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

• URL: http://arxiv.org/abs/2605.17958v1
• Date: Mon, 18 May 2026 07:12:44 GMT
• Title: Enhancing the Code Reasoning Capabilities of LLMs via Consistency-based Reinforcement Learning
• Abstract summary: Code reasoning refers to the task of predicting the output of a program given its source code and specific inputs.<n>Existing work has verified the effectiveness of reinforcement learning on the task.<n>We propose CodeThinker, a consistency-driven reinforcement learning framework for code reasoning.
• Score: 17.688379880523566
• License: http://creativecommons.org/licenses/by/4.0/
• 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.

Related papers

• The Imitation Game: When LLMs Learn to Reason Like Programs via Code-Centric Reasoning Data Synthesis [50.473217152006875]
  MIMIC is a framework that leverages executable code as a rigorous medium for reasoning data synthesis.<n>MIMIC transforms algorithms into verifiable reasoning trajectories through narrative fusion, code-guided test synthesis, and dynamic code instrumentation.<n>Our method significantly elevates accuracy across general reasoning, complex mathematical benchmarks, and fine-grained deterministic tasks.
  arXiv  Detail & Related papers  (2026-09-13T17:04:16Z)
• When Do Intrinsic Rewards Work for Code Reasoning? A Comprehensive Study [77.64957118012423]
  Reinforcement learning with verifiable rewards (RLVR) has driven substantial progress in large language model reasoning.<n>Recent work addresses this by deriving rewards from a model's own signals, such as majority voting or confidence-based scores.<n>We present a systematic empirical study of intrinsic reward methods for code generation.
  arXiv  Detail & Related papers  (2026-06-18T19:15:50Z)
• What Really Improves Mathematical Reasoning: Structured Reasoning Signals Beyond Pure Code [72.9921566968371]
  We revisit the claim that code improves reasoning through controlled pretraining experiments on a 10T-token corpus with fine-grained domain separation.<n>Code substantially improves programming ability but does not act as a general reasoning enhancer.<n>Cross-domain structured reasoning traces, such as code-text and math-text mixtures, are better explained.
  arXiv  Detail & Related papers  (2026-05-19T12:37:01Z)
• Evaluating and Achieving Controllable Code Completion in Code LLM [89.64782747840225]
  We present the first instruction-guided code completion benchmark, Controllable Code Completion Benchmark (C3-Bench)<n>We reveal substantial gaps in instruction-following capabilities between open-source and advanced proprietary models during code completion tasks.<n>The resulting model, Qwen2.5-Coder-C3, achieves state-of-the-art performance on C3-Bench.
  arXiv  Detail & Related papers  (2026-01-22T11:40:04Z)
• Readability-Robust Code Summarization via Meta Curriculum Learning [53.44612630063336]
  In the real world, code is often poorly structured or obfuscated, significantly degrading model performance.<n>We propose RoFTCodeSum, a novel fine-tuning method that enhances the robustness of code summarization against poorly readable code.
  arXiv  Detail & Related papers  (2026-01-09T02:38:24Z)
• Chain of Execution Supervision Promotes General Reasoning in Large Language Models [48.100128916029064]
  We introduce TracePile, a large-scale corpus of 2.6 million samples that transforms code execution into explicit, step-by-step chain-of-thought-style rationales.<n>We evaluate TracePile using three training setups: continue-pretraining, instruction tuning after pretraining, and two-stage finetuning.<n> Notably, TracePile boosts LLaMA3.1-8B by 7.1% on average across nine math datasets and delivers clear gains on LiveCodeBench, CRUX, and MMLU.
  arXiv  Detail & Related papers  (2025-10-24T02:21:11Z)
• CodeBoost: Boosting Code LLMs by Squeezing Knowledge from Code Snippets with RL [28.43882967593511]
  Code large language models (LLMs) have become indispensable tools for building efficient and automated coding pipelines.<n>Existing models are typically post-trained using reinforcement learning (RL) from general-purpose LLMs using "human instruction-final answer" pairs.<n>We propose CodeBoost, a framework that enhances code LLMs purely from code snippets, without relying on human-annotated instructions.
  arXiv  Detail & Related papers  (2025-08-07T10:31:24Z)
• CodeReasoner: Enhancing the Code Reasoning Ability with Reinforcement Learning [8.197518276987989]
  Code reasoning is a fundamental capability for large language models (LLMs) in the code domain.<n>Prior approaches mainly rely on supervised fine-tuning to improve performance in code reasoning tasks.<n>We argue this is due to two core issues: the low quality of training data and the limitations of supervised fine-tuning.<n>We propose CodeReasoner, a framework that spans both dataset construction and a two-stage training process.
  arXiv  Detail & Related papers  (2025-07-23T14:26:58Z)
• Teaching LLM to Reason: Reinforcement Learning from Algorithmic Problems without Code [76.80306464249217]
  We propose TeaR, which aims at teaching LLMs to reason better.<n>TeaR leverages careful data curation and reinforcement learning to guide models in discovering optimal reasoning paths through code-related tasks.<n>We conduct extensive experiments using two base models and three long-CoT distillation models, with model sizes ranging from 1.5 billion to 32 billion parameters, and across 17 benchmarks spanning Math, Knowledge, Code, and Logical Reasoning.
  arXiv  Detail & Related papers  (2025-07-10T07:34:05Z)
• Which Data Attributes Stimulate Math and Code Reasoning? An Investigation via Influence Functions [8.540135660509058]
  Large language models (LLMs) have demonstrated remarkable reasoning capabilities in math and coding.<n>We leverage influence functions to attribute LLMs' reasoning ability on math and coding to individual training examples, sequences, and tokens.<n>High-difficulty math examples improve both math and code reasoning, while low-difficulty code tasks most effectively benefit code reasoning.
  arXiv  Detail & Related papers  (2025-05-26T13:15:26Z)
• Improving LLM Reasoning through Scaling Inference Computation with Collaborative Verification [52.095460362197336]
  Large language models (LLMs) struggle with consistent and accurate reasoning. LLMs are trained primarily on correct solutions, reducing their ability to detect and learn from errors. We propose a novel collaborative method integrating Chain-of-Thought (CoT) and Program-of-Thought (PoT) solutions for verification.
  arXiv  Detail & Related papers  (2024-10-05T05:21:48Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.