Pith. sign in

REVIEW 2 major objections 4 minor 21 references

DHRCL claims that ordering code-RL rewards by prerequisite—syntax, execution, then tests and structure—with stage durations set from validation trends, raises average Pass@1 to 50.3±0.2 on Qwen3-8B, beating the strongest verifiable dense-re

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-05 04:21 UTC pith:NVAPJTJZ

load-bearing objection A well-run empirical paper with a modest but credible gain; the novel scheduler's statistics are fragile enough that it needs a sensitivity analysis before I'd trust the 0.7-point advantage over fixed-stage. the 2 major comments →

arxiv 2607.26457 v2 pith:NVAPJTJZ submitted 2026-07-29 cs.LG

DHRCL:Training Code LLMs with Dense Hierarchical Rewards and Curriculum Learning

Shuhang Wang , Ziming Li , Hui Cheng This is my paper
classification cs.LG
keywords reinforcement learningcode generationcurriculum learningdense rewardsLLM post-trainingAST structural similaritygroup relative policy optimizationQwen3
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.

Reinforcement learning for code generation usually gives the model one scalar outcome—did the tests pass—or a fixed blend of signals. This paper argues that syntax validity, executability, unit-test pass rate, and program structure are prerequisite skills that should be trained in order, not weighted together from the start. Its method, DHRCL, schedules these four rewards through a three-stage curriculum, decides when to advance stages from validation trends instead of hand-set thresholds, and redistributes each trajectory's gradient across tokens according to how established the tokens are. Under a matched Qwen3-8B and KodCode protocol, DHRCL reports an average Pass@1 of 50.3±0.2, 1.1 points above VeRPO, the strongest baseline, and the advantage persists across Qwen3-4B, 8B, and 14B. If right, it suggests staged, verifiable dense rewards can substitute for learned reward models in code RL.

Core claim

The paper's central claim is that code-generation RL should not present syntax validity, executability, unit-test pass rate, and AST-based structural similarity as one fixed reward mixture. These four signals describe progressively dependent capabilities, so DHRCL trains them in a fixed order—Syntax, then Execution, then Pass & Structural—while keeping the pass-rate reward active throughout. Stage boundaries come from a one-sided linear-trend test on the most recent eight held-out validation observations, with a transition only after two consecutive failed tests, and the best checkpoint of the stage is restored before advancing. Within each stage, trajectory-level advantages are redistribute

What carries the argument

The load-bearing object is the stage-aware hierarchical reward coupled with the curriculum rule. Syntax validation rate is computed by an error-tolerant parser over statement- and block-level units weighted by token length; execution success is a 0/1 indicator; pass rate is the fraction of unit tests passed; AST similarity is tree-edit distance between normalized ASTs of generated and reference programs. Each stage combines pass rate with one primary signal and gates the remaining terms by a batch-level competence coefficient, set to the batch mean of that stage's primary reward. Stage transitions are decided by a one-sided test on the slope of the most recent eight validation observations a

Load-bearing premise

The trend-based stage switch assumes the last eight validation measurements of the stage metric follow a straight line, and that two consecutive failed positive-trend tests reliably mean the metric has plateaued, even though the measurements overlap and are tested repeatedly.

What would settle it

Re-run DHRCL with stage durations set by an oracle or by non-overlapping validation windows with multiplicity-corrected trend tests; if the Pass@1 advantage over VeRPO falls below the reported 95% CI [0.6,1.5], the automatic progression rule, not the hierarchical rewards, is carrying the gain.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Heterogeneous code rewards can be ordered by prerequisite rather than weighted statically, with functional correctness kept active at every stage.
  • Stage duration can be set by validation-trend monitoring instead of manual thresholds; the rule selected median transition iterations of 140/280 on Qwen3-8B, and varied with model scale.
  • Trajectory-level credit can be redistributed by token confidence without an external reward model, lowering the average degenerate group ratio to 0.20.
  • The full DHRCL recipe reaches the GRPO reference validation Pass@1 in about half the iterations (0.49×) and about 55% of the GPU hours.
  • DHRCL reduces syntax errors, runtime errors, and timeouts relative to both the base model and VeRPO; executable-but-incorrect programs remain the dominant residual failure mode.

