Pith. sign in

REVIEW 1 major objections 7 minor 27 references

Reviewed by Pith at T0; open to challenge.

T0 means a machine referee read the full paper against a public rubric. The mark states how deep the mechanical check went, never who wrote it. the ladder, T0–T4 →

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

T0 review · glm-5.2

Thinking can make AI less factual — now there's a fix

2026-07-08 22:08 UTC pith:GJJTWWJP

load-bearing objection MARGO mixes thinking and non-thinking rollouts in the same GRPO group — a clean, genuinely new mechanism for a real problem. The math is correct, the method is lightweight, and math ability is preserved. The main concern is evaluation validity: one judge model (Qwen3-32B) is used for data selection, training rewards, and all test evaluation, with no error bars. Still worth a serious referee. the 1 major comments →

arxiv 2607.05861 v1 pith:GJJTWWJP submitted 2026-07-07 cs.CL cs.LG

Mitigating Factual Hallucination in Large Reasoning Models via Mixed-Mode Advantage Regularization

Kaishen Wang , Tong Zheng , Xuehao Cui , Ruibo Chen , Tianyi Xiong , Heng Huang This is my paper
classification cs.CL cs.LG
keywords thinkingtextitunderlineexplicitfactualanswersmargomodel
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.

Large reasoning models generate explicit thinking traces before answering questions. This paper identifies a failure mode where thinking actually corrupts correct answers: the model knows the right answer without thinking, but generating a reasoning trace introduces entity confusion, unsupported associations, and factual drift that flips the correct answer into a wrong one. The authors call this 'thinking-induced hallucination.' They show this happens in 5-8% of factual QA questions across multiple model scales, and is largely absent in math reasoning. To address it, they propose MARGO, a reinforcement learning method that places both thinking and non-thinking responses to the same question into one comparison group during training. This lets the optimizer see whether thinking actually adds factual value over direct answering, rather than just ranking thinking trajectories against each other. The key mathematical insight is that mixing modes introduces a 'residual value' term into the advantage computation: if thinking helps, it gets reinforced; if thinking hurts, it gets penalized relative to the non-thinking reference.

Core claim

The paper's central contribution is the decomposition of mixed-mode advantage into two terms: a within-mode advantage (how good this thinking trajectory is compared to other thinking trajectories) and a residual-value adjustment (whether thinking as a whole helps or hurts factuality for this question, measured against the non-thinking baseline). This second term is absent in standard GRPO, which only compares thinking trajectories to each other. By including non-thinking rollouts in the same group, MARGO automatically suppresses thinking that corrupts correct direct answers while preserving thinking that recovers missing knowledge, without requiring a separate classifier or hand-crafted mode

What carries the argument

MARGO constructs mixed rollout groups containing both thinking and non-thinking trajectories for each question, then applies standard GRPO advantage normalization across the mixed group. The thinking ratio α controls the proportion of thinking vs non-thinking samples (set to 0.75 in experiments, meaning 6 thinking and 2 non-thinking rollouts per group).

Load-bearing premise

The same model (Qwen3-32B) serves as both the training reward signal and the evaluation judge, using the same prompt. If this judge has systematic biases toward certain answer formats or phrasings that MARGO's outputs happen to match, the reported gains could be partially artifactual rather than reflecting genuine factual improvement.

What would settle it

Run MARGO with a different judge model for evaluation than for training, or add human evaluation on a sample of outputs, and check whether the gains persist.

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

1 major / 7 minor

Summary. This paper introduces MARGO, a mixed-mode GRPO variant that places thinking and non-thinking trajectories in the same rollout group for factuality-oriented QA. The core idea is that non-thinking responses serve as same-model references in advantage estimation, so that thinking trajectories are penalized when they degrade factuality relative to direct answering. The advantage decomposition (Proposition 1, Appendix A) shows that the mixed-mode advantage of a thinking trajectory decomposes into a within-thinking advantage plus a residual-value adjustment proportional to the thinking/non-thinking reward gap. Experiments on Qwen3-4B and Qwen3-8B across six factuality QA benchmarks show consistent improvements over fixed-mode and adaptive baselines, while mathematical reasoning ability is preserved. The transition analysis (Tables 1, 6–8) documenting thinking-induced hallucination is a useful empirical contribution.

