Pith. sign in

REVIEW 3 major objections 5 minor 42 references

Standard on-policy self-distillation collapses a model’s intermediate thinking on hard reasoning; AD-OPSD stops that by shielding high-entropy forks.

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 · grok-4.5

2026-07-14 09:07 UTC pith:GHIPPVSP

load-bearing objection Clean multi-scale diagnosis of OPSD degradation on reasoning models plus a usable student-side gate; the ET-density causal story is thinner than claimed, but the empirical package still holds up. the 3 major comments →

arxiv 2607.10805 v1 pith:GHIPPVSP submitted 2026-07-12 cs.CL cs.LG

Diagnosing and Mitigating Thinking Collapse in On-Policy Self-Distillation

Keqin Peng , Chen Li , Yuanxin Ouyang , Yancheng Yuan , Liang Ding This is my paper
classification cs.CL cs.LG
keywords on-policy self-distillationthinking collapseepistemic-token densityhigh-entropy decision forkspointwise KL gatingdual-perspective targetmathematical reasoningLLM post-training
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.

On-policy self-distillation (OPSD) trains a student on its own rollouts toward a ground-truth-conditioned teacher, and works well for general alignment. On complex math reasoning, however, the same procedure can lower accuracy. The authors show this is tied to Thinking Collapse: a sharp drop in the density of intermediate “thinking” markers (wait, hmm, perhaps, and similar epistemic tokens per thousand tokens generated). Diagnostics localize the damage to high-student-entropy decision forks, where the teacher often replaces the student’s epistemic top-1 token with a non-epistemic one, and where those tokens sit in high pointwise student–teacher divergence. AD-OPSD answers with a dual-perspective target: inside a top-p% entropy sandbox, tokens under active suppression are soft-gated by pointwise KL toward a frozen base-model prior, while ordinary teacher correction is left full strength elsewhere. Across Qwen3 and DeepSeek-R1-distill scales on MATH, AIME, and HMMT, this recovers thinking density and improves over unmasked OPSD by up to 4.1 percentage points absolute average accuracy, with gains also under non-thinking training modes.

Core claim

The paper establishes that OPSD’s reasoning degradation is Thinking Collapse—suppression of native intermediate reasoning, measured as epistemic-token density—and that this collapse is concentrated at high-student-entropy forks via Student-ET to Teacher-Non-ET target discrepancies in high pointwise-divergence regions. Adaptive Dual-Perspective OPSD (AD-OPSD) mitigates it by asymmetrically blending a frozen base prior into the teacher target for sandboxed high-risk tokens, preserving exploratory paths while keeping teacher error correction elsewhere.

What carries the argument

AD-OPSD’s dual-perspective target π*_i = U_i π_base + (1−U_i) π_t, with U_i set by an asymmetrical sigmoid of the pointwise KL contribution only inside a top-p% student-entropy sandbox when log P_s > log P_t (default p=20%, τ=1).

Load-bearing premise

The claim leans on a fixed ten-word list of epistemic markers (wait, hmm, perhaps, and so on) as a length-normalized stand-in for the intermediate reasoning that actually drives accuracy.

What would settle it

Train AD-OPSD and unmasked OPSD on the same math setup, then check whether accuracy gains disappear when thinking density is measured with a different epistemic lexicon or when high-entropy forks are left unanchored while density is forced back by other means.

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

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

3 major / 5 minor

Summary. The paper identifies Thinking Collapse in on-policy self-distillation (OPSD) on reasoning LLMs: a sharp drop in epistemic-token density (ET per 1k, Eq. 1–2) that correlates with accuracy degradation under Think/Think OPSD (Fig. 1). Diagnostics localize the effect to high-student-entropy forks via gradient masking (Fig. 2a–b), initialization Student-ET→Teacher-Non-ET target discrepancies (Fig. 2c), and concentration of ETs under active suppression in high pointwise KL regions (Fig. 3, Eq. 3). The proposed AD-OPSD interpolates the teacher target with a frozen base prior inside a top-p% entropy sandbox using an asymmetrical sigmoid pointwise-KL gate (Eqs. 4–6), retaining full teacher correction elsewhere. On MATH/AIME/HMMT with Qwen3-1.7B/4B and DS-R1-Distill-1.5B, AD-OPSD improves over unmasked OPSD by up to +4.1% absolute average accuracy (Table 1), partially recovers ET density (Table 2), and generalizes to NT/NT training (Table 3).