Where Pith is reading between the lines

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

  • Editorial inference: the AST reward's contribution (+0.6 Pass@1, +0.06 AST similarity) suggests that any reference-conditioned structural prior, such as dataflow graphs or type annotations, could behave similarly, but this is not tested here.
  • Editorial inference: because the margin over VeRPO narrows from 1.3 to 0.7 points as the model grows from 4B to 14B, the curriculum may matter most where the base policy is weakest; larger models might need fewer or shorter stages.
  • Editorial inference: the probability-based credit redistribution recipe is general enough to apply to any RL task with staged verifiable objectives—for example theorem proving or data transformation—though the paper only evaluates code generation.
  • Editorial inference: the trend-based progression rule could be made more robust by change-point detection or multiplicity-corrected tests, and the paper's own overlapping-window design leaves that as an open improvement.

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

2 major / 4 minor

Summary. The paper proposes DHRCL, a reinforcement-learning post-training framework for code LLMs. It decomposes program-level feedback into four dense signals (syntax validity, execution success, unit-test pass rate, and AST-based structural similarity) and arranges them into a three-stage curriculum (Syntax; Execution; Pass & Structural). Stage boundaries are not fixed in advance; they are determined by a trend test on a held-out validation split. The paper also introduces stage-dependent token-level credit reweighting: confidence-based weighting in the syntax stage, uniform weighting in the execution stage, and uncertainty-based weighting in the final stage. Experiments use a matched Qwen3-8B/KodCode protocol with all baselines retrained, three seeds, problem-level paired-bootstrap CIs, and scale studies on Qwen3-4B/8B/14B. DHRCL reports the highest average Pass@1 (50.3±0.2 on Qwen3-8B) and consistent gains over the strongest baseline VeRPO across scales.

Significance. If the reported gains hold, DHRCL is a valuable empirical contribution: it shows that a hierarchical dense reward combined with a curriculum can improve code-RL performance without an external reward model, and it provides a carefully matched comparison protocol. The experimental design is a clear strength: all baselines are retrained under the same data, model, decoding, and budget conditions; three seeds are used; paired-bootstrap CIs support the headline comparisons; and the ablations isolate reward composition, curriculum scheduling, and token weighting. The central concern is that one of the paper's claimed contributions, the trend-based automatic stage progression, rests on a statistically fragile scheduler, and the ablation evidence for it lacks inferential tests. The empirical margin attributable to this component is small (about 0.7 Pass@1 points over a fixed schedule), so the validity of the scheduler is load-bearing for the paper's central claim.

major comments (2)
  1. [Trend-Based Automatic Stage Progression] The stage scheduler uses a one-sided OLS trend test on the most recent eight validation observations, with validation performed every two iterations. Consecutive windows overlap by seven observations, so the OLS standard errors are not valid under the independence assumption. The test is also repeated at every evaluation without multiplicity adjustment, and no sensitivity analysis is given for the window length (8), the significance level (0.05), or the two-consecutive-failure rule. This is not a purely cosmetic issue: Table 7 shows trend-based progression at 50.3 vs. 49.6 for fixed-stage, so the scheduler contributes only about 0.7 points over a fixed schedule. If that contribution is an artifact of the invalid trend test, the head-to-head margin over VeRPO (Table 1: 1.1 points) could shrink to roughly 0.4 points, which may no longer be significant. I ask for autocorrelation-robust infe
  2. [Ablation Studies] The headline comparisons against baselines are supported by paired-bootstrap CIs, but the ablations used to substantiate the individual components report only means and seed standard deviations. The key contrasts are all small: Trend-based vs. Fixed-stage (50.3 vs. 49.6, Table 7), Full vs. No Weighting (50.3 vs. 49.4, Table 7), and Full Reward vs. Pass+Syntax+Exec (50.3 vs. 49.7, Table 6/Table 4). With three seeds, these differences may or may not be statistically reliable; the manuscript does not provide inferential evidence. Since these ablations are load-bearing for the claims that the curriculum and the token-weighting ordering are beneficial, I request paired-bootstrap CIs (or equivalent) for the ablation contrasts as well.
minor comments (4)
  1. [Comparison Experiments] In Tables 1 and 2, the column 'Ext. RM' is not defined in the text or captions. It appears to refer to 'external reward model', but this should be stated explicitly.
  2. [Training-Efficiency and Reward-Signal Metrics] The DGR comparison across methods is potentially confounded by reward scale: the degeneracy threshold is an absolute value (Std < 1e-2) applied to raw rewards, and different methods use different reward scales and normalization schemes. Consider normalizing rewards to a common scale before computing DGR, or softening the cross-method DGR comparison.
  3. [Trend-Based Automatic Stage Progression] Only median transition iterations are reported (e.g., 140/280 for Qwen3-8B in Table 7). Given the importance of the scheduler, report the full distribution across seeds (e.g., min/max or individual seed values).
  4. [Stage-Aware Probability-Based Token Credit Redistribution] The phrase 'established token patterns' is informal. The mechanism uses detached token probability, and the paper later acknowledges that token probability does not identify semantic correctness. A short formal definition of what 'established' means would help reproducibility.