Significance. The paper addresses a real and timely problem: thinking-induced hallucination in large reasoning models on factuality tasks. The mixed-mode advantage decomposition (Proposition 1) is a clean, correct mathematical result that directly motivates the method. The transition analysis across multiple benchmarks and model scales (Tables 1, 6–8, including the GSM8K contrast in Table 8) provides solid empirical grounding for the phenomenon. The data selection ablation (Table 3) and the transition-ratio analysis (Figure 2a–b) are well-designed controls that isolate the contribution of mixed-mode regularization from RL or data selection alone. The method is simple to implement and the experimental setup is reasonably comprehensive.

major comments (1)
  1. §4.1, Appendix B, Appendix C.1: The same judge model (Qwen3-32B) with the same prompt is used for (a) training data filtering (Eqs. 25–32), (b) reward computation during RL training (Eq. 13), and (c) all six test-benchmark evaluations. This creates a measurement loop: if Qwen3-32B has systematic biases in semantic matching (e.g., favoring certain answer formats, lengths, or phrasings), MARGO could learn to exploit these during RL training, and the same biases would inflate test scores. The paper acknowledges that the judge performs ground-truth-conditioned matching rather than open-ended verification (Appendix B), which mitigates but does not eliminate this concern. The SimpleQA-V column in Table 2 is particularly striking: on Qwen3-4B, MARGO jumps from 4.10% (FixedThink) to 10.00% (+5.90 absolute, +144% relative), while FixedThink+RL stays flat at 4.10%. This is more than double the per
minor comments (7)
  1. Table 2: No error bars, confidence intervals, or significance tests are reported for any benchmark. Given that several gains are in the 1–3% range (e.g., TriviaQA 4B: +1.68% over FixedThink+RL), it is difficult to assess whether these are within sampling noise. Adding bootstrap confidence intervals or at least reporting results over multiple random seeds would strengthen the claims.
  2. §4.1: The evaluation uses temperature 0.6 for thinking and 0.7 for non-thinking mode. It would help to clarify whether the reported numbers are single-run or averaged over multiple samples, and whether the same decoding settings were used for all baselines.
  3. Table 3 (data selection ablation) is only reported on Qwen3-4B. Including the 8B ablation would make the data-selection claim more robust.
  4. Figure 2a–b: The y-axis labels and bar values are somewhat hard to read. Consider using a table format or increasing font size.
  5. §C.1, Eqs. 29–30: The threshold for thinking-favored examples on Qwen3-4B requires Δ(x) ≥ 1.0 and S_T(x) ≥ 1.0. Since S_T is computed from N=6 samples, S_T(x) ≥ 1.0 means all 6 thinking samples are correct. This is a very strict threshold; it would help to report how many examples satisfy each condition.
  6. The paper would benefit from a brief discussion of how α (the thinking ratio) was chosen. Only α=0.75 is tested; a small ablation over α values would clarify sensitivity.
  7. References: Several entries have incomplete formatting (e.g., 'Guibin Zhang et al.' missing venue/year; 'Pingzhi Li et al.' missing year).

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the careful and constructive review. The referee identifies a valid concern about the measurement loop created by using the same judge model (Qwen3-32B) for training data filtering, RL reward computation, and test evaluation. We agree this is a legitimate methodological concern and will address it in revision by adding an independent evaluation judge and discussing the limitations of the shared-judge design. The core contributions—the thinking-induced hallucination phenomenon, the advantage decomposition, and the mixed-mode regularization mechanism—are independent of the specific judge used, but we acknowledge the referee's point that the current evaluation setup does not fully rule out judge-bias exploitation.

read point-by-point responses

  1. Referee: §4.1, Appendix B, Appendix C.1: The same judge model (Qwen3-32B) with the same prompt is used for (a) training data filtering, (b) reward computation during RL training, and (c) all six test-benchmark evaluations. This creates a measurement loop where MARGO could learn to exploit Qwen3-32B's systematic biases during RL training, and the same biases would inflate test scores. The SimpleQA-V result on Qwen3-4B (4.10% to 10.00%, +144% relative) is particularly striking.

