Pith. sign in

REVIEW 25 cited by

CodeRL: Mastering Code Generation through Pretrained Models and Deep Reinforcement Learning

Not yet reviewed by Pith; the record is open.

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

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2207.01780 v3 pith:V325N4CR submitted 2022-07-05 cs.LG cs.CLcs.PL

Hung Le , Yue Wang , Akhilesh Deepak Gotmare , Silvio Savarese , Steven C.H. Hoi This is my paper · ORCID
classification cs.LGcs.CLcs.PL
keywords generationmodelresultscodelearningpretrainedproblemprogram
verification ladder T0 review T1 audit T2 compute T3 formal

Signed reviews

No signed human review yet.

0 comments
read the original abstract

Program synthesis or code generation aims to generate a program that satisfies a problem specification. Recent approaches using large-scale pretrained language models (LMs) have shown promising results, yet they have some critical limitations. In particular, they often follow a standard supervised fine-tuning procedure to train a code generation model only from the pairs of natural-language problem descriptions and ground-truth programs. Such paradigm largely ignores some important but potentially useful signals in the problem specification such as unit tests, which thus often results in poor performance when solving complex unseen coding tasks. To address the limitations, we propose "CodeRL", a new framework for program synthesis tasks through pretrained LMs and deep reinforcement learning (RL). Specifically, during training, we treat the code-generating LM as an actor network, and introduce a critic network that is trained to predict the functional correctness of generated programs and provide dense feedback signals to the actor. During inference, we introduce a new generation procedure with a critical sampling strategy that allows a model to automatically regenerate programs based on feedback from example unit tests and critic scores. For the model backbones, we extended the encoder-decoder architecture of CodeT5 with enhanced learning objectives, larger model sizes, and better pretraining data. Our method not only achieves new SOTA results on the challenging APPS benchmark, but also shows strong zero-shot transfer capability with new SOTA results on the simpler MBPP benchmark.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 25 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Self-Improving Large Language Models via Progressive Experience Evolution

    cs.CL 2026-08 conditional novelty 7.0 of 10

    SPEE improves LLM math reasoning by explicitly evolving a global pool of transferable experience and distilling it into the policy before applying GRPO reinforcement learning.

  2. Embarrassingly Simple Self-Distillation Improves Code Generation

    cs.CL 2026-04 conditional novelty 7.0 of 10

    Simple self-distillation—fine-tuning a code model on its own temperature-sampled, truncated outputs—raises LiveCodeBench pass@1 substantially without verifiers, teachers, or RL.

  3. Training Language Models to Generate Quality Code with Program Analysis Feedback

    cs.CL 2025-05 conditional novelty 7.0 of 10

    REAL trains LLMs with rewards from program analysis and unit tests, improving joint functionality and code quality over prior methods.

  4. Consilience for Verifier-Free Test-Time Scaling

    cs.CL 2026-08 accept novelty 6.0 of 10

    Selecting LLM rollouts with low early confidence and high final confidence outperforms confidence-maximizing methods on hard math and free-form code benchmarks.

  5. Security Tests as Executable Specifications for LLM Code Generation: Benefits, Trade-offs, and Coverage Limits

    cs.SE 2026-08 conditional novelty 6.0 of 10

    Showing failing security tests to an LLM during repair fixes a subset of insecure candidates, but visible test success leaves a measurable hidden coverage gap.

  6. RLPF: Reinforcement Learning from Performance Feedback for Code Generation

    cs.LG 2026-07 conditional novelty 6.0 of 10

    RLPF's staged performance reward lifts Qwen3-32B on PerfCodeBench from 11.1% to 54.6% correct-and-runnable and from 8.1% to 38.6% relative efficiency.

  7. NKI-Agent: Domain-Specific Fine-Tuning and Agentic Tool Use for Neuron Kernel Generation

    cs.LG 2026-07 conditional novelty 6.0 of 10

    An agent with compile/verify tools reaches 77.3% NKI kernel pass rate on real Trn1 hardware with Opus 4.8, versus 6% single-shot; SFT Qwen3-30B hits 25% at ~1/100th cost, and binary-reward GRPO fails to beat SFT.

  8. Re:Form -- Reducing Human Annotations in Scalable Formal Software Verification with RL in LLMs: A Preliminary Study on Dafny

    cs.CL 2025-07 conditional novelty 6.0 of 10

    A minimal-prior pipeline with automated data curation and verifier-driven RL lets small LLMs generate verifiable Dafny specifications and beat larger proprietary models on a synthetic compositional benchmark.

  9. Dr. Boot: Bootstrapping Program Synthesis Language Models to Perform Repairing

    cs.SE 2025-07 conditional novelty 6.0 of 10

    A bootstrapping algorithm that trains program synthesis models on their own successful outputs and on repaired failures modestly improves pass@k over regular fine-tuning on MBPP, with mixed results on APPS.

  10. Are They All Good? Evaluating the Quality of CoTs in LLM-based Code Generation

    cs.SE 2025-07 reject novelty 6.0 of 10

    An empirical study of 1,023 CoT-code pairs shows that 76.4% of LLM-generated CoTs are low quality and that CoT correctness does not guarantee code correctness.