Circularity Check

0 steps flagged

No circularity found: DHRCL's claims are empirical, evaluated on held-out external benchmarks, and its design choices are explicitly heuristic rather than fitted predictions or self-citation-dependent derivations.

full rationale

The paper's central claim is an empirical comparison: DHRCL achieves higher average Pass@1 than VeRPO and other baselines under a matched Qwen3/KodCode protocol, with final evaluation on HumanEval, HumanEval+, BigCodeBench-Full, BigCodeBench-Hard, LiveCodeBench V6, and CodeElo. These benchmarks are external to training and are never used for reward fitting, checkpoint selection, or stage scheduling. The held-out KodCode validation split drives trend-based stage progression and early stopping, which is legitimate model selection rather than a fitted input being renamed as a prediction. The trend test in 'Trend-Based Automatic Stage Progression' fits a linear slope to recent validation observations to decide when to switch stages; this is a scheduling heuristic, not a claim that the test predicts final benchmark performance. The final Pass@1 is measured on separate external problems, so no equation reduces the reported result to its own inputs. The three-stage capability order (Syntax, Execution, Pass & Structural) is acknowledged in the Limitations as 'the three-stage capability order remains manually specified.' This is an honest admission of a design assumption, not a circular derivation. The token-weighting schemes (confidence, uniform, uncertainty) are presented as heuristics with trajectory-level mean one; they are not derived from the evaluation metric and do not define Pass@1. There is no load-bearing self-citation: the cited baselines (GRPO, AceCoder, VeRPO) are external prior work, and the paper does not rely on a uniqueness theorem or an ansatz imported from the authors' own prior publications. The skeptic's concern about overlapping-window OLS, autocorrelated validation observations, and repeated testing without multiplicity adjustment is a statistical robustness issue: if the scheduler makes noisy transition decisions, the empirical margins could change, but the evaluation would still be external and non-circular. The paper is self-contained against external benchmarks, so no significant circularity is present.

Axiom & Free-Parameter Ledger

1 free parameters · 4 axioms · 0 invented entities

The central claim rests on several hand-chosen heuristics and domain assumptions: the manually specified stage order, the interpretation of token probability as decision confidence, the linear-trend plateau detector, and the use of reference solutions as a structural prior. None of these is derived from first principles; the paper validates them through ablations.