    Authors: We agree with the referee that using the same judge model for training data filtering, RL reward computation, and test evaluation creates a potential measurement loop. This is a legitimate methodological concern. We will address it through the following revisions: (1) We will add an independent evaluation using a different judge model (e.g., GPT-4o or Llama-3.3-70B) on at least a subset of benchmarks to verify that MARGO's improvements are not artifacts of Qwen3-32B's biases. (2) We will add an explicit discussion of this limitation in the paper, acknowledging the shared-judge concern and explaining why ground-truth-conditioned semantic matching (as opposed to open-ended verification) mitigates—but does not eliminate—the risk. (3) Regarding the SimpleQA-V jump specifically: we note that FixedThink+RL, which uses the same judge for reward computation and the same training data, does NOT show this gain (staying at 4.10%), which suggests the improvement is attributable to the mixed-mode advantage mechanism rather than simple judge exploitation. If MARGO were merely learning to exploit judge biases, we would expect FixedThink+RL to benefit similarly since it shares the same reward signal. That said, we acknowledge this argument is not fully conclusive without an independent judge, and we will provide the cross-judge evaluation to properly address the concern. revision: yes

Circularity Check

0 steps flagged

No circularity found: advantage decomposition is a genuine algebraic identity; judge reuse is a measurement concern, not a circularity.

full rationale

The paper's central theoretical result (Proposition 1, Eq. 19/22) is a straightforward algebraic identity: substituting the mixed-baseline definition μ_mix = αμ_T + (1-α)μ_N into R(y_T) - μ_mix and rearranging yields the decomposition [R(y_T) - μ_T] + (1-α)Δ_res. This is not circular—it is a valid mathematical derivation from stated definitions. The training data filtering (Appendix C.1, Eqs. 25-32) selects examples with large thinking/non-thinking factuality gaps, which is a legitimate data curation choice, not a fitted input renamed as prediction. The experimental results compare MARGO against multiple baselines (FixedThink+RL, FixedNoThink+RL, AdaptiveRL, etc.) trained on the same data, isolating the mixed-mode advantage mechanism from RL alone. The reader's concern about Qwen3-32B serving as judge for both training rewards and evaluation is a valid measurement-validity concern (potential systematic bias inflating results), but it is not circularity in the sense of a derivation reducing to its inputs by construction. The judge performs ground-truth-conditioned semantic matching, not open-ended factual verification, and the same judge is applied to all baselines equally. No self-citation chain is load-bearing for the central claim. The paper is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

7 free parameters · 4 axioms · 0 invented entities

The paper introduces no new entities (particles, forces, dimensions, etc.). The 'thinking residual' and 'residual value' are conceptual reformulations of the difference between two expected rewards under different prompting modes, not new postulated objects. All free parameters are standard RL hyperparameters or data filtering thresholds. The most consequential axiom is the judge reliability assumption, which is load-bearing for both training and evaluation.

free parameters (7)
  • α (thinking ratio in rollout group) = 0.75
    6 thinking + 2 non-thinking rollouts per group. Not tuned via systematic search; the paper does not report ablations over α.
  • Learning rate = 1e-6
    Standard for RL fine-tuning of reasoning models; not tuned.
  • KL loss coefficient = 0.001
    Standard stabilization; not tuned.
  • Data selection thresholds (Qwen3-4B) = Δ≤-0.7, SN≥0.7 / Δ≥1.0, ST≥1.0
    Model-specific thresholds for filtering training examples with large thinking/non-thinking gaps. Chosen heuristically; different thresholds for 8B model.
  • Data selection thresholds (Qwen3-8B) = Δ≤-0.5, SN≥0.5 / Δ≥0.7, ST≥0.7
    Relaxed thresholds for larger model. Justified informally ('gap becomes less extreme') but not systematically validated.
  • Format reward bonus = 0.05
    Small bonus for valid response formatting; not tuned.
  • Correctness reward = +1/-1
    Binary reward scale; not tuned.
axioms (4)
  • standard math GRPO with group-relative advantage normalization is a valid RL training procedure for LLMs
    Established by Shao et al. 2024; used as the base framework.
  • domain assumption Qwen3-32B can reliably judge semantic equivalence between model predictions and ground-truth answers in factual QA
    Used as both training reward signal and evaluation metric (§4.1, Appendix B). No human evaluation or alternative judge validates this assumption.
  • domain assumption The thinking/non-thinking mode distinction (via prompting template) captures a meaningful behavioral difference in the model's factual tendency
    §3.1 defines modes as template-determined; the residual-value formulation (§3.3) depends on non-thinking responses reflecting 'direct factual tendency.'
  • ad hoc to paper Training examples with large thinking/non-thinking factuality gaps are more informative for mixed-mode optimization than random examples
    Appendix C.1 defines the filtering criteria; Table 3 ablation supports this but only compares against random selection, not against alternative filtering strategies.

pith-pipeline@v1.1.0-glm · 23500 in / 3227 out tokens · 576892 ms · 2026-07-08T22:08:18.145567+00:00 · methodology

0 comments
read the original abstract

Large reasoning models (LRMs) improve language model capabilities by generating explicit thinking traces before final answers. In factuality-oriented question answering (QA), such thinking often improves overall performance by helping the model recover relevant knowledge and refine its answers. However, we find that this benefit is not uniform at the instance level: explicit thinking can also overturn correct non-thinking answers and lead to factual drift. We refer to this failure mode as \emph{thinking-induced hallucination}. To explain this phenomenon, we formulate explicit thinking in factuality QA as a thinking residual over the model's direct-answer tendency, which can either recover missing knowledge or introduce unsupported associations. Based on this formulation, we propose MARGO, \underline{\textit{M}}ixed-Mode \underline{\textit{A}}dvantage \underline{\textit{R}}egularization for \underline{\textit{G}}rounded \underline{\textit{O}}ptimization, a reinforcement learning framework that uses non-thinking rollouts as same-model references in advantage estimation. By constructing mixed-mode rollout groups with both thinking and non-thinking trajectories, MARGO evaluates whether explicit thinking adds factual value beyond direct answering, thereby suppressing hallucination-prone thinking while preserving beneficial thinking behaviors. Experiments across multiple factuality-oriented QA benchmarks demonstrate that MARGO improves factual reliability over strong baselines, while evaluations on mathematical benchmarks show that it preserves general reasoning ability.

Figures

Figures reproduced from arXiv: 2607.05861 by Heng Huang, Kaishen Wang, Ruibo Chen, Tianyi Xiong, Tong Zheng, Xuehao Cui.

Figure 1
Figure 1. Figure 1: Overview of MARGO. (a) Explicit thinking can induce hallucination by overturning a correct non-thinking answer into an incorrect prediction. (b) We view explicit thinking in LRMs as a residual over the model’s direct-answer tendency, which can be either helpful or harmful. (c) Unlike standard GRPO, which compares only thinking rollouts, MARGO uses non-thinking rollouts as same-model references to suppress … view at source ↗
Figure 2
Figure 2. Figure 2: Analysis of transition-ratio changes and mathematical generalization. (a) Changes in [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative case study comparing Fixed [PITH_FULL_IMAGE:figures/full_fig_p008_3.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

27 extracted references · 27 canonical work pages · 14 internal anchors

  1. [1]

    Improving text-to-image generation with input-side inference-time scaling.arXiv preprint arXiv:2510.12041, 2025a

    Ruibo Chen, Jiacheng Pan, Heng Huang, and Zhenheng Yang. Improving text-to-image generation with input-side inference-time scaling.arXiv preprint arXiv:2510.12041, 2025a. Xilun Chen, Ilia Kulikov, Vincent-Pierre Berges, Barlas O ˘guz, Rulin Shao, Gargi Ghosh, Jason Weston, and Wen-tau Yih. Learning to reason for factuality.arXiv preprint arXiv:2508.05618,...

    work page arXiv

  2. [2]

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

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

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    Simpleqa verified: A reliable factuality benchmark to measure parametric knowledge.arXiv preprint arXiv:2509.07968,

    Lukas Haas, Gal Yona, Giovanni D’Antonio, Sasha Goldshtein, and Dipanjan Das. Simpleqa verified: A reliable factuality benchmark to measure parametric knowledge.arXiv preprint arXiv:2509.07968,

    work page arXiv

  4. [4]

    Your models have thought enough: Training large reasoning models to stop overthinking.arXiv preprint arXiv:2509.23392,

    Jinyi Han, Ying Huang, Ying Liao, Zishang Jiang, Xikun Lu, Haiquan Zhao, Xinyi Wang, Guanghao Zhou, Sihang Jiang, Jiaqing Liang, et al. Your models have thought enough: Training large reasoning models to stop overthinking.arXiv preprint arXiv:2509.23392,

    work page arXiv

  5. [5]

    DeBERTaV3: Improving DeBERTa using ELECTRA-Style Pre-Training with Gradient-Disentangled Embedding Sharing

    Pengcheng He, Jianfeng Gao, and Weizhu Chen. Debertav3: Improving deberta using electra-style pre-training with gradient-disentangled embedding sharing.arXiv preprint arXiv:2111.09543,

    work page internal anchor Pith review Pith/arXiv arXiv

  6. [6]

    T1: Advancing Language Model Reasoning through Reinforcement Learning and Inference Scaling

    Zhenyu Hou, Xin Lv, Rui Lu, Jiajie Zhang, Yujiang Li, Zijun Yao, Juanzi Li, Jie Tang, and Yuxiao Dong. T1: Advancing language model reasoning through reinforcement learning and inference scaling.arXiv preprint arXiv:2501.11651,

    work page internal anchor Pith review Pith/arXiv arXiv

  7. [7]

    OpenAI o1 System Card

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

    work page internal anchor Pith review Pith/arXiv arXiv

  8. [8]

    Not All Turns Are Equally Hard: Adaptive Thinking Budgets For Efficient Multi-Turn Reasoning

    Neharika Jali, Anupam Nayak, and Gauri Joshi. Not all turns are equally hard: Adaptive thinking budgets for efficient multi-turn reasoning.arXiv preprint arXiv:2604.05164,

    work page internal anchor Pith review Pith/arXiv arXiv

  9. [9]

    Fabio Petroni, Tim Rocktäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. Language models as knowledge bases? InProceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP), pages 2463–2473,

    work page 2019

  10. [10]

    Adam Roberts, Colin Raffel, and Noam Shazeer. How much knowledge can you pack into the parameters of a language model? InProceedings of the 2020 conference on empirical methods in natural language processing (EMNLP), pages 5418–5426,

    work page 2020

  11. [11]

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathemat- ical reasoning in open language models.arXiv preprint arXiv:2402.03300,

    work page internal anchor Pith review Pith/arXiv arXiv

  12. [12]

    Between Underthinking and Overthinking: An Empirical Study of Reasoning Length and correctness in LLMs

    Jinyan Su, Jennifer Healey, Preslav Nakov, and Claire Cardie. Between underthinking and over- thinking: An empirical study of reasoning length and correctness in llms.arXiv preprint arXiv:2505.00127,

    work page internal anchor Pith review Pith/arXiv arXiv

  13. [13]

    Detection and Mitigation of Hallucination in Large Reasoning Models: A Mechanistic Perspective

    Zhongxiang Sun, Qipeng Wang, Haoyu Wang, Xiao Zhang, and Jun Xu. Detection and miti- gation of hallucination in large reasoning models: A mechanistic perspective.arXiv preprint arXiv:2505.12886,

    work page internal anchor Pith review Pith/arXiv arXiv

  14. [14]

    AdapThink: Adaptive Thinking Preferences for Reasoning Language Model

    Xu Wan, Wei Wang, Wenyue Xu, Wotao Yin, Jie Song, and Mingyang Sun. Adapthink: Adaptive thinking preferences for reasoning language model.arXiv preprint arXiv:2506.18237,

    work page internal anchor Pith review Pith/arXiv arXiv

  15. [15]

    Unsafe by reciprocity: How generation-understanding coupling undermines safety in unified multimodal models.arXiv preprint arXiv:2603.27332,

    Kaishen Wang and Heng Huang. Unsafe by reciprocity: How generation-understanding coupling undermines safety in unified multimodal models.arXiv preprint arXiv:2603.27332,

    work page arXiv

  16. [16]

    Self-Consistency Improves Chain of Thought Reasoning in Language Models

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

    work page internal anchor Pith review Pith/arXiv arXiv

  17. [17]