Significance. If the diagnosis and method hold, the work supplies a concrete, student-side control for a documented failure mode of reasoning OPSD and a practical recipe (entropy sandbox + pointwise divergence gate) that is easy to implement on top of existing OPSD pipelines. Strengths include a multi-scale empirical chain (masking, static target audit, divergence mapping), clear ablations of static vs dynamic blending and alternative gates (Table 4, Fig. 4), and explicit limitations on scale and domain. The contribution is primarily empirical and engineering rather than theoretical; its value rests on whether ET density is a valid intervention target and whether the accuracy gains generalize beyond the reported math/LoRA setting.

major comments (3)
  1. §3.1 and Fig. 2a–b already show that recovering ET density is not sufficient for accuracy: Top-20%/50% entropy gradient masking restores density from 7.9 to 9.8–10.7 ET/1k while accuracy stays flat or worse (31.4–33.9%). Table 2 then shows AD-OPSD only partially recovers density (8.4–8.6 vs Base ~10.3–10.6) yet posts the largest accuracy gains. This pattern is consistent with the gate acting as a general high-entropy regularizer toward the frozen base prior rather than specifically restoring causal intermediate reasoning. The central narrative that “mitigating Thinking Collapse” (via ET density) is the operative mechanism therefore needs either (i) additional process-level metrics (e.g., verification/backtrack quality, search-tree diversity, or human-rated intermediate steps) that move with accuracy, or (ii) an explicit reframing that AD-OPSD’s benefit is soft regularization of high-entr
  2. Eqs. 1–2 and §2 define thinking density via a fixed 10-word marker set (wait, hmm, perhaps, maybe, actually, alternatively, seems, might, likely, check). This set is load-bearing for localization (Fig. 2c, Fig. 3) and for the claim that the sandbox protects “native thinking capacity.” The manuscript does not report (a) correlation of these markers with final accuracy or with independent process rewards, (b) sensitivity to marker choice/expansion, or (c) qualitative checks that protected tokens correspond to useful search rather than stylistic hedging. Without such validation, both the diagnostic story and the justification for pointwise-KL gating remain under-supported even when accuracy improves. A short ablation or correlation analysis on held-out rollouts would substantially strengthen the claim.
  3. Table 1: gains over OPSD are consistent, but absolute improvement over Base is modest and model-dependent (Qwen3-4B: 70.2 vs 69.2; DS-R1-Distill-1.5B: 35.8 vs 35.6). The abstract’s “up to +4.1%” is relative to OPSD, not Base. The paper should more carefully separate “recovery from OPSD-induced degradation” from “net improvement over the unaligned thinking prior,” and discuss when AD-OPSD is expected to beat Base rather than merely undo collapse. This is especially relevant for the harder AIME/HMMT columns where the practical value of the method is claimed.
minor comments (5)
  1. Fig. 1 and Fig. 2 plot accuracy and density on dual axes but do not always state the exact evaluation protocol (samples per prompt, temperature) in the caption; cross-reference Table 7 for clarity.
  2. Eq. 5 uses both Ui and ui; the lowercase ui is defined only in the surrounding text. A single consistent notation for the log-ratio would reduce ambiguity.
  3. §8.3 cites concurrent Purified OPSD (Shen et al., 2026b). A short explicit comparison of student-side vs teacher-side intervention (compute, when each is preferable) would help readers position the contribution.
  4. Table 6 lists “DS-OPSD Static Weight”; this appears to be a typo for SD-OPSD.
  5. Limitations correctly flag math-only and LoRA/small-scale settings; a one-sentence note on whether the entropy-sandbox percentile (default 20%) transferred without retuning across the three model families would be useful.