free parameters (1)
  • trend test hyperparameters = significance 0.05, window 8, min observations 12, consecutive failures 2
    Hand-chosen values for the stage-progression rule. The paper does not vary these or report sensitivity, so the reported benefit of trend-based progression could depend on this specific configuration.
axioms (4)
  • domain assumption The three-stage capability order (Syntax, Execution, Pass & Structural) is the correct prerequisite order for code generation RL.
    Stated in the Method section as the predefined capability order; the paper itself notes in Limitations that this order is manually specified.
  • domain assumption Token probability under the rollout policy measures how 'established' a token decision is, and weighting gradients by it improves optimization without harmful bias.
    Introduced in 'Stage-Aware Probability-Based Token Credit Redistribution'; the paper asserts this heuristic but does not analyze the induced bias in the gradient estimator.
  • domain assumption The most recent 8 validation observations of the primary metric follow a linear trend, and two consecutive failures to reject H0: b<=0 reliably indicate a plateau.
    Assumed in 'Trend-Based Automatic Stage Progression'; overlapping windows and multiple comparisons make this a fragile statistical premise.
  • domain assumption Reference solutions in KodCode are correct and the AST structural reward computed against them is a useful prior that generalizes to external benchmarks.
    Used in the AST reward definition; the paper treats it as auxiliary, but the reward is computed from reference solutions and could overfit to KodCode-style structure.

pith-pipeline@v1.3.0-alltime-deepseek · 17323 in / 17513 out tokens · 181686 ms · 2026-08-05T04:21:42.693278+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of DHRCL:Training Code LLMs with Dense Hierarchical Rewards and Curriculum Learning." pith.science (2026). https://pith.science/paper/NVAPJTJZ

@misc{pith2026260726457,
  author       = {Pith},
  title        = {Pith review of: DHRCL:Training Code LLMs with Dense Hierarchical Rewards and Curriculum Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NVAPJTJZ}},
  note         = {Machine review of arXiv:2607.26457}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Reinforcement learning is a natural post-training paradigm for code-oriented large language models because generated programs can be evaluated through parsing, execution, unit tests, and structural analysis. However, existing methods often rely on sparse outcome rewards or statically combine heterogeneous dense signals, even though syntax validity, executability, functional correctness, and structural organization describe different and progressively dependent programming capabilities. We propose DHRCL, a reinforcement learning framework with Dense Hierarchical Rewards and Curriculum Learning. DHRCL decomposes feedback into syntax validation, execution success, unit-test pass rate, and AST-based structural similarity, and organizes these signals through a three-stage Syntax, Execution, Pass & Structural curriculum. Stage duration is determined automatically from recent validation trends rather than manually specified capability thresholds. We further introduce stage-aware probability-based token credit redistribution. The mechanism follows a consolidation-to-refinement principle: it emphasizes established token patterns during syntax-oriented optimization, applies uniform propagation for non-local execution feedback, and allocates more credit or blame to less-established token decisions during final functional optimization. Under a unified Qwen3-8B and KodCode protocol, the experiments compare DHRCL with binary, pass-rate, reward-model-based, and verifiable dense-reward baselines. We further evaluate DHRCL across Qwen3-4B, Qwen3-8B, and Qwen3-14B backbones, showing that its advantage remains consistent as model capacity increases.

Figures

Figures reproduced from arXiv: 2607.26457 by Hui Cheng, Shuhang Wang, Ziming Li.

Figure 1
Figure 1. Figure 1: Overall pipeline of DHRCL. where SimAST ∈ [0, 1] is the implemented AST-similarity function. If the generated program cannot be strictly parsed, we set rAST(y) = 0. Implementation details of the AST structural reward are provided in the Supplementary Docu￾ment, Section “Implementation Details of the AST Structural Reward.” AST similarity is not interpreted as semantic equivalence, maintainability, or compu… view at source ↗
Figure 2
Figure 2. Figure 2: Training dynamics of DHRCL on Qwen3-8B [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Degenerate group ratio on Qwen3-8B. term stage-wise progression. Across three seeds, the median transition iterations are 140/280 for Qwen3-8B, 180/340 for Qwen3-4B, and 120/260 for Qwen3-14B, indicating that the selected stage duration varies with model scale. Reward-signal discriminability. DGR measures how of￾ten a rollout group receives nearly identical rewards and therefore little relative-advantage i… view at source ↗

Review history (2 revisions) →

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

21 extracted references · 10 canonical work pages

  1. [5]

    CurriculumLearningforSmallