pith. sign in

arxiv: 2605.17958 · v1 · pith:OQPXG764new · submitted 2026-05-18 · 💻 cs.LG · cs.PL

Enhancing the Code Reasoning Capabilities of LLMs via Consistency-based Reinforcement Learning

Zhanyue Qin , Jia Feng , Yibo Lyu , Yun Peng , Dianbo Sui , Cuiyun Gao , Qing Liao This is my paper

Pith reviewed 2026-05-20 13:09 UTC · model grok-4.3

classification 💻 cs.LG cs.PL
keywords code reasoningreinforcement learninglarge language modelsconsistency rewardstepwise reasoningreward hackingprogram output predictiondynamic beam sampling
0
0 comments X

The pith

Consistency checks across reasoning steps improve how well LLMs predict program outputs.

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

The paper seeks to establish that standard reinforcement learning for code reasoning produces weak training signals because rewards depend only on final answers or loose cues. This leads to sparse feedback and models that exploit shortcuts instead of learning proper execution logic. CodeThinker counters this by synthesizing data that records consistent intermediate steps from code and inputs to the predicted output, then rewards the model for keeping those steps aligned. The resulting training process yields higher accuracy on code reasoning benchmarks and carries over to math and code generation tasks without further tuning. A sympathetic reader would see this as a way to make RL rewards denser and more faithful to the actual reasoning chain.

Core claim

CodeThinker is a consistency-driven reinforcement learning framework with three parts: a stepwise reasoning-aware training module that uses a consistency tracing paradigm to synthesize data capturing the step-by-step process, a dynamic beam sampling strategy that improves output quality under fixed budgets, and a consistency reward that reduces hacking. When applied to multiple base models, the framework reaches state-of-the-art accuracy on code reasoning benchmarks and delivers average gains of 5.33 points on mathematical reasoning and 3.11 points on code tasks spanning 17 languages, all without additional training.

What carries the argument

The consistency tracing paradigm, which acts as a template to generate training examples that record aligned intermediate reasoning steps for predicting a program's output given its code and inputs.

Load-bearing premise

The consistency tracing method produces synthetic training data that correctly reflects the true step-by-step reasoning process without adding systematic errors or biases.

What would settle it

A clear test would be to check whether models trained under the consistency reward show no accuracy gain, or even lower accuracy, on a fresh set of programs where each intermediate reasoning step can be verified independently against actual execution traces.

Figures

Figures reproduced from arXiv: 2605.17958 by Cuiyun Gao, Dianbo Sui, Jia Feng, Qing Liao, Yibo Lyu, Yun Peng, Zhanyue Qin.

Figure 1
Figure 1. Figure 1: (A) An example of code reasoning task. (B) An example of code reasoning using vanilla [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Workflow of the CodeThinker framework. Consistency Tracing (ConTra) paradigm ( [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: (A): Changes in the model’s depth of thinking during the CodeThinker training. (B): [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Changes in sample scores and standard deviations of scores during CodeThinker and [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Results of different regular sampling and dynamic beam sampling ratios. To investigate the impact of the ratio between regular sampling and dynamic beam sampling within CodeThinker, we compare various sam￾pling combinations. The results are presented in [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Cosine similarity of code in the LeetCodeReasoning dataset with code in three mainstream [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: (A): Changes in the model’s depth of thinking during the CodeThinker training. (B): [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: (A): Changes in the model’s depth of thinking during the CodeThinker training. (B): [PITH_FULL_IMAGE:figures/full_fig_p021_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: (A): Changes in the model’s depth of thinking during the CodeThinker training. (B): [PITH_FULL_IMAGE:figures/full_fig_p021_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: (A): Changes in the model’s depth of thinking during the CodeThinker training. (B): [PITH_FULL_IMAGE:figures/full_fig_p021_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: A sample of code in dataset proposed by Tang et al. [32]. [PITH_FULL_IMAGE:figures/full_fig_p026_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Case 1: a CodeReasoner GRPO rollout. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Case 2: an ExecVerify GRPO rollout. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Case 3: a CodeThinker GRPO rollout. 28 [PITH_FULL_IMAGE:figures/full_fig_p028_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: The prompt used in Unit Test Generation of the pipeline. [PITH_FULL_IMAGE:figures/full_fig_p029_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: The prompt used to extract CoT of ConTra during the SFT phase. [PITH_FULL_IMAGE:figures/full_fig_p030_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: The prompt used to extract CoT of ConTra during the SFT phase. [PITH_FULL_IMAGE:figures/full_fig_p031_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: An example of the code reasoning of CodeThinker. [PITH_FULL_IMAGE:figures/full_fig_p032_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: An example of the code reasoning of CodeIO. [PITH_FULL_IMAGE:figures/full_fig_p033_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: An example of the code reasoning of SemCoder. [PITH_FULL_IMAGE:figures/full_fig_p034_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: An example of extracting vanilla tracing. [PITH_FULL_IMAGE:figures/full_fig_p035_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: An example of the code reasoning of CodeReasoner. [PITH_FULL_IMAGE:figures/full_fig_p036_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: An example of the code reasoning of ExecVerify. [PITH_FULL_IMAGE:figures/full_fig_p037_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: An example of the code reasoning of CWM. [PITH_FULL_IMAGE:figures/full_fig_p038_24.png] view at source ↗
read the original abstract

Code reasoning refers to the task of predicting the output of a program given its source code and specific inputs. It can measure the reasoning capability of large language models (LLMs) and also benefit downstream tasks such as code generation and mathematical reasoning. Existing work has verified the effectiveness of reinforcement learning on the task. However, these methods design rewards solely based on final outputs or coarse-grained signals, and neglect the inherent consistency of the stepwise reasoning process in the task. Therefore, these methods often result in sparse reward or reward hacking, which limits the full play of enhanced learning capabilities. To alleviate these issues, we propose CodeThinker, a consistency-driven reinforcement learning framework for code reasoning. Specifically, CodeThinker has three key components: (1) a stepwise reasoning-aware model training module, which utilizes a consistency tracing paradigm as a template to synthesize training data that captures the stepwise reasoning process; (2) a dynamic beam sampling strategy, which aims to improve the quality of sampled outputs under a fixed sampling budget; and (3) a consistency reward mechanism that can effectively alleviate reward hacking. Experiments on three popular benchmarks show that CodeThinker achieves state-of-the-art performance across multiple LLMs. For instance, it outperforms the strongest baseline by 4.3% in accuracy when deployed on Qwen2.5-Coder-7B-Instruct. We also validate the effectiveness of CodeThinker on downstream tasks. Results show that, without additional training, CodeThinker obtains average accuracy gains of 5.33 and 3.11 percentage points on mathematical reasoning and code reasoning tasks covering 17 programming languages, respectively.

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

3 major / 2 minor

Summary. The manuscript proposes CodeThinker, a consistency-driven reinforcement learning framework to enhance LLMs' code reasoning. It comprises three components: (1) a stepwise reasoning-aware training module that uses a consistency tracing paradigm to synthesize training trajectories capturing intermediate reasoning steps; (2) a dynamic beam sampling strategy to improve output quality under fixed compute; and (3) a consistency reward mechanism intended to mitigate sparse rewards and reward hacking. Experiments on three benchmarks report state-of-the-art results across multiple LLMs, including a 4.3% accuracy gain over the strongest baseline on Qwen2.5-Coder-7B-Instruct, plus downstream gains of 5.33 and 3.11 percentage points on mathematical reasoning and multi-language code tasks without further training.

Significance. If the reported gains prove robust and the consistency signal is shown to be independent of base-model biases, the work would meaningfully advance RL-based reasoning methods by moving beyond final-answer rewards. The downstream transfer results without additional training would be a notable strength, as would any machine-checked or reproducible elements in the experimental pipeline.

major comments (3)
  1. [§3.1] §3.1 (consistency tracing paradigm): The synthesis procedure is described as prompting an LLM to generate stepwise trajectories followed by consistency filtering, but no ablation or diagnostic experiment is presented to test whether the resulting trajectories are independent of the generator model's own reasoning artifacts. This directly bears on whether the consistency reward in component (3) supplies a corrective signal or merely reinforces self-consistent but potentially incorrect patterns, undermining the central claim that the 4.3% gain reflects genuine capability improvement rather than self-reinforcement.
  2. [§4] §4 (experimental results): The 4.3% accuracy improvement on Qwen2.5-Coder-7B-Instruct and the downstream gains are reported without specifying the exact set of baselines, number of random seeds or evaluation runs, data splits, or statistical significance tests. These omissions make it impossible to assess whether the gains are load-bearing evidence for the framework or could be explained by variance or baseline selection.
  3. [§3.3] §3.3 (consistency reward): The reward is defined to penalize inconsistency across sampled trajectories, yet the manuscript provides no analysis of how this interacts with the dynamic beam sampling in §3.2 under the same sampling budget; if the two components are correlated by construction, the claimed alleviation of reward hacking may be overstated.
minor comments (2)
  1. [§3] Notation for the consistency score and beam width parameters is introduced without a consolidated table of symbols, making it harder to follow the reward formulation across sections.
  2. [Figure 2] Figure 2 (training curves) lacks error bars or shading for multiple runs, reducing clarity on stability of the reported improvements.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment point by point below with honest responses, proposing specific revisions to the manuscript where the concerns are valid.

read point-by-point responses

  1. Referee: [§3.1] §3.1 (consistency tracing paradigm): The synthesis procedure is described as prompting an LLM to generate stepwise trajectories followed by consistency filtering, but no ablation or diagnostic experiment is presented to test whether the resulting trajectories are independent of the generator model's own reasoning artifacts. This directly bears on whether the consistency reward in component (3) supplies a corrective signal or merely reinforces self-consistent but potentially incorrect patterns, undermining the central claim that the 4.3% gain reflects genuine capability improvement rather than self-reinforcement.

    Authors: We agree this diagnostic is important for validating the core claim. The consistency tracing paradigm uses multiple independent generations followed by filtering to favor trajectories with internal agreement, which we designed to reduce error reinforcement. However, without an explicit test against generator artifacts, the distinction remains unproven. In the revised manuscript we will add a new ablation subsection that (i) regenerates trajectories using a different base LLM, (ii) measures overlap and step-level accuracy against a small set of human-annotated reasoning traces, and (iii) compares downstream RL performance with and without cross-model filtering. This will directly test whether the consistency signal supplies a corrective effect beyond self-reinforcement. revision: yes

  2. Referee: [§4] §4 (experimental results): The 4.3% accuracy improvement on Qwen2.5-Coder-7B-Instruct and the downstream gains are reported without specifying the exact set of baselines, number of random seeds or evaluation runs, data splits, or statistical significance tests. These omissions make it impossible to assess whether the gains are load-bearing evidence for the framework or could be explained by variance or baseline selection.

    Authors: We acknowledge that the current experimental reporting lacks the necessary rigor and transparency. In the revised version we will expand Section 4 and add an appendix that explicitly lists every baseline with citations and hyper-parameters, reports all main results as means and standard deviations over three independent random seeds, details the precise train/validation/test splits for each benchmark, and includes statistical significance tests (paired t-tests on per-instance accuracy) with p-values for the key comparisons. These additions will allow readers to evaluate the robustness of the reported gains. revision: yes

  3. Referee: [§3.3] §3.3 (consistency reward): The reward is defined to penalize inconsistency across sampled trajectories, yet the manuscript provides no analysis of how this interacts with the dynamic beam sampling in §3.2 under the same sampling budget; if the two components are correlated by construction, the claimed alleviation of reward hacking may be overstated.

    Authors: We recognize that the interaction between dynamic beam sampling and the consistency reward was not analyzed, leaving open the possibility that their benefits are not fully independent. Although the two components address different stages (exploration efficiency versus reward density), we will add a dedicated analysis subsection in the revision. This will include controlled experiments that fix the sampling budget while toggling dynamic versus uniform beam sampling, then measure reward hacking indicators (e.g., divergence between consistency reward and final-answer correctness) and final task accuracy. The results will clarify whether the claimed mitigation of reward hacking holds when the sampling strategy is held constant. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical RL framework evaluated on external benchmarks

full rationale

The paper presents CodeThinker as a three-component empirical method (consistency tracing for data synthesis, dynamic beam sampling, consistency reward) whose central claims are validated through accuracy gains on three independent benchmarks and downstream tasks across 17 languages. No equations, derivations, or first-principles results are shown to reduce by construction to fitted parameters or self-generated inputs; the consistency reward is applied to improve upon baseline RL rather than tautologically re-deriving the synthesis process. External benchmark evaluation provides independent falsifiability, keeping the work self-contained without load-bearing self-citation chains or renamed empirical patterns.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Only the abstract is available, so the ledger is necessarily incomplete. The method rests on the domain assumption that consistency signals in reasoning steps provide a reliable training signal beyond final-answer correctness.

free parameters (1)
  • RL training hyperparameters (learning rate, beam width, reward scaling)
    Standard in any RL setup; values not reported in abstract but required for the training module and sampling strategy.
axioms (1)
  • domain assumption Consistency across reasoning steps is a valid proxy for genuine reasoning quality and reduces reward hacking
    Invoked to justify the consistency reward mechanism and the claim that it alleviates sparse-reward and hacking problems.

pith-pipeline@v0.9.0 · 5842 in / 1344 out tokens · 54708 ms · 2026-05-20T13:09:25.121915+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

69 extracted references · 69 canonical work pages · 21 internal anchors

  1. [1]

    Aho, Monica S

    Alfred V . Aho, Monica S. Lam, Ravi Sethi, and Jeffrey D. Ullman.Compilers: Principles, Techniques, and Tools (2nd Edition). Addison-Wesley Longman Publishing Co., Inc., USA,

    work page

  2. [2]

    Program Synthesis with Large Language Models

    Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie Cai, Michael Terry, Quoc Le, and Charles Sutton. Program synthesis with large language models, 2021. URLhttps://arxiv.org/abs/2108.07732

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  3. [3]

    Ashok, and Shashank Shet

    Ramakrishna Bairi, Atharv Sonwane, Aditya Kanade, Vageesh D C, Arun Iyer, Suresh Parthasarathy, Sriram Rajamani, B. Ashok, and Shashank Shet. Codeplan: Repository-level coding using llms and planning, 2023. URLhttps://arxiv.org/abs/2309.12499

    work page arXiv 2023

  4. [4]

    Reasoning runtime behavior of a program with llm: How far are we?, 2024

    Junkai Chen, Zhiyuan Pan, Xing Hu, Zhenhao Li, Ge Li, and Xin Xia. Reasoning runtime behavior of a program with llm: How far are we?, 2024. URL https://arxiv.org/abs/ 2403.16437

    work page arXiv 2024

  5. [5]

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian...

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  6. [6]

    Training Verifiers to Solve Math Word Problems

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. Training verifiers to solve math word problems, 2021. URL https://arxiv.org/ abs/2110.14168

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  7. [7]

    Sycophancy to Subterfuge: Investigating Reward-Tampering in Large Language Models

    Carson Denison, Monte MacDiarmid, Fazl Barez, David Duvenaud, Shauna Kravec, Samuel Marks, Nicholas Schiefer, Ryan Soklaski, Alex Tamkin, Jared Kaplan, Buck Shlegeris, Samuel R. Bowman, Ethan Perez, and Evan Hubinger. Sycophancy to subterfuge: Inves- tigating reward-tampering in large language models, 2024. URL https://arxiv.org/abs/ 2406.10162

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  8. [8]

    Min, Gail Kaiser, Junfeng Yang, and Baishakhi Ray

    Yangruibo Ding, Jinjun Peng, Marcus J. Min, Gail Kaiser, Junfeng Yang, and Baishakhi Ray. Semcoder: Training code language models with comprehensive semantics reasoning, 2024. URLhttps://arxiv.org/abs/2406.01006

    work page arXiv 2024

  9. [9]

    Abstract syntax and variable binding

    Marcelo Fiore, Gordon Plotkin, and Daniele Turi. Abstract syntax and variable binding. In Proceedings of the 14th Annual IEEE Symposium on Logic in Computer Science, LICS ’99, page 193, USA, 1999. IEEE Computer Society. ISBN 0769501583

    work page 1999

  10. [10]

    PAL: Program-aided Language Models

    Luyu Gao, Aman Madaan, Shuyan Zhou, Uri Alon, Pengfei Liu, Yiming Yang, Jamie Callan, and Graham Neubig. Pal: Program-aided language models, 2023. URL https://arxiv.org/ abs/2211.10435

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  11. [11]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ah- mad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston Zhang, Aurelien Rodriguez, Austen Gregerson, Ava...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  12. [12]

    Bowman, and Evan Hubinger

    Ryan Greenblatt, Carson Denison, Benjamin Wright, Fabien Roger, Monte MacDiarmid, Sam Marks, Johannes Treutlein, Tim Belonax, Jack Chen, David Duvenaud, Akbir Khan, Julian Michael, Sören Mindermann, Ethan Perez, Linda Petrini, Jonathan Uesato, Jared Kaplan, Buck Shlegeris, Samuel R. Bowman, and Evan Hubinger. Alignment faking in large language models,

    work page

  13. [13]

    URLhttps://arxiv.org/abs/2412.14093

    work page internal anchor Pith review Pith/arXiv arXiv

  14. [14]

    Alex Gu, Baptiste Rozière, Hugh Leather, Armando Solar-Lezama, Gabriel Synnaeve, and Sida I. Wang. Cruxeval: A benchmark for code reasoning, understanding and execution, 2024. URLhttps://arxiv.org/abs/2401.03065

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  15. [15]

    Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Y . Wu, Y . K. Li, Fuli Luo, Yingfei Xiong, and Wenfeng Liang. Deepseek-coder: When the large language model meets programming – the rise of code intelligence, 2024. URL https://arxiv.org/abs/2401.14196

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  16. [16]

    Measuring mathematical problem solving with the math dataset,

    Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset,

    work page

  17. [17]

    URLhttps://arxiv.org/abs/2103.03874

    work page internal anchor Pith review Pith/arXiv arXiv

  18. [18]

    LoRA: Low-Rank Adaptation of Large Language Models

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021. URL https://arxiv.org/abs/2106.09685

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  19. [19]

    Qwen2.5-Coder Technical Report

    Binyuan Hui, Jian Yang, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Lei Zhang, Tianyu Liu, Jiajun Zhang, Bowen Yu, Keming Lu, Kai Dang, Yang Fan, Yichang Zhang, An Yang, Rui Men, Fei Huang, Bo Zheng, Yibo Miao, Shanghaoran Quan, Yunlong Feng, Xingzhang Ren, Xuancheng Ren, Jingren Zhou, and Junyang Lin. Qwen2.5-coder technical report, 2024. URL https: //arxiv.org/...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  20. [20]

    LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Code

    Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Ar- mando Solar-Lezama, Koushik Sen, and Ion Stoica. Livecodebench: Holistic and contamination free evaluation of large language models for code, 2024. URL https://arxiv.org/abs/ 2403.07974. 12

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  21. [21]

    Gonzalez, Hao Zhang, and Ion Stoica

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention, 2023. URL https://arxiv.org/abs/2309. 06180

    work page 2023

  22. [22]

    Codei/o: Condensing reasoning patterns via code input-output prediction, 2025

    Junlong Li, Daya Guo, Dejian Yang, Runxin Xu, Yu Wu, and Junxian He. Codei/o: Condensing reasoning patterns via code input-output prediction, 2025. URL https://arxiv.org/abs/ 2502.07316

    work page arXiv 2025

  23. [23]

    Let's Verify Step by Step

    Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s verify step by step, 2023. URL https://arxiv.org/abs/2305.20050

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  24. [24]

    Is Your Code Generated by ChatGPT Really Correct? Rigorous Evaluation of Large Language Models for Code Generation

    Jiawei Liu, Chunqiu Steven Xia, Yuyao Wang, and Lingming Zhang. Is your code generated by chatgpt really correct? rigorous evaluation of large language models for code generation, 2023. URLhttps://arxiv.org/abs/2305.01210

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  25. [25]

    Reposcope: Leveraging call chain-aware multi-view context for repository-level code generation, 2025

    Yang Liu, Li Zhang, Fang Liu, Zhuohang Wang, Donglin Wei, Zhishuo Yang, Kechi Zhang, Jia Li, and Lin Shi. Reposcope: Leveraging call chain-aware multi-view context for repository-level code generation, 2025. URLhttps://arxiv.org/abs/2507.14791

    work page arXiv 2025

  26. [26]

    Understanding R1-Zero-Like Training: A Critical Perspective

    Zichen Liu, Changyu Chen, Wenjun Li, Penghui Qi, Tianyu Pang, Chao Du, Wee Sun Lee, and Min Lin. Understanding r1-zero-like training: A critical perspective, 2025. URL https: //arxiv.org/abs/2503.20783

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  27. [27]

    Natural emergent misalignment from reward hacking in production rl, 2025

    Monte MacDiarmid, Benjamin Wright, Jonathan Uesato, Joe Benton, Jon Kutasov, Sara Price, Naia Bouscal, Sam Bowman, Trenton Bricken, Alex Cloud, Carson Denison, Johannes Gasteiger, Ryan Greenblatt, Jan Leike, Jack Lindsey, Vlad Mikulik, Ethan Perez, Alex Rodrigues, Drake Thomas, Albert Webson, Daniel Ziegler, and Evan Hubinger. Natural emergent misalignmen...

    work page arXiv 2025

  28. [28]

    Thomas J. McCabe. A complexity measure. InProceedings of the 2nd International Conference on Software Engineering, ICSE ’76, page 407, Washington, DC, USA, 1976. IEEE Computer Society Press

    work page 1976

  29. [29]

    Catcoder: Repository-level code generation with relevant code and type context, 2025

    Zhiyuan Pan, Xing Hu, Xin Xia, and Xiaohu Yang. Catcoder: Repository-level code generation with relevant code and type context, 2025. URLhttps://arxiv.org/abs/2406.03283

    work page arXiv 2025

  30. [30]

    Proximal Policy Optimization Algorithms

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms, 2017. URLhttps://arxiv.org/abs/1707.06347

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  31. [31]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y . K. Li, Y . Wu, and Daya Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models, 2024. URL https://arxiv.org/abs/ 2402.03300

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  32. [32]

    ISBN 9798400711961

    Guangming Sheng, Chi Zhang, Zilingfeng Ye, Xibin Wu, Wang Zhang, Ru Zhang, Yanghua Peng, Haibin Lin, and Chuan Wu. Hybridflow: A flexible and efficient rlhf framework. In Proceedings of the Twentieth European Conference on Computer Systems, EuroSys ’25, page 1279–1297. ACM, Mar 2025. doi: 10.1145/3689031.3696075. URL http://dx.doi.org/ 10.1145/3689031.3696075

    work page doi:10.1145/3689031.3696075 2025

  33. [33]

    Codereasoner: Enhancing the code reasoning ability with reinforcement learning, 2025

    Lingxiao Tang, He Ye, Zhongxin Liu, Xiaoxue Ren, and Lingfeng Bao. Codereasoner: Enhancing the code reasoning ability with reinforcement learning, 2025. URL https: //arxiv.org/abs/2507.17548

    work page arXiv 2025

  34. [34]

    Execverify: White-box rl with verifiable stepwise rewards for code execution reasoning,

    Lingxiao Tang, He Ye, Zhaoyang Chu, Muyang Ye, Zhongxin Liu, Xiaoxue Ren, and Lingfeng Bao. Execverify: White-box rl with verifiable stepwise rewards for code execution reasoning,

    work page

  35. [35]

    URLhttps://arxiv.org/abs/2603.11226

    work page arXiv

  36. [36]

    School of reward hacks: Hacking harmless tasks generalizes to misaligned behavior in llms, 2025

    Mia Taylor, James Chua, Jan Betley, Johannes Treutlein, and Owain Evans. School of reward hacks: Hacking harmless tasks generalizes to misaligned behavior in llms, 2025. URL https: //arxiv.org/abs/2508.17511. 13

    work page arXiv 2025

  37. [37]

    FAIR CodeGen team, Jade Copet, Quentin Carbonneaux, Gal Cohen, Jonas Gehring, Jacob Kahn, Jannik Kossen, Felix Kreuk, Emily McMilin, Michel Meyer, Yuxiang Wei, David Zhang, Kunhao Zheng, Jordi Armengol-Estapé, Pedram Bashiri, Maximilian Beck, Pierre Chambon, Abhishek Charnalia, Chris Cummins, Juliette Decugis, Zacharias V . Fisches, François Fleuret, Fabi...

    work page

  38. [38]

    URLhttps://arxiv.org/abs/2510.02387

    work page arXiv

  39. [39]

    When reward hacking rebounds: Understanding and mitigating it with representation-level signals, 2026

    Rui Wu and Ruixiang Tang. When reward hacking rebounds: Understanding and mitigating it with representation-level signals, 2026. URLhttps://arxiv.org/abs/2604.01476

    work page arXiv 2026

  40. [40]

    Leetcodedataset: A temporal dataset for robust evaluation and efficient training of code llms, 2025

    Yunhui Xia, Wei Shen, Yan Wang, Jason Klein Liu, Huifeng Sun, Siyue Wu, Jian Hu, and Xiaolong Xu. Leetcodedataset: A temporal dataset for robust evaluation and efficient training of code llms, 2025. URLhttps://arxiv.org/abs/2504.14655

    work page arXiv 2025

  41. [41]

    Cruxeval-x: A benchmark for multilingual code reasoning, understanding and execution, 2025

    Ruiyang Xu, Jialun Cao, Yaojie Lu, Ming Wen, Hongyu Lin, Xianpei Han, Ben He, Shing-Chi Cheung, and Le Sun. Cruxeval-x: A benchmark for multilingual code reasoning, understanding and execution, 2025. URLhttps://arxiv.org/abs/2408.13001

    work page arXiv 2025

  42. [42]

