patching activations of a handful of content tokens restores most performance, suggesting reasoning information is spread across the trace rather than concentrated in low- or high-entropy spots.

Core claim

The paper's central claim is that the predictability (entropy) of a token or sentence in a reasoning trace is not a reliable guide to whether that piece of text is needed for the final answer. In a sweep of six models across math, logic, and science benchmarks, sentence-level pruning by either low or high entropy never reliably outperformed deleting the same number of sentences at random. At token level, low-entropy selection appeared to help on math, but the paper shows the gain is inherited from numeric tokens, which are both low-entropy and content-bearing; a low-entropy selection that excludes digits loses the advantage. Finally, restoring the original hidden activations of a few retaine

What carries the argument

The central objects are the selection criteria themselves: random, low-entropy, high-entropy, numbers, low-entropy-no-numbers, plus pattern-based (newline/end-of-sentence) selectors, compared through relative performance retention (RPR) and its AUC difference from random. The decisive machinery is the paired comparison of 'numbers' versus 'low-entropy no numbers' to strip the entropy effect from the digit confound, and a causal activation-patching intervention that reuses cached full-trace hidden states at retained positions. Together they separate surface-text selection from the hidden-state context that actually carries the reasoning.

Load-bearing premise

The load-bearing premise is that token entropy computed from only the top-20 next-token probabilities faithfully ranks the uncertainty of each position; if tail probabilities are large and uneven across token classes, the entropy ordering—and the null result—could be an artifact of that truncated estimate.

What would settle it

Take a single math benchmark and model, recompute token entropies over the full output distribution (or with k much larger than 20), then rerun the low-entropy, numbers, and low-entropy-no-numbers selectors with and without patching. If low-entropy-no-numbers ever beats random under full-vocabulary entropy, the paper's claim that the math gains are purely a numeric-token confound would be falsified.

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

If this is right

  • Entropy-based compression claims should be re-benchmarked against random pruning at matched retention rates; the paper's results predict that most reported advantages will shrink to noise.
  • On mathematical reasoning, keeping numeric tokens is what matters, and it matters only when the model can access their original activations; without patching, a trace of bare digits is not enough.
  • The failure of entropy selection on logic and science tasks suggests domain transfer cannot be assumed for this class of compression heuristics.
  • Activation patching at a small set of content-rich tokens can recover most of a full trace's accuracy, so efficient deployment might cache hidden states for selected positions rather than regenerating from compressed text.

Where Pith is reading between the lines

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

  • A testable extension: on other numeric-heavy domains (code, tabular data, scientific formulas), the numbers-vs-low-entropy dissociation should replicate; if it does not, the domain boundary of the explanation is narrower than stated.
  • The paper's top-20 entropy proxy could itself be a confound; recomputing entropies over the full vocabulary on a small subset would show whether the null result survives a more faithful entropy estimate.
  • If reasoning information truly is distributed, methods that try to locate 'critical' tokens or steps (early exit, step saliency) may be chasing a mirage; the useful signal may instead be the set of positions whose hidden states the model actually uses.
  • Because patching helps even random selection (from 73% to 32% of the trace needed for 90% retention), the practical lesson is that retrieving hidden states is a compression lever independent of any entropy heuristic.

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 tests whether entropy-based selection of CoT units (sentences or tokens) is a valid compression signal. Across six open-weight models (gpt-oss, Gemma-4, Qwen3) and six reasoning benchmarks (AIME 2024–2026, MATH-500, GPQA-Diamond, ZebraLogic), it compares low-entropy, high-entropy, numbers, low-entropy-no-numbers, and random selection under both unpatched and activation-patched compression, measuring relative performance retention (RPR) and area under the RPR curve. The main findings are: (1) sentence-level entropy selection rarely beats random and often underperforms it; (2) token-level low-entropy selection appears to help on math benchmarks, but this is attributed to overlap with numeric tokens; (3) patching retained tokens with their original activations recovers near-full performance for numbers and low-entropy selectors on math, while other selectors do not. The paper concludes that entropy is not a reliable heuristic for locating task-relevant content and that task information is not concentrated in easily identifiable tokens.