  11. AutoTriton: Automatic Triton Programming with Reinforcement Learning in LLMs

    cs.LG 2025-07 conditional novelty 6.0 of 10

    An 8B model trained with SFT and GRPO reinforcement learning generates Triton kernels that match or exceed frontier model performance on TritonBench and KernelBench correctness metrics.

  12. Integrating Symbolic Execution into the Fine-Tuning of Code-Generating LLMs

    cs.SE 2025-04 conditional novelty 6.0 of 10

    Adding symbolic-execution-generated tests to APPS improves reward model accuracy for code generation, but barely changes final code-generation quality relative to CodeRL.

  13. ACECODER: Acing Coder RL via Automated Test-Case Synthesis

    cs.SE 2025-02 conditional novelty 6.0 of 10

    ACECODER shows that automatically synthesized test cases can serve as scalable reward signals for code RL, yielding consistent gains on HumanEval, MBPP, BigCodeBench, and LiveCodeBench.

  14. FullStack Bench: Evaluating LLMs as Full Stack Coders

    cs.AI 2024-11 conditional novelty 6.0 of 10

    FullStack Bench and SandboxFusion offer a broader, execution-based benchmark for code LLMs with 3,374 problems in 16 languages and 11 domains.

  15. Semantically-Aware Rewards for Open-Ended R1 Training in Free-Form Generation

    cs.CL 2025-06 conditional novelty 5.0 of 10

    PrefBERT, a 150M-parameter reward model trained on human quality ratings, outperforms ROUGE-L and BERTScore as a GRPO reward signal for open-ended long-form generation.

  16. Improving LLM-Generated Code Quality with GRPO

    cs.AI 2025-06 conditional novelty 5.0 of 10

    Adding a hand-built code-quality score to the GRPO reward makes a 3B model produce code that blind raters prefer, but evidence rests on one model and a synthetic benchmark.

  17. CRScore++: Reinforcement Learning with Verifiable Tool and AI Feedback for Code Review

    cs.SE 2025-05 conditional novelty 5.0 of 10

    A two-stage SFT+DPO pipeline using linter and code-smell tool outputs plus an LLM judge improves generated code review comments and appears to transfer from Python to Java and JavaScript, though the judge-based evalua...

  18. RLSR: Reinforcement Learning from Self Reward

    cs.LG 2025-05 conditional novelty 5.0 of 10

    A self-judging reinforcement learning loop, where a language model verifies its own integration answers by differentiating them, let a 7B model reach 75.5% on MIT Integration Bee 2025 qualifying problems.

  19. DeepDistill: Enhancing LLM Reasoning Capabilities via Large-Scale Difficulty-Graded Data Training

    cs.CL 2025-04 conditional novelty 5.0 of 10

    A two-stage SFT recipe using pass-rate and coefficient-of-variation based data selection from a 40M-response distilled dataset lifts Qwen2.5-72B to 79.2% on AIME2024, nearly matching RL-trained reasoning models.

  20. Accelerating Reinforcement Learning Algorithms Convergence using Pre-trained Large Language Models as Tutors With Advice Reusing

    cs.LG 2025-09 conditional novelty 4.0 of 10

    LLM tutoring modestly accelerates RL convergence on average, with advice reuse saving wall-clock time but reducing stability.

  21. CRPE: Expanding The Reasoning Capability of Large Language Model for Code Generation

    cs.SE 2025-05 conditional novelty 4.0 of 10

    A three-stage training pipeline (synthetic problems, multi-agent chain-of-thought data, step-DPO self-improvement) raises LiveCodeBench pass@1 for Qwen2.5-Coder bases to 21.88 (7B) and 35.08 (32B).

  22. AM-Thinking-v1: Advancing the Frontier of Reasoning at 32B Scale

    cs.CL 2025-05 conditional novelty 4.0 of 10

    AM-Thinking-v1, a 32B dense model fine-tuned and RL-trained from Qwen2.5-32B, reports AIME 2024 85.3, AIME 2025 74.4, LiveCodeBench 70.3, and Arena-Hard 92.5, beating DeepSeek-R1 on the math and code benchmarks.

  23. Surveying the Effects of Quality, Diversity, and Complexity in Synthetic Data From Large Language Models

    cs.LG 2024-12 conditional novelty 4.0 of 10

    This survey organizes LLM synthetic data research around quality, diversity, and complexity, claiming quality mainly helps in-distribution generalization, diversity mainly helps out-of-distribution generalization, and...

  24. Exploring the Potential of Offline RL for Reasoning in LLMs: A Preliminary Study

    cs.CL 2025-05 conditional novelty 3.0 of 10

    Applying LD-DPO to the DeepDistill-32B model improves average benchmark scores by 3.3 points, but the gains are uneven, the tables are inconsistent, and the evidence is preliminary.

  25. Optimizing AI-Assisted Code Generation

    cs.SE 2024-12 conditional novelty 2.0 of 10

    A literature review and conceptual design for a secure, user-friendly AI code generation platform, with no implementation or experimental validation.

Pith tools