    Qwen3 Technical Report

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jing Zhou, Jingren Zhou, Junyang Lin, Kai Dang, Keqin Bao, Kexin Yang, ...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  43. [43]

    DAPO: An Open-Source LLM Reinforcement Learning System at Scale

    Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Weinan Dai, Tiantian Fan, Gaohong Liu, Lingjun Liu, Xin Liu, Haibin Lin, Zhiqi Lin, Bole Ma, Guangming Sheng, Yuxuan Tong, Chi Zhang, Mofan Zhang, Wang Zhang, Hang Zhu, Jinhua Zhu, Jiaze Chen, Jiangjie Chen, Chengyi Wang, Hongli Yu, Yuxuan Song, Xiangpeng Wei, Hao Zhou, Jingjing Liu, W...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  44. [44]

    Repocoder: Repository-level code completion through iterative retrieval and generation, 2023

    Fengji Zhang, Bei Chen, Yue Zhang, Jacky Keung, Jin Liu, Daoguang Zan, Yi Mao, Jian-Guang Lou, and Weizhu Chen. Repocoder: Repository-level code completion through iterative retrieval and generation, 2023. URLhttps://arxiv.org/abs/2303.12570

    work page arXiv 2023

  45. [45]

    LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models

    Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. Llamafactory: Unified efficient fine-tuning of 100+ language models, 2024. URLhttps://arxiv.org/abs/2403.13372. 14 A Related Work A.1 Evaluating LLMs on Code Execution Reasoning Task CRUXEval [13]CRUXEval introduces 800 short, self-contained Python functions...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  46. [46]

    Function f is called with s ='abc'and t ='a'

    work page

  47. [47]

    Calculate lengths: n = len(s) = 3, m = len(t) = 1

    work page

  48. [48]

    Initialize indices: i = 0, j = 0

    work page

  49. [49]

    Initialize rem = False

    work page

  50. [50]

    Enter the while loop (i < n and j < m): - First iteration: - i = 0, j = 0 - Check if s[i] != t[j]:'a'!='a'-> False - Since they are equal, execute`else`block: j += 1 -> j = 1 - i += 1 -> i = 1 - Second iteration: - Check if i < n (1 < 3) and j < m (1 < 1) -> False - Exit the loop

    work page

  51. [51]

    [/THOUGHT] [ANSWER] assert f('abc','a') == 1 [/ANSWER] Figure 12: Case 1: a CodeReasoner GRPO rollout

    Return j, which is 1. [/THOUGHT] [ANSWER] assert f('abc','a') == 1 [/ANSWER] Figure 12: Case 1: a CodeReasoner GRPO rollout. 26 Case 2: an ExecVerify GRPO rollout. Input (abridged): [SYSTEM_PROMPT] ... [/SYSTEM_PROMPT] [INSTRUCTION] ... [/INSTRUCTION] Here is the code content: 1 def f(money, children): 2 if money < children: 3 return -1 4 if money > 8 * c...

    work page

  52. [52]

    - Include index 0 in the result list

    **First iteration (i=0, w='abc')**: - Check if'a'in'abc'-> True. - Include index 0 in the result list

    work page

  53. [53]

    - Exclude index 1

    **Second iteration (i=1, w='bcd')**: - Check if'a'in'bcd'-> False. - Exclude index 1

    work page

  54. [54]

    - Include index 2 in the result list

    **Third iteration (i=2, w='aaaa')**: - Check if'a'in'aaaa'-> True. - Include index 2 in the result list

    work page

  55. [55]

    # THINKING

    **Fourth iteration (i=3, w='cbc')**: - Check if'a'in'cbc'-> False. - Exclude index 3. The final result list is [0, 2]. Local variables now include i=3 and w='cbc' (from the last iteration). [/THOUGHT] [LOCALS] {'words': ['abc','bcd','aaaa','cbc'],'x':'a','i': 3,'w':'cbc'} [/LOCALS] [RETURN] assert f(['abc','bcd','aaaa','cbc'],'a') == [0, 2] [/RETURN] [/TR...

    work page

  56. [56]

    # THINKING

    Replace the "# THINKING" marker with your detailed step-by-step execution reasoning

    work page

  57. [57]

    [ANSWER]

    Keep the "[ANSWER]" block exactly unchanged

    work page

  58. [58]

    [COT]" and

    Output ONLY the completed result enclosed in "[COT]" and "[/COT]" tags. Do NOT output any other extra information outside the specified tags. Your reasoning in "[THOUGHT]" should: - Execute the code line by line with the given input values. - Track how variables change at each step. - For loops, trace each iteration explicitly. - For conditionals, explain...

    work page

  59. [59]

    The function f is called with nums = [1, -2, 3, -4, 5, -6]

    work page

  60. [60]

    i = 0, j = 1

    ans = [0] * len(nums) = [0] * 6 = [0, 0, 0, 0, 0, 0]. i = 0, j = 1

    work page

  61. [61]

    ans = [1, 0, 0, 0, 0, 0]

    Loop over each element x in nums: - x = 1: 1 > 0 is True, so ans[0] = 1, i = 2. ans = [1, 0, 0, 0, 0, 0]. - x = -2: -2 > 0 is False, so ans[1] = -2, j = 3. ans = [1, -2, 0, 0, 0, 0]. - x = 3: 3 > 0 is True, so ans[2] = 3, i = 4. ans = [1, -2, 3, 0, 0, 0]. - x = -4: -4 > 0 is False, so ans[3] = -4, j = 5. ans = [1, -2, 3, -4, 0, 0]. - x = 5: 5 > 0 is True,...

    work page

  62. [62]

    The function returns ans = [1, -2, 3, -4, 5, -6]

    The loop ends. The function returns ans = [1, -2, 3, -4, 5, -6]. [/THOUGHT] [ANSWER] assert f([1, -2, 3, -4, 5, -6]) == [1, -2, 3, -4, 5, -6] [/ANSWER] [/COT] Figure 21: An example of extracting vanilla tracing. 35 An example of the code reasoning of CodeReasoner. Input: [SYSTEM_PROMPT] ... [/SYSTEM_PROMPT] [INSTRUCTION] ... [/INSTRUCTION] [PYTHON] def f(...

    work page

  63. [63]

    **Function Call**:`f(['leet','code'],'e')`

    work page

  64. [64]

    **Initialization**: -`words = ['leet','code']` -`x ='e'` -`res = []`

    work page

  65. [65]

    - Append`i`(0) to`res`:`res = [0]`

    **Loop Iteration**: - **First Iteration** (`i=0`,`w='leet'`): - Check if`x`('e') is in`w`('leet'):`'e'in'leet'`evaluates to `True`. - Append`i`(0) to`res`:`res = [0]`. - **Second Iteration** (`i=1`,`w='code'`): - Check if`x`('e') is in`w`('code'):`'e'in'code'`evaluates to `True`. - Append`i`(1) to`res`:`res = [0, 1]`

    work page

  66. [66]

    [ANSWER] assert f(['leet','code'],'e') == [0, 1] [/ANSWER] Figure 22: An example of the code reasoning of CodeReasoner.[correct] 36 An example of the code reasoning of ExecVerify

    **Return Statement**: After processing all elements, return`res = [0, 1]`. [ANSWER] assert f(['leet','code'],'e') == [0, 1] [/ANSWER] Figure 22: An example of the code reasoning of CodeReasoner.[correct] 36 An example of the code reasoning of ExecVerify. Input: [SYSTEM_PROMPT] ... [/SYSTEM_PROMPT] [INSTRUCTION] ... [/INSTRUCTION] [PYTHON] def f(words, x):...

    work page

  67. [67]

    Function`f`is called with`words = ['leet','code']`and`x ='e'`

    work page

  68. [68]

    Since`'e'`is in `'leet'`, append`0`to`l`

    Iterate over`range(len(words))`, which is`range(2)`(indices 0 and 1): - **Iteration 1 (i = 0)**: - Check if`'e'in words[0]`:`words[0]`is`'leet'`. Since`'e'`is in `'leet'`, append`0`to`l`. Now`l = [0]`. - **Iteration 2 (i = 1)**: - Check if`'e'in words[1]`:`words[1]`is`'code'`. Since`'e'`is in `'code'`, append`1`to`l`. Now`l = [0, 1]`

    work page

  69. [69]

    words":

    Return`l`, which is`[0, 1]`. [ANSWER] assert f(['leet','code'],'e') == [0, 1] [/ANSWER] Figure 23: An example of the code reasoning of ExecVerify.[correct] 37 An example of the code reasoning of CWM. Input: <|begin_of_text|><|trace_context_start|> def f(words, x): res = [] for i, w in enumerate(words): if x in w: res.append(i) return res def main(): # << ...

    work page