pith. sign in

arxiv: 2606.29481 · v1 · pith:KV26OFPBnew · submitted 2026-06-28 · 💻 cs.CL · cs.AI

To Reason or to Fabricate: Reasoning Without Shortcuts via Hint-Anchored Pairwise Aggregation

Jiuheng Lin , Chen Zhang , Yansong Feng This is my paper

Pith reviewed 2026-06-30 07:33 UTC · model grok-4.3

classification 💻 cs.CL cs.AI
keywords LLM reasoningreinforcement learningdata overlapshortcut learningpairwise reward modelhint injectionprocess reward modelsgeneralization
0
0 comments X

The pith

Hint injection creates explicit anchors that let a pairwise judge separate genuine LLM reasoning from post-hoc fabrication driven by data overlap.

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

The paper introduces HIPPO, an RL framework that deliberately injects hints into reasoning traces to trigger the shortcut behaviors caused by overlap between RL data and pretraining or SFT corpora. These triggered traces then serve as clear reference points for a pairwise reward model that learns to prefer authentic deduction over fabricated rationalizations. Because the comparison is pairwise rather than absolute, the resulting preference signals are more stable and allow the policy to acquire reasoning that transfers to out-of-distribution tasks. A reader would care if this approach actually produces models whose reasoning steps are not just memorized answers dressed up after the fact.

Core claim

By injecting hints to surface overlap-induced shortcuts and then performing pairwise aggregation on the resulting traces, HIPPO supplies a lightweight judge model with highly discriminable signals that distinguish genuine step-by-step deduction from post-hoc fabrication, yielding policies whose reasoning generalizes beyond the training distribution.

What carries the argument

Hint-anchored pairwise aggregation: hint injection surfaces shortcut traces that become explicit anchors for pairwise preference labeling, which trains the reward model to score authentic reasoning higher than fabricated alternatives.

If this is right

  • Standard process reward models are replaced by a pairwise formulation that avoids unstable absolute scoring.
  • Reasoning policies acquire skills that transfer to general out-of-distribution tasks rather than remaining tied to memorized patterns.
  • The same hint-anchoring technique can be applied to any RL setup where data overlap encourages post-hoc rationalization.
  • Training becomes more robust because the judge receives explicit positive and negative examples anchored to the same hint.

Where Pith is reading between the lines

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

  • The method could be extended to automatically discover effective hints instead of requiring manual injection.
  • If the pairwise signal proves reliable, it might reduce reliance on expensive curation of non-overlapping training sets.
  • Similar anchoring ideas might help in other domains where models exploit surface correlations, such as code synthesis or factual question answering.

Load-bearing premise

Deliberately injected hints will reliably elicit the overlap-driven shortcut behaviors and make them distinguishable enough for the pairwise judge to learn a stable separation between real and fabricated reasoning.

What would settle it

A controlled experiment in which models trained with HIPPO show no gain over standard RL baselines on held-out tasks that have documented high overlap with the pretraining corpus, or lose their advantage once the hints are removed at test time.

Figures

Figures reproduced from arXiv: 2606.29481 by Chen Zhang, Jiuheng Lin, Yansong Feng.

Figure 1
Figure 1. Figure 1: An illustration of Pre-RL data overlap, using Deep [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The overall framework of HIPPO. Given a task, HIPPO performs standard RL rollouts paired with hint-injected [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Response quality of different methods on MedQA [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Score comparison between the direct and hint [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of PNT statistics between the direct [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Logit Lens visualizations of answer probability [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
read the original abstract

While reinforcement learning (RL) significantly enhances LLM reasoning, its efficacy is severely undermined by Pre-RL data overlap, where RL datasets overlap with pretraining or SFT corpora, causing models to exploit shortcuts by memorizing correct answers and fabricating post-hoc reasoning. To address this, we introduce HIPPO, a novel RL framework that integrates hint-injected aggregation with a tailored pairwise reward model. By utilizing hint injection to deliberately trigger overlap-induced behaviors, the resulting traces naturally serve as explicit anchors for pairwise comparison. This provides highly discriminable preference signals, enabling a lightweight judge model to reliably distinguish genuine reasoning deduction from shortcut-driven rationalization, while the pairwise formulation ensures stable and robust optimization compared to standard PRMs. Extensive experiments demonstrate that HIPPO yields substantial improvements over standard baselines and generalizes effectively to out-of-distribution general tasks, showing it extracts authentic, transferable reasoning skills rather than superficial shortcut patterns.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 0 minor

Summary. The manuscript introduces HIPPO, a reinforcement learning framework for LLM reasoning that integrates hint-injected aggregation with a pairwise reward model. Hint injection is used to deliberately trigger overlap-induced shortcut behaviors from pre-RL data overlap, which then serve as explicit anchors for pairwise comparison. This is claimed to enable a lightweight judge model to distinguish genuine reasoning deduction from shortcut-driven rationalization, yielding more stable optimization than standard PRMs. The paper asserts that HIPPO produces substantial improvements over baselines and effective generalization to out-of-distribution tasks by extracting authentic, transferable reasoning skills.