Significance. If the findings withstand scrutiny, this is an important negative result for the CoT-compression literature, which has largely lacked random baselines. The paper's strengths include a broad and systematic grid (6 models × 6 datasets × sentence/token × patched/unpatched), bootstrap CIs on the main curves, an independent semantic-verifier check, an explicit numeric-token confound analysis with worked examples, and a diagnostic activation-patching protocol. The release of code and instructions is a plus. However, the headline claim is sensitive to the top-20 entropy proxy and to the absence of CIs on aggregated ΔAUC tables; these need to be addressed before the negative conclusion can be taken as established.

major comments (4)
  1. [§4, 'Entropy measures and sentence segmentation' (footnote 7; §3 'Entropy definitions')] All rankings and therefore all conclusions are computed with H_i = −Σ_{t=1}^{20} p_i(t) log p_i(t), using top-20 log-probs from a temperature-0.7/top-p-0.9 sampler. No evidence is provided that this truncated, renormalized quantity preserves the ordering of the full next-token distribution. If tail mass is substantial for some token classes (e.g., rare tokens in natural language vs. near-deterministic numeric tokens), the proxy error is class-dependent and can invert low/high entropy rankings; a noisy proxy biases all comparisons toward the null, which is exactly the headline result. The low-entropy-no-numbers control uses the same proxy and cannot rule this out. Please add a calibration: compute full-logit entropy (or at least top-100) on a subset and report rank correlation with the k=20 proxy, tail-mass magnitude by token class, and stability of the low-vs-high entropy separation unde
  2. [§5, Tables 1, 5–7; Appendix C] The central negative claims are stated in terms of ΔAUC = AUC_s − AUC_random, but Tables 1, 5, 6, 7 report only point estimates. Appendix C describes a bootstrap for AUC but no intervals or significance tests are shown for the deltas; phrases such as 'within noise' and 'reliably outperforming chance' are therefore unsupported. In addition, 'random' appears to be a single stochastic draw for each (model, dataset, selector); the bootstrap over traces does not include variability due to the random selection itself. Please report bootstrap CIs for each ΔAUC, and either average random selection over multiple seeds or incorporate it into the CI. This is necessary for a null claim.
  3. [Abstract; §6 Conclusions vs. Tables 5–6] The abstract says 'entropy offers no advantage over random pruning in any evaluated setting' and the Conclusions say 'random pruning dominates both low- and high-entropy selection across every model, scale and domain.' These statements are contradicted by the paper's own tables. Sentence-level Table 5 has positive Δlow for Qwen3-14B on AIME24 (+0.02), AIME26 (+0.01), and GPQA-Diamond (+0.02); Table 6 shows large positive Δlow for unpatched token-level low-entropy selection on AIME benchmarks (e.g., Qwen3-4B +0.32, Qwen3-14B +0.32). The later analysis attributes those token-level gains to numeric tokens, so the qualified finding is defensible, but the unqualified 'no advantage'/'dominates' claims need to be revised to match the evidence, e.g., 'no reliable advantage once numeric-token overlap is accounted for'.
  4. [Abstract/§6 and Appendix D.3, patching discussion] The abstract claims the patching results provide 'causal evidence that task information is not concentrated in a small set of CoT tokens ... but rather distributed across the full reasoning chain.' The patching experiment replaces the hidden states at retained positions with their full-trace activations; these activations are contextual summaries of all previous tokens. As the paper itself notes in Appendix D.3, patching 'restores full-context activations at every retained position regardless of which positions were chosen.' Therefore the result that numbers patches recover 90% RPR with 9% tokens does not show that task information is distributed over the surface tokens; it shows that a few hidden states can carry global context. The conclusion should be reworded to avoid overinterpretation of the diagnostic, or an additional intervention (e.g., patching with activations from a corrupted
minor comments (5)
  1. [§4 sentence segmentation] Sentence boundaries are defined only by periods followed by whitespace; '!' and '?' are excluded. Please justify or note the impact.
  2. [Figure 2] The legend uses 'low entropy, no numbers' while the text uses 'low-entropy no numbers'; please unify the terminology.
  3. [Footnote 12] The footnote cites 'DeepScaleR (Nalaie et al., 2022)', but the reference Nalaie et al. (2022) is about DeepScale for multi-object tracking on cameras/edge servers; this does not appear to be the DeepScaleR reasoning dataset. Please correct the citation.
  4. [References] The reference list entry for OpenAI et al. (2025) has an anomalous placeholder ':' in the author field.
  5. [Limitations] Non-AIME subsets are only 50–100 questions; this is acknowledged in the Limitations, but the abstract's 'across various ... reasoning tasks' should be qualified by this sample size.