Circularity Check

0 steps flagged

No significant circularity: empirical diagnostics and external-benchmark evaluation; AD-OPSD gate is a stop-gradient design choice, not a fitted prediction of the reported accuracy.

full rationale

The paper's chain is observational and engineering, not a first-principles derivation that reduces to its own inputs. Thinking Collapse is defined as a measured drop in a fixed, length-normalized epistemic-token density (Eq. 1–2, markers taken from prior external work) observed under standard OPSD on held-out rollouts; the localization (high-student-entropy forks, Student-ET→Teacher-Non-ET top-1 discrepancies, concentration in high pointwise KL) is established by separate diagnostic experiments (entropy gradient masking, static target analysis at step 0, pointwise-KL histograms) that do not use the final accuracy numbers. AD-OPSD then constructs an interpolated target π* via a stop-gradient asymmetrical sigmoid of the same pointwise contribution inside a top-p% entropy sandbox (Eqs. 4–6); Ui is never fitted to the MATH/AIME/HMMT accuracies that are later reported. Those accuracies come from external competition benchmarks never used to define the loss or the markers. Self-citations (authors' prior overthinking/token-level work) appear only in related-work discussion and are not load-bearing uniqueness theorems or ansatzes that force the central claim. No parameter is fitted on a subset and then re-presented as a prediction of a closely related quantity; hyper-parameters (p%=20%, τ=1) are ablated openly. The derivation is therefore self-contained against external evidence and exhibits none of the six circular patterns.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 2 invented entities

The central claim rests on standard KL distillation math, an empirical definition of thinking density via a fixed marker list, the modeling choice that the frozen base policy is a good epistemic anchor, and a handful of hand-set control hyperparameters (sandbox percentile, gate temperature, static blend weight). No new physical entities are postulated; “Thinking Collapse,” Ui, and the dual-perspective target are operational constructs defined inside the paper.

free parameters (4)
  • Entropy sandbox percentile p%
    Default top-20% student-entropy tokens for gating; swept in {10,20,30,40,100}% and chosen where AIME accuracy peaks. Directly controls which tokens receive base anchoring.
  • Gate temperature τ
    Default 1.0 in the sigmoid of pointwise KL contribution; scales how aggressively Ui rises with suppression risk.
  • Static blend weight α (SD-OPSD)
    Default 0.5 fixed Ui inside active-suppression sandbox for the static ablation variant.
  • LoRA rank/alpha and training horizon
    r=64, α=128, 188 steps, max completion 1024 (default), LR 5e-6; all fixed shared training knobs that affect absolute numbers though not the qualitative claim.
axioms (5)
  • domain assumption Forward KL distillation of a ground-truth-conditioned teacher on student rollouts is a valid dense supervision objective (standard OPSD).
    Taken as the baseline framework from Zhao et al. 2026 and related OPD literature; §1 and §4.
  • domain assumption A fixed set of 10 English epistemic markers is a length-normalized proxy for intermediate exploratory reasoning intensity.
    Eq. 1–2 and §2, following Kim et al. 2026b; load-bearing for the Thinking Collapse narrative.
  • domain assumption High-entropy tokens are the critical decision forks where reasoning path diversity is steered.
    Invoked via Wang et al. 2026 / Xu et al. 2026 to justify entropy masking and the sandbox (§3.1, §4.2).
  • ad hoc to paper The frozen initial/base model distribution is an appropriate unsuppressed reference prior for native thinking.
    Core of dual-perspective target π* (Eq. 4); not independently proven, but ablated via SD/AD comparisons.
  • ad hoc to paper Pointwise student-teacher KL contribution under Ps>Pt marks tokens at high risk of thinking suppression.
    Justified empirically in §3.3 (Fig. 3) and used as the real-time gate score in Eq. 5.
invented entities (2)
  • Thinking Collapse no independent evidence
    purpose: Name the measured pathology: precipitous drop in ET-per-1k under reasoning OPSD correlated with accuracy loss.
    Operational definition in §2; independent_evidence is empirical within this paper’s setups, not a new physical object.
  • Teacher Unreliability Index Ui / Adaptive Dual-Perspective target π* no independent evidence
    purpose: Token-wise soft mixture of base prior and teacher for sandboxed high-risk tokens.
    Defined in Eq. 4–5 as the control mechanism of AD-OPSD; no external measurement outside the training objective.

pith-pipeline@v1.1.0-grok45 · 23623 in / 3918 out tokens · 62618 ms · 2026-07-14T09:07:15.002608+00:00 · methodology

0 comments
read the original abstract

On-Policy Self-Distillation (OPSD) has emerged as a crucial paradigm for enhancing and aligning Large Language Models (LLMs). However, in complex reasoning tasks, OPSD paradoxically degrades downstream performance. In this paper, we systematically investigate this pathology and identify a severe optimization trap we define as \textbf{Thinking Collapse} -- a sharp decline in the model's native intermediate reasoning behavior, measured by epistemic-token density (ET per 1k). Through entropy-based gradient masking and token-level target analysis, we show that this collapse is triggered by aggressive teacher gradients at high-student-entropy decision forks, where student epistemic tokens are frequently suppressed into teacher non-epistemic targets and are highly concentrated in high pointwise student-teacher divergence regions. To resolve this optimization pathology, we propose \textbf{Adaptive Dual-Perspective OPSD (AD-OPSD)}, a robust control framework that dynamically moderates the self-distillation objective. AD-OPSD selectively anchors high-suppression-risk sandboxed tokens to a reference prior derived from the frozen base model via an asymmetrical pointwise divergence gate, preserving native thinking capacity while retaining OPSD's error-correcting power. Extensive experiments across competitive mathematical benchmarks show that AD-OPSD improves over standard OPSD by up to \textbf{+4.1\%} absolute average accuracy across diverse model scales and datasets. Further analysis demonstrates that AD-OPSD mitigates thinking collapse and generalizes robustly to different post-training paradigms.

Figures

Figures reproduced from arXiv: 2607.10805 by Chen Li, Keqin Peng, Liang Ding, Yancheng Yuan, Yuanxin Ouyang.

Figure 1
Figure 1. Figure 1: AIME25 Accuracy vs. Thinking Density (ET [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Consolidated OPSD Diagnostics under Token-Level Masking and Microscopic Target Discrepancy. Subfigures (a) and (b) illustrate OPSD performance under entropy-based masking (LHS: Thinking Density, RHS: Accuracy %). Subfigure (c) depicts the cumulative student ET count (LHS, blue curve) and cumulative ET-to-Non￾ET target discrepancy transitions (RHS, orange curve) across student entropy percentiles sorted in … view at source ↗
Figure 3
Figure 3. Figure 3: Distribution and Concentration of Epis￾temic Tokens across Pointwise KL Percentiles within High-Entropy Regions (Top 20% Student Entropy) un￾der Active Suppression (Ps > Pt). Pointwise KL serves as a highly robust guidance indicator of suppression risk, allowing AD-OPSD to protect vital thinking steps by shielding regions of high pointwise divergence. teacher’s, yielding a substantial log probability ra￾ti… view at source ↗
Figure 4
Figure 4. Figure 4: Consolidated discussion and hyperparameter study for Qwen3-1.7B on AIME benchmarks. Subfigures (a) and (b) present the downstream reasoning accuracy of Standard OPSD and our proposed AD-OPSD across varying maximum context lengths (Lmax ∈ {512, 1024, 2048}) with absolute gains (∆) clearly annotated. Subfigure (c) presents the hyperparameter study on the Entropy Sandbox percentile ratio (p% ∈ {10%, 20%, 30%,… 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

42 extracted references · 2 canonical work pages

  1. [1]

    arXiv preprint arXiv:2603.24472 , year=

    Why Does Self-Distillation (Sometimes) Degrade the Reasoning Capability of LLMs? , author=. arXiv preprint arXiv:2603.24472 , year=

    Pith/arXiv arXiv

  2. [2]

    arXiv preprint arXiv:2603.15500 , year=

    Understanding Reasoning in LLMs through Strategic Information Allocation under Uncertainty , author=. arXiv preprint arXiv:2603.15500 , year=

    Pith/arXiv arXiv

  3. [3]

    NIPS , url=

    Beyond the 80/20 rule: High-entropy minority tokens drive effective reinforcement learning for llm reasoning , author=. NIPS , url=

  4. [4]

    arXiv preprint arXiv:2505.09388 , url=

    Qwen3 technical report , author=. arXiv preprint arXiv:2505.09388 , url=

    Pith/arXiv arXiv

  5. [5]