Significance. If the empirical claims hold, the work would address an important practical problem in RL for reasoning models, where data overlap leads to memorization and post-hoc fabrication rather than authentic deduction. The pairwise formulation for preference signals is presented as a technical advantage for stability. However, the abstract supplies no experimental details, baselines, metrics, controls, or quantitative results, preventing any assessment of whether the method actually achieves the claimed separation of reasoning from shortcuts or the reported gains.

major comments (1)
  1. [Abstract] Abstract: the central claims of 'substantial improvements over standard baselines' and 'generalizes effectively to out-of-distribution general tasks' are presented without any experimental details, datasets, baselines, metrics, ablations, or quantitative results. This absence makes the load-bearing empirical assertions unverifiable and prevents evaluation of whether the hint-anchored pairwise mechanism actually extracts authentic reasoning rather than fitting to the injected hints.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the detailed feedback. The primary concern raised is the lack of experimental specifics in the abstract, which we address directly below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claims of 'substantial improvements over standard baselines' and 'generalizes effectively to out-of-distribution general tasks' are presented without any experimental details, datasets, baselines, metrics, ablations, or quantitative results. This absence makes the load-bearing empirical assertions unverifiable and prevents evaluation of whether the hint-anchored pairwise mechanism actually extracts authentic reasoning rather than fitting to the injected hints.

    Authors: We agree that the abstract, constrained by length, omits specific experimental details and thus does not allow standalone verification of the claims. The full manuscript (Sections 4 and 5) supplies the missing information: datasets (GSM8K, MATH, and OOD variants), baselines (standard RLHF, PRM-based methods), metrics (accuracy, reasoning trace quality, OOD transfer), ablations isolating the hint-injection and pairwise components, and quantitative results showing the reported gains. To improve verifiability at the abstract level, we will revise the abstract to incorporate concise quantitative highlights (e.g., percentage improvements and OOD scores) while preserving brevity. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation self-contained in design choices

full rationale

The provided abstract and description introduce HIPPO as a framework using hint injection to create anchors for pairwise reward modeling. This is presented as an engineering solution to shortcut behaviors rather than a derivation that reduces to fitted parameters, self-definitions, or self-citation chains. No equations, uniqueness theorems, or renamings of known results are shown that collapse by construction. The central claim of extracting authentic reasoning rests on empirical generalization, which is externally falsifiable and not internally forced. Full manuscript not supplied, but nothing in visible text triggers any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no explicit free parameters, axioms, or invented entities; all details are at the level of high-level description.

pith-pipeline@v0.9.1-grok · 5686 in / 1051 out tokens · 27989 ms · 2026-06-30T07:33:28.006194+00:00 · methodology

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

114 extracted references · 18 canonical work pages · 11 internal anchors

  1. [1]

    ReTool: Reinforcement Learning for Strategic Tool Use in LLMs

    Feng, Jiazhan and Huang, Shijue and Qu, Xingwei and Zhang, Ge and Qin, Yujia and Zhong, Baoquan and Jiang, Chengquan and Chi, Jinxin and Zhong, Wanjun , year =. doi:10.48550/arXiv.2504.11536 , urldate =. arXiv , langid =:2504.11536 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2504.11536

  2. [2]

    RAG-RL: Advancing Retrieval-Augmented Generation via RL and Curriculum Learning

    Huang, Jerry and Madala, Siddarth and Sidhu, Risham and Niu, Cheng and Hockenmaier, Julia and Zhang, Tong , year =. doi:10.48550/arXiv.2503.12759 , urldate =. arXiv , annotation =:2503.12759 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2503.12759

  3. [3]

    Jin, Bowen and Zeng, Hansi and Yue, Zhenrui and Wang, Dong and Zamani, Hamed and Han, Jiawei , year =. Search-. doi:10.48550/arXiv.2503.09516 , urldate =. arXiv , langid =:2503.09516 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2503.09516

  4. [4]

    Song, Huatong and Jiang, Jinhao and Min, Yingqian and Chen, Jie and Chen, Zhipeng and Zhao, Wayne Xin and Fang, Lei and Wen, Ji-Rong , year =. R1-. doi:10.48550/arXiv.2503.05592 , urldate =. arXiv , langid =:2503.05592 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2503.05592

  5. [5]

    ZeroSearch: Incentivize the Search Capability of LLMs without Searching

    Sun, Hao and Qiao, Zile and Guo, Jiayan and Fan, Xuanbo and Hou, Yingyan and Jiang, Yong and Xie, Pengjun and Huang, Fei and Zhang, Yan , year =. doi:10.48550/arXiv.2505.04588 , urldate =. arXiv , langid =:2505.04588 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2505.04588

  6. [6]

    Wang, Jiaan and Meng, Fandong and Zhou, Jie , year =. Deep. doi:10.48550/arXiv.2504.10187 , urldate =. arXiv , annotation =:2504.10187 , primaryclass =

    work page doi:10.48550/arxiv.2504.10187

  7. [7]

    Xie, Tian and Gao, Zitian and Ren, Qingnan and Luo, Haoming and Hong, Yuqian and Dai, Bryan and Zhou, Joey and Qiu, Kai and Wu, Zhirong and Luo, Chong , year =. Logic-. doi:10.48550/arXiv.2502.14768 , urldate =. arXiv , langid =:2502.14768 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2502.14768

  8. [8]

    Zhou, Ruochen and Xu, Minrui and Chen, Shiqi and Liu, Junteng and Li, Yunqi and Lin, Xinxin and Chen, Zhengyu and He, Junxian , year =. Does. doi:10.48550/arXiv.2507.04391 , urldate =. arXiv , annotation =:2507.04391 , primaryclass =

    work page doi:10.48550/arxiv.2507.04391

  9. [9]

    2025 , eprint=

    ReasonFlux: Hierarchical LLM Reasoning via Scaling Thought Templates , author=. 2025 , eprint=

    2025

  10. [10]

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

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

    work page internal anchor Pith review Pith/arXiv arXiv

  11. [11]

    QwQ-32B: The Power of Scaling RL , url =

    Qwen , month =. QwQ-32B: The Power of Scaling RL , url =

  12. [12]

    Kimi k1.5: Scaling Reinforcement Learning with LLMs

    Kimi k1. 5: Scaling reinforcement learning with llms , author=. arXiv preprint arXiv:2501.12599 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  13. [13]

    2024 , eprint=

    GPT-4o System Card , author=. 2024 , eprint=

    2024

  14. [14]

    2024 , eprint=

    HuatuoGPT-o1, Towards Medical Complex Reasoning with LLMs , author=. 2024 , eprint=

    2024

  15. [15]

    2024 , eprint=

    OpenAI o1 System Card , author=. 2024 , eprint=

    2024

  16. [16]

    ArXiv , year=

    Proximal Policy Optimization Algorithms , author=. ArXiv , year=

  17. [17]

    ArXiv , year=

    REINFORCE++: A Simple and Efficient Approach for Aligning Large Language Models , author=. ArXiv , year=

  18. [18]

    2024 , eprint=

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models , author=. 2024 , eprint=

    2024

  19. [19]

    Open R1: A fully open reproduction of DeepSeek-R1 , url =

    HuggingFace , month =. Open R1: A fully open reproduction of DeepSeek-R1 , url =

  20. [20]

    2025 , howpublished=

    Open-Reasoner-Zero: An Open Source Approach to Scaling Reinforcement Learning on the Base Model , author=. 2025 , howpublished=

    2025

  21. [21]

    Jiayi Pan and Junjie Zhang and Xingyao Wang and Lifan Yuan and Hao Peng and Alane Suhr , title =

  22. [22]

    2025 , eprint=

    Search-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learning , author=. 2025 , eprint=

    2025

  23. [23]

    2025 , eprint=

    R1-Searcher: Incentivizing the Search Capability in LLMs via Reinforcement Learning , author=. 2025 , eprint=

    2025

  24. [24]

    2025 , eprint=

    RAG-RL: Advancing Retrieval-Augmented Generation via RL and Curriculum Learning , author=. 2025 , eprint=

    2025

  25. [25]

    2024 , eprint=

    Towards Consistent Natural-Language Explanations via Explanation-Consistency Finetuning , author=. 2024 , eprint=

    2024

  26. [26]

    2024 , eprint=

    Making Reasoning Matter: Measuring and Improving Faithfulness of Chain-of-Thought Reasoning , author=. 2024 , eprint=

    2024

  27. [27]

    2025 , eprint=

    Truthful or Fabricated? Using Causal Attribution to Mitigate Reward Hacking in Explanations , author=. 2025 , eprint=

    2025

  28. [28]

    2025 , eprint=

    Bias-Augmented Consistency Training Reduces Biased Reasoning in Chain-of-Thought , author=. 2025 , eprint=

    2025

  29. [29]

    2025 , eprint=

    Chain-of-Probe: Examining the Necessity and Accuracy of CoT Step-by-Step , author=. 2025 , eprint=

    2025

  30. [30]

    2024 , eprint=

    Prover-Verifier Games improve legibility of LLM outputs , author=. 2024 , eprint=

    2024

  31. [31]

    2025 , eprint=

    Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation , author=. 2025 , eprint=

    2025

  32. [32]

    2025 , eprint=

    Reasoning Models Don't Always Say What They Think , author=. 2025 , eprint=

    2025

  33. [33]

    2025 , eprint=

    Frontier Models are Capable of In-context Scheming , author=. 2025 , eprint=

    2025

  34. [34]

    2025 , eprint=

    Chain-of-Thought Reasoning In The Wild Is Not Always Faithful , author=. 2025 , eprint=

    2025

  35. [35]

    2025 , eprint=

    FineMedLM-o1: Enhancing Medical Knowledge Reasoning Ability of LLM from Supervised Fine-Tuning to Test-Time Training , author=. 2025 , eprint=

    2025

  36. [36]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    JEC-QA: a legal-domain question answering dataset , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  37. [37]

    2025 , eprint=

    Crossing the Reward Bridge: Expanding RL with Verifiable Rewards Across Diverse Domains , author=. 2025 , eprint=

    2025

  38. [38]

    Reward Hacking in Reinforcement Learning

    Weng, Lilian. Reward Hacking in Reinforcement Learning. lilianweng.github.io. 2024

    2024

  39. [39]

    POLARIS: A Post-Training Recipe for Scaling Reinforcement Learning on Advanced Reasoning Models , url =

    An, Chenxin and Xie, Zhihui and Li, Xiaonan and Li, Lei and Zhang, Jun and Gong, Shansan and Zhong, Ming and Xu, Jingjing and Qiu, Xipeng and Wang, Mingxuan and Kong, Lingpeng , year =. POLARIS: A Post-Training Recipe for Scaling Reinforcement Learning on Advanced Reasoning Models , url =

  40. [40]

    2025 , eprint=

    LIMR: Less is More for RL Scaling , author=. 2025 , eprint=

    2025

  41. [41]

    2023 , eprint=

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models , author=. 2023 , eprint=

    2023

  42. [42]

    Case method in american legal education: Its origins and objectives, the , author=. J. Legal Educ. , volume=. 1951 , publisher=

    1951

  43. [43]

    2020 , eprint=

    What Disease does this Patient Have? A Large-scale Open Domain Question Answering Dataset from Medical Exams , author=. 2020 , eprint=

    2020

  44. [44]

    Advances in Neural Information Processing Systems , volume=

    Lexeval: A comprehensive chinese legal benchmark for evaluating large language models , author=. Advances in Neural Information Processing Systems , volume=

  45. [45]

    2025 , eprint=

    Qwen2.5 Technical Report , author=. 2025 , eprint=

    2025

  46. [46]

    ArXiv , year=

    DeepSeek-V3 Technical Report , author=. ArXiv , year=

  47. [47]

    2025 , eprint=

    A Survey on LLM-as-a-Judge , author=. 2025 , eprint=

    2025

  48. [48]

    HybridFlow: A flexible and efficient RLHF framework

    Sheng, Guangming and Zhang, Chi and Ye, Zilingfeng and Wu, Xibin and Zhang, Wang and Zhang, Ru and Peng, Yanghua and Lin, Haibin and Wu, Chuan , year=. HybridFlow: A Flexible and Efficient RLHF Framework , url=. doi:10.1145/3689031.3696075 , booktitle=

    work page doi:10.1145/3689031.3696075

  49. [49]

    Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles , year=

    Efficient Memory Management for Large Language Model Serving with PagedAttention , author=. Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles , year=

  50. [50]

    2025 , eprint=

    GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models , author=. 2025 , eprint=

    2025

  51. [51]

    2025 , eprint=

    CodeBoost: Boosting Code LLMs by Squeezing Knowledge from Code Snippets with RL , author=. 2025 , eprint=

    2025

  52. [52]

    2025 , eprint=

    Posterior-GRPO: Rewarding Reasoning Processes in Code Generation , author=. 2025 , eprint=

    2025

  53. [53]

    2025 , eprint=

    Real-Time Detection of Hallucinated Entities in Long-Form Generation , author=. 2025 , eprint=

    2025

  54. [54]

    2025 , eprint=

    TRIDENT: Benchmarking LLM Safety in Finance, Medicine, and Law , author=. 2025 , eprint=

    2025

  55. [55]

    2025 , eprint=

    AGENTICT ^2 S:Robust Text-to-SPARQL via Agentic Collaborative Reasoning over Heterogeneous Knowledge Graphs for the Circular Economy , author=. 2025 , eprint=

    2025

  56. [56]

    2024 , eprint=

    Faithful Logical Reasoning via Symbolic Chain-of-Thought , author=. 2024 , eprint=

    2024

  57. [57]

    2025 , eprint=

    Reasoning-Table: Exploring Reinforcement Learning for Table Reasoning , author=. 2025 , eprint=

    2025

  58. [58]

    2025 , eprint=

    Table-R1: Inference-Time Scaling for Table Reasoning , author=. 2025 , eprint=

    2025

  59. [59]

    2025 , eprint=

    MedEthicsQA: A Comprehensive Question Answering Benchmark for Medical Ethics Evaluation of LLMs , author=. 2025 , eprint=

    2025

  60. [60]

    2025 , eprint=

    Select to Know: An Internal-External Knowledge Self-Selection Framework for Domain-Specific Question Answering , author=. 2025 , eprint=

    2025

  61. [61]

    2019 , eprint=

    BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding , author=. 2019 , eprint=

    2019

  62. [62]

    2019 , eprint=

    RoBERTa: A Robustly Optimized BERT Pretraining Approach , author=. 2019 , eprint=

    2019

  63. [63]

    2019 , eprint=

    PubMedQA: A Dataset for Biomedical Research Question Answering , author=. 2019 , eprint=

    2019

  64. [64]

    2025 , eprint=

    DAPO: An Open-Source LLM Reinforcement Learning System at Scale , author=. 2025 , eprint=

    2025

  65. [65]

    2025 , eprint=

    Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model? , author=. 2025 , eprint=

    2025

  66. [66]

    2026 , eprint=

    Detecting Data Contamination from Reinforcement Learning Post-training for Large Language Models , author=. 2026 , eprint=

    2026

  67. [67]

    2026 , eprint=

    On The Fragility of Benchmark Contamination Detection in Reasoning Models , author=. 2026 , eprint=

    2026

  68. [68]

    2025 , eprint=

    Reasoning or Memorization? Unreliable Results of Reinforcement Learning Due to Data Contamination , author=. 2025 , eprint=

    2025

  69. [69]

    2026 , eprint=

    Spurious Rewards Paradox: Mechanistically Understanding How RLVR Activates Memorization Shortcuts in LLMs , author=. 2026 , eprint=

    2026

  70. [70]

    2025 , eprint=

    CLARity: Reasoning Consistency Alone Can Teach Reinforced Experts , author=. 2025 , eprint=

    2025

  71. [71]

    2026 , eprint=

    RL-PLUS: Countering Capability Boundary Collapse of LLMs in Reinforcement Learning with Hybrid-policy Optimization , author=. 2026 , eprint=

    2026

  72. [72]

    2025 , eprint=

    Echo Chamber: RL Post-training Amplifies Behaviors Learned in Pretraining , author=. 2025 , eprint=

    2025

  73. [73]

    IEEE Transactions on Information Theory , volume=

    Estimating divergence functionals and the likelihood ratio by convex risk minimization , author=. IEEE Transactions on Information Theory , volume=

  74. [74]

    2024 , eprint=

    A Minimaximalist Approach to Reinforcement Learning from Human Feedback , author=. 2024 , eprint=

    2024

  75. [75]

    2026 , eprint=

    Gained in Translation: Privileged Pairwise Judges Enhance Multilingual Reasoning , author=. 2026 , eprint=

    2026

  76. [76]

    2025 , eprint=

    Investigating Non-Transitivity in LLM-as-a-Judge , author=. 2025 , eprint=

    2025

  77. [77]

    2026 , eprint=

    Coupled Variational Reinforcement Learning for Language Model General Reasoning , author=. 2026 , eprint=

    2026

  78. [78]

    2025 , eprint=

    Are DeepSeek R1 And Other Reasoning Models More Faithful? , author=. 2025 , eprint=

    2025

  79. [79]

    2026 , eprint=

    Lie to Me: How Faithful Is Chain-of-Thought Reasoning in Reasoning Models? , author=. 2026 , eprint=

    2026

  80. [80]

    2025 , eprint=

    Unspoken Hints: Accuracy Without Acknowledgement in LLM Reasoning , author=. 2025 , eprint=

    2025

Showing first 80 references.