Circularity Check

0 steps flagged

No circularity: the paper's claims are empirical comparisons; the sole self-citation is a non-load-bearing implementation detail.

full rationale

The paper contains no derivation that reduces to its inputs. Entropy estimates use top-k=20 log-probabilities, citing Scalena et al. (2026b), an author-overlapping work, but this citation only motivates a truncation choice forced by vLLM constraints; it is not used to justify the central conclusion, and the paper's null result is an empirical comparison against a random baseline rather than a fitted prediction. The numeric-token confound analysis ('low-entropy no numbers' and 'numbers' selectors) is a controlled ablation, not a circular redefinition: numbers is defined by digit content, not by entropy. Activation patching is explicitly framed as a diagnostic tool and not as deployable compression, so the recovery under patching is an experimental finding, not a consequence of the entropy definition. The only potential weaknesses (truncation bias in the entropy proxy; patching injecting full-context information into retained positions) are threats to validity or interpretation, not circularity of the sort this pass targets. No self-citation is load-bearing; no prediction is equivalent to a fitted parameter.

Axiom & Free-Parameter Ledger

1 free parameters · 5 axioms · 0 invented entities

The paper is empirical; it introduces no new theory, so the ledger lists the measurement and intervention assumptions the conclusions rely on. The only hand-chosen numeric input that directly feeds the central ranking is the top-k entropy truncation.

free parameters (1)
  • top-k log-probability truncation k = 20
    Hand-picked due to vLLM API constraints; every entropy-based selector in the paper ranks tokens/sentences using this truncated distribution. The paper provides no control showing that top-20 entropy approximates full-vocabulary entropy well.
axioms (5)
  • domain assumption Shannon entropy over the top-k predictive distribution is a meaningful measure of token-level uncertainty/relevance.
    Underpins all entropy-based selectors (§3, entropy definitions).
  • domain assumption Exact-match extraction after the suffix 'Therefore, the answer is \boxed{' reflects task accuracy.
    Primary metric throughout; mitigated by math-verify semantic check in Appendix B.
  • domain assumption Residual-stream activation patching at all layers for retained positions transfers the relevant full-trace context to the compressed sequence.
    Central to the causal claim in Section 5 and Figure 4; the method is standard in interpretability (NNsight) but the paper asserts rather than validates its sufficiency.
  • domain assumption Random sentence/token sampling without replacement is a fair 'no-information' baseline.
    Defines the null model for ΔAUC; the comparison depends on this.
  • domain assumption Traces exceeding the 16384-token budget are discarded without biasing entropy distributions.
    Section 4 trace sampling; acknowledged in Limitations as a possible understatement of variance.

pith-pipeline@v1.3.0-alltime-deepseek · 19758 in / 14804 out tokens · 140257 ms · 2026-08-03T00:35:50.664748+00:00 · methodology

0 comments
read the original abstract

Entropy-based pruning has been proposed as an effective method for compressing Chain-of-Thought (CoT) reasoning with negligible accuracy loss. We test the robustness of low- and high-entropy CoT step selection methods across various models and reasoning tasks, showing that entropy offers no advantage over random pruning in any evaluated setting. Moving from sentences to tokens, we then show that retaining low-entropy tokens seems effective only on mathematical benchmarks. We find this is due to the inherently low-entropy nature of numeric tokens, which also convey semantic content in such problems. Finally, we demonstrate that patching a subset of a few CoT tokens with their original activations recovers near-perfect full-trace performance, providing causal evidence that task information is not concentrated in a small set of CoT tokens identifiable by heuristics, but rather distributed across the full reasoning chain.

Figures

Figures reproduced from arXiv: 2607.28707 by Daniel Scalena, Elisabetta Fersini, Gabriele Sarti, Luca Bortolussi, Malvina Nissim, Sara Candussio.

Figure 1
Figure 1. Figure 1: Random dominates both entropy-based selectors across the entire compression range. RPR vs. Compression Rate for gpt-oss-20b on AIME25 for various selection strategies. Shaded bands are 95% bootstrap Confidence Intervals (Appendix C). 7We use k = 20 as a proxy of the full logit distribution due to vLLM constraints, following Scalena et al. (2026b). 8Exclamation and question marks are consequently not treate… view at source ↗
Figure 2
Figure 2. Figure 2: Low-entropy token-level compression out￾performs random selection only on mathematical tasks when patching is performed, thanks to nu￾merical tokens. On MATH-500, low-entropy no num￾bers drops sharply below both low-entropy and ran￾dom, while numbers is the most performant selector even at extreme compressions. The advantage of low￾entropy on mathematical reasoning is driven by the lim￾ited choice of numer… view at source ↗
Figure 3
Figure 3. Figure 3: Patching closes the gap for low-entropy and numbers across all models on AIME25, while high-entropy and low-entropy no numbers remain largely unaffected or worse. ity of low-entropy steps are indeed redundant and thus can be discarded without accuracy degrada￾tion, while this is not the case for high-entropy and random sentence selection. Our results do also con￾firm that pruning strategies can lead to imp… view at source ↗
Figure 4
Figure 4. Figure 4: Minimum percentage of trace tokens needed to reach and sustain a [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Sentence-level full grid: Relative Performance Retention vs. compression rate for all six models (rows) × six datasets (columns), selectors random, low-entropy, high-entropy. Shaded bands are 95% bootstrap CIs (Appendix C). Random is at or above both entropy-based selectors across nearly every panel, consistent with [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Token-level full grid, no patching: Relative Performance Retention vs. compression rate, selec￾tors random, low-entropy, high-entropy, low-entropy no numbers, numbers. Model ordering: gpt-oss-20b, gpt-oss-120b, gemma-4-E4B-it, gemma-4-26B-A4B-it, Qwen3-14B, Qwen3-4B [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Token-level patched full grid: Relative Performance Retention vs. compression rate, selectors random, low-entropy, high-entropy, low-entropy no numbers, numbers. Model ordering: gpt-oss-20b, gpt-oss-120b, gemma-4-E4B-it, gemma-4-26B-A4B-it, Qwen3-14B, Qwen3-4B [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Sentence-level full grid under math-verify: Relative Performance Retention vs. compression rate for all six models (rows) × six datasets (columns), selectors random, low-entropy, high-entropy. The pattern matches [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Token-level full grid under math-verify (no patching): Relative Performance Retention vs. compression rate for all six models (rows) × six datasets (columns), selectors random, low-entropy, high-entropy. The pattern matches [PITH_FULL_IMAGE:figures/full_fig_p020_9.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

19 extracted references · 13 linked inside Pith

  1. [3]

    Preprint, arXiv:2407.14561

    Nnsight and ndif: Democratizing access to open-weight foundation model internals. Preprint, arXiv:2407.14561. Chen Huang, Wei Lu, and Wenxuan Zhang

    Pith/arXiv arXiv

  2. [4]

    What is the smallest positive perfect cube that can be written as the sum of three consecutive integers?

    and traces that do not reach[EOT]within the token budget are discarded (Section 4), n represents the upper bound of the considered traces. We resam- ple this set of n traces with replacement B= 1000 times; at each bootstrap iterationb, we draw indices {i1, . . . , in} uniformly with replacement from the original n traces and recompute RPR(b), discard- ing...

    2024

  3. [7]

    Ziqing Qiao, Yongheng Deng, Jiali Zeng, Dong Wang, Lai Wei, Fandong Meng, Jie Zhou, Ju Ren, and Yaoxue Zhang

    gpt-oss-120b and gpt- oss-20b model card.Preprint, arXiv:2508.10925. Ziqing Qiao, Yongheng Deng, Jiali Zeng, Dong Wang, Lai Wei, Fandong Meng, Jie Zhou, Ju Ren, and Yaoxue Zhang

    Pith/arXiv arXiv

  4. [8]

    David Rein, Betty Li Hou, Asa Cooper Stickland, Jack- son Petty, Richard Yuanzhe Pang, Julien Dirani, Ju- lian Michael, and Samuel R

    Concise: Confidence- guided compression in step-by-step efficient reason- ing.ArXiv, abs/2505.04881. David Rein, Betty Li Hou, Asa Cooper Stickland, Jack- son Petty, Richard Yuanzhe Pang, Julien Dirani, Ju- lian Michael, and Samuel R. Bowman

    arXiv

  5. [10]

    Preprint, arXiv:2408.03314

    Scaling llm test-time compute optimally can be more effective than scaling model parameters. Preprint, arXiv:2408.03314. Gemma Team, Sherif El Abd, Vaibhav Aggarwal, Robin Algayres, Alek Andreev, Olivier Bachem, Ian Bal- lantyne, Cormac Brick, Victor C ˘arbune, Michelle Casbon, Mayank Chaturvedi, Victor Cotruta, Alice Coucke, Phil Culliton, Robert Dadashi...

    Pith/arXiv arXiv

  6. [11]

    Preprint, arXiv:2607.02770

    Gemma 4 technical report. Preprint, arXiv:2607.02770. Jean-Francois Ton, Muhammad Faaiz Taufiq, and Yang Liu

    Pith/arXiv arXiv

  7. [12]

    InForty-second Interna- tional Conference on Machine Learning

    Understanding chain-of-thought in LLMs through information theory. InForty-second Interna- tional Conference on Machine Learning. Shenzhi Wang, Le Yu, Chang Gao, Chujie Zheng, Shix- uan Liu, Rui Lu, Kai Dang, Xionghui Chen, Jianxin Yang, Zhenru Zhang, Yuqiong Liu, An Yang, An- drew Zhao, Yang Yue, Shiji Song, Bowen Yu, Gao Huang, and Junyang Lin. 2025a. B...

    Pith/arXiv arXiv

  8. [13]

    Xuan Xiong, Huan Liu, Li Gu, Zhixiang Chi, Yue Qiu, Yuanhao Yu, and Yang Wang

    Tokenskip: Control- lable chain-of-thought compression in llms.ArXiv, abs/2502.12067. Xuan Xiong, Huan Liu, Li Gu, Zhixiang Chi, Yue Qiu, Yuanhao Yu, and Yang Wang

    arXiv

  9. [14]

    Preprint, arXiv:2604.05355

    Etr: Entropy trend reward for efficient chain-of-thought reasoning. Preprint, arXiv:2604.05355. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Day- iheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, and 41 others

    Pith/arXiv arXiv

  10. [15]

    Qwen3 technical report.Preprint, arXiv:2505.09388. An Yang, Beichen Zhang, Binyuan Hui, Bofei Gao, Bowen Yu, Chengpeng Li, Dayiheng Liu, Jian- hong Tu, Jingren Zhou, Junyang Lin, Keming Lu, Mingfeng Xue, Runji Lin, Tianyu Liu, Xingzhang Ren, and Zhenru Zhang

    Pith/arXiv arXiv

  11. [16]

    Yifan Zhang and Team Math-AI

    Qwen2.5-math tech- nical report: Toward mathematical expert model via self-improvement.Preprint, arXiv:2409.12122. Yifan Zhang and Team Math-AI

    Pith/arXiv arXiv

  12. [17]

    A Suffix ablation To rule out that our choice of answer-eliciting suf- fix (Therefore the answer is \boxed { , foot- note of Section

    Shorthand for thought: Com- pressing llm reasoning via entropy-guided superto- kens.Preprint, arXiv:2604.26355. A Suffix ablation To rule out that our choice of answer-eliciting suf- fix (Therefore the answer is \boxed { , foot- note of Section

    Pith/arXiv arXiv

  13. [18]

    does not itself bias the reported accuracies, we run a small ablation on a subset of N= 25 questions (we limit this test to one trace per question) drawn from one mathematical benchmark (AIME25) and one non-mathematical one (ZebraLogic) across all six models, comparing the default suffix against K= 3 alternative phras- ings: \boxed{, Based only on the abo...

    2026

  14. [2021]

    DeepSeek-AI

    Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168. DeepSeek-AI

    Pith/arXiv arXiv

  15. [2022]

    In 2022 IEEE/ACM Seventh International Conference on Internet-of-Things Design and Implementation (IoTDI), pages 67–79

    Deepscale: Online frame size adaptation for multi- object tracking on smart cameras and edge servers. In 2022 IEEE/ACM Seventh International Conference on Internet-of-Things Design and Implementation (IoTDI), pages 67–79. OpenAI, :, Sandhini Agarwal, Lama Ahmad, Jason Ai, Sam Altman, Andy Applebaum, Edwin Arbus, Rahul K. Arora, Yu Bai, Bowen Baker, Haimin...

    2022

  16. [2023]

    Bill Yuchen Lin, Ronan Le Bras, Kyle Richardson, Ashish Sabharwal, Radha Poovendran, Peter Clark, and Yejin Choi

    Let’s verify step by step.Preprint, arXiv:2305.20050. Bill Yuchen Lin, Ronan Le Bras, Kyle Richardson, Ashish Sabharwal, Radha Poovendran, Peter Clark, and Yejin Choi

    Pith/arXiv arXiv

  17. [2024]

    In First Conference on Language Modeling

    GPQA: A graduate-level google-proof q&a benchmark. In First Conference on Language Modeling. Daniel Scalena, Sara Candussio, Luca Bortolussi, Elis- abetta Fersini, Malvina Nissim, and Gabriele Sarti. 2026a. Beyond the commitment boundary: Probing epiphenomenal chain-of-thought in large reasoning models.Preprint, arXiv:2606.13603. Daniel Scalena, Leonidas ...

    Pith/arXiv arXiv

  18. [2025]

    arXiv preprint arXiv:2501.12948

    Deepseek-r1: Incentivizing rea- soning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948. Jaden Fiotto-Kaufman, Alexander R. Loftus, Eric Todd, Jannik Brinkmann, Koyena Pal, Dmitrii Troitskii, Michael Ripa, Adam Belfki, Can Rager, Caden Juang, Aaron Mueller, Samuel Marks, Arnab Sen Sharma, Francesca Lucchetti, Nikhil Prakash, ...

    Pith/arXiv arXiv

  19. [2026]

    InFindings of the Association for Computational Linguistics: EACL 2026, pages 5000–5006, Rabat, Morocco

    Think just enough: Leveraging self-assessed con- fidence for adaptive reasoning in language models. InFindings of the Association for Computational Linguistics: EACL 2026, pages 5000–5006, Rabat, Morocco. Association for Computational Linguistics. Hynek Kydlí ˇcek

    2026