    Measuring short-form factuality in large language models

    Jason Wei, Nguyen Karina, Hyung Won Chung, Yunxin Joy Jiao, Spencer Papay, Amelia Glaese, John Schulman, and William Fedus. Measuring short-form factuality in large language models. arXiv preprint arXiv:2411.04368,

    work page internal anchor Pith review Pith/arXiv arXiv

  18. [18]

    Qwen3 Technical Report

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388,

    work page internal anchor Pith review Pith/arXiv arXiv

  19. [19]

    Hotpotqa: A dataset for diverse, explainable multi-hop question answering

    Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question answering. InProceedings of the 2018 conference on empirical methods in natural language processing, pages 2369–2380,

    work page 2018

  20. [20]

    Are Reasoning Models More Prone to Hallucination?

    Zijun Yao, Yantao Liu, Yanxu Chen, Jianhui Chen, Junfeng Fang, Lei Hou, Juanzi Li, and Tat-Seng Chua. Are reasoning models more prone to hallucination?arXiv preprint arXiv:2505.23646,

    work page internal anchor Pith review Pith/arXiv arXiv

  21. [21]

    Hallurnn: Mitigating hallucinations via recurrent cross-layer reasoning in large vision-language models

    Le Yu, Kaishen Wang, Jianlong Xiong, Yue Cao, Lei Zhang, and Zhang Yi Tao He. Hallurnn: Mitigating hallucinations via recurrent cross-layer reasoning in large vision-language models. arXiv preprint arXiv:2506.17587,

    work page arXiv

  22. [22]

    Unraveling hallucination in large reasoning models: A topological perspective

    Guibin Zhang, Yuxiang Zhang, Moayad Aloqaily, Haolang Lu, Kun Wang, Jing Liang, Xingjun Ma, Yu-Gang Jiang, and Qingsong Wen. Unraveling hallucination in large reasoning models: A topological perspective. Jiajie Zhang, Nianyi Lin, Lei Hou, Ling Feng, and Juanzi Li. Adaptthink: Reasoning models can learn when to think. InProceedings of the 2025 Conference o...

    work page 2025

  23. [23]

    Progressive-Hint Prompting Improves Reasoning in Large Language Models

    Chuanyang Zheng, Zhengying Liu, Enze Xie, Zhenguo Li, and Yu Li. Progressive-hint prompting improves reasoning in large language models.arXiv preprint arXiv:2304.09797,

    work page internal anchor Pith review Pith/arXiv arXiv

  24. [24]

    Parallel-probe: Towards efficient parallel thinking via 2d probing.arXiv preprint arXiv:2602.03845,

    Tong Zheng, Chengsong Huang, Runpeng Dai, Yun He, Rui Liu, Xin Ni, Huiwen Bao, Kaishen Wang, Hongtu Zhu, Jiaxin Huang, et al. Parallel-probe: Towards efficient parallel thinking via 2d probing.arXiv preprint arXiv:2602.03845,

    work page arXiv

  25. [25]

    A”, “B”, or “C

    12 Limitations This work focuses on improving the factual reliability of LRMs in factuality-oriented QA. While our experiments cover multiple benchmarks and model scales, future work can further explore the applicability of mixed-mode advantage regularization to broader scenarios, such as open-ended generation, multi-turn interaction, retrieval-augmented ...

    work page 1998

  26. [26]

    it is possible that

    ,→ ,→ ``` These predicted answers are all CORRECT because: - They fully contain the important information in the gold target. - They do not contain any information that contradicts the gold target. - Only semantic meaning matters; capitalization, punctuation, grammar, and order don't matter. - Hedging and guessing are permissible, provided that the gold t...

    work page 2025

  27. [27]

    Specifically, we report Mean@16 and Pass@16 on AMC23, AIME24, and AIME25 for both Qwen3-4B and Qwen3-8B. Compared with 17 Method AMC23 AIME24 AIME25 AverageMean@16 Pass@16 Mean@16 Pass@16 Mean@16 Pass@16 Qwen3-4B FixedN oT hink 67.0 92.5 23.3 53.3 21.7 46.7 37.3 FixedT hink 95.8 100.0 73.8 83.3 64.4 86.7 78.0 MARGO96.2100.073.890.065.490.078.5 Qwen3-8B Fi...

    work page 2021