    First Workshop on Foundations of Reasoning in Language Models , year=

    OpenThoughts: Data Recipes for Reasoning Models , author=. First Workshop on Foundations of Reasoning in Language Models , year=

  6. [6]

    arXiv preprint arXiv:2601.18734 , url=

    Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models , author=. arXiv preprint arXiv:2601.18734 , url=

    Pith/arXiv arXiv

  7. [7]

    ICLR , year=

    LoRA: Low-Rank Adaptation of Large Language Models , author=. ICLR , year=

  8. [8]

    arXiv preprint arXiv:2605.11609 , url=

    Anti-Self-Distillation for Reasoning RL via Pointwise Mutual Information , author=. arXiv preprint arXiv:2605.11609 , url=

    Pith/arXiv arXiv

  9. [9]

    arXiv preprint arXiv:2605.10781 , url=

    Rebellious student: Reversing teacher signals for reasoning exploration with self-distilled rlvr , author=. arXiv preprint arXiv:2605.10781 , url=

    Pith/arXiv arXiv

  10. [10]

    arXiv preprint arXiv:2501.12948 , url=

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning , author=. arXiv preprint arXiv:2501.12948 , url=

    Pith/arXiv arXiv

  11. [11]

    ICML 2026 Workshop on Foundations of Deep Generative Models: Understanding Memorization, Generalization, and Reasoning , url=

    Rethinking On-Policy Self-Distillation for Thinking Models , author=. ICML 2026 Workshop on Foundations of Deep Generative Models: Understanding Memorization, Generalization, and Reasoning , url=

    2026

  12. [12]

    arXiv preprint arXiv:2601.19897 , url=

    Self-Distillation Enables Continual Learning , author=. arXiv preprint arXiv:2601.19897 , url=

    Pith/arXiv arXiv

  13. [13]

    arXiv preprint arXiv:2601.20802 , url=

    Reinforcement Learning via Self-Distillation , author=. arXiv preprint arXiv:2601.20802 , url=

    Pith/arXiv arXiv

  14. [14]

    arXiv preprint arXiv:2604.14084 , url=

    Tip: Token importance in on-policy distillation , author=. arXiv preprint arXiv:2604.14084 , url=

    Pith/arXiv arXiv

  15. [15]

    ICLR , year=

    Let's verify step by step , author=. ICLR , year=

  16. [16]

    2026 , eprint=

    Beyond Benchmarks: MathArena as an Evaluation Platform for Mathematics with LLMs , author=. 2026 , eprint=

    2026

  17. [17]

    American Invitational Mathematics Examination (AIME) 2024 , author=

    2024

  18. [18]

    American Invitational Mathematics Examination (AIME) 2025 , author=

    2025

  19. [19]

    arXiv preprint arXiv:1503.02531 , year=

    Distilling the Knowledge in a Neural Network , author=. arXiv preprint arXiv:1503.02531 , year=

    Pith/arXiv arXiv

  20. [20]

    NIPS , volume=

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models , author=. NIPS , volume=. 2022 , url=

    2022

  21. [21]

    ICLR , year=

    Self-Consistency Improves Chain of Thought Reasoning in Language Models , author=. ICLR , year=

  22. [22]

    arXiv preprint arXiv:2110.14168 , year=

    Training Verifiers to Solve Math Word Problems , author=. arXiv preprint arXiv:2110.14168 , year=

    Pith/arXiv arXiv

  23. [23]

    arXiv preprint arXiv:2408.03314 , year=

    Scaling LLM Test-Time Compute Optimally Can Be More Effective Than Scaling Model Parameters , author=. arXiv preprint arXiv:2408.03314 , year=

    Pith/arXiv arXiv

  24. [24]

    arXiv preprint arXiv:2402.03300 , year=

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models , author=. arXiv preprint arXiv:2402.03300 , year=

    Pith/arXiv arXiv

  25. [25]

    Findings of the Association for Computational Linguistics: ACL 2023 , year=

    Distilling Step-by-Step! Outperforming Larger Language Models with Less Training Data and Smaller Model Sizes , author=. Findings of the Association for Computational Linguistics: ACL 2023 , year=

    2023

  26. [26]

    arXiv preprint arXiv:2604.03128 , url=

    Self-distilled rlvr , author=. arXiv preprint arXiv:2604.03128 , url=

    Pith/arXiv arXiv

  27. [27]

    Dao, Tri , booktitle=. Flash

  28. [28]

    ICLR , url=

    Decoupled Weight Decay Regularization , author=. ICLR , url=

  29. [29]

    arXiv preprint arXiv:2605.27028 , url=

    Less is More: Early Stopping Rollout for On-Policy Distillation , author=. arXiv preprint arXiv:2605.27028 , url=

    Pith/arXiv arXiv

  30. [30]

    Findings of ACL , url=

    Fast and effective on-policy distillation from reasoning prefixes , author=. Findings of ACL , url=

  31. [31]

    ICLR , year=

    On-Policy Distillation of Language Models: Learning from Self-Generated Mistakes , author=. ICLR , year=

  32. [32]

    arXiv preprint arXiv:2604.13016 , year=

    Rethinking On-Policy Distillation of Large Language Models: Phenomenology, Mechanism, and Recipe , author=. arXiv preprint arXiv:2604.13016 , year=

    Pith/arXiv arXiv

  33. [33]

    arXiv preprint arXiv:2605.11182 , year=

    The Many Faces of On-Policy Distillation: Pitfalls, Mechanisms, and Fixes , author=. arXiv preprint arXiv:2605.11182 , year=

    Pith/arXiv arXiv

  34. [34]

    arXiv preprint arXiv:2602.12275 , year=

    On-Policy Context Distillation for Language Models , author=. arXiv preprint arXiv:2602.12275 , year=

    Pith/arXiv arXiv

  35. [35]

    arXiv preprint arXiv:2604.12002 , year=

    Self-Distillation Zero: Self-Revision Turns Binary Rewards into Dense Supervision , author=. arXiv preprint arXiv:2604.12002 , year=

    Pith/arXiv arXiv

  36. [36]

    arXiv preprint arXiv:2607.02234 , year=

    Purified OPSD: On-Policy Self-Distillation Without Losing How to Think , author=. arXiv preprint arXiv:2607.02234 , year=

    Pith/arXiv arXiv

  37. [37]

    arXiv preprint arXiv:2603.25562 , year=

    Revisiting On-Policy Distillation: Empirical Failure Modes and Simple Fixes , author=. arXiv preprint arXiv:2603.25562 , year=

    Pith/arXiv arXiv

  38. [38]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , pages=

    Token-Level Self-Evolution Training for Sequence-to-Sequence Learning , author=. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , pages=. 2023 , address=. doi:10.18653/v1/2023.acl-short.73 , url=

    work page doi:10.18653/v1/2023.acl-short.73 2023

  39. [39]

    arXiv preprint arXiv:2505.23480 , year=

    Revisiting Overthinking in Long Chain-of-Thought from the Perspective of Self-Doubt , author=. arXiv preprint arXiv:2505.23480 , year=. doi:10.48550/arXiv.2505.23480 , url=

    Pith/arXiv arXiv doi:10.48550/arxiv.2505.23480

  40. [40]

    arXiv preprint arXiv:2602.02099 , year=

    Think Dense, Not Long: Dynamic Decoupled Conditional Advantage for Efficient Reasoning , author=. arXiv preprint arXiv:2602.02099 , year=. doi:10.48550/arXiv.2602.02099 , url=

    doi:10.48550/arxiv.2602.02099

  41. [41]

    arXiv preprint arXiv:2605.24998 , year=

    Better, Faster: Harnessing Self-Improvement in Large Reasoning Models , author=. arXiv preprint arXiv:2605.24998 , year=. doi:10.48550/arXiv.2605.24998 , url=

    Pith/arXiv arXiv doi:10.48550/arxiv.2605.24998

  42. [42]

    arXiv preprint arXiv:2511.12661 , year=

    Reason-KE++: Aligning the Process, Not Just the Outcome, for Faithful LLM Knowledge Editing , author=. arXiv preprint arXiv:2511.12661 , year=. doi:10.48550/arXiv.2511.12661 , url=

    work page doi:10.48550/arxiv.2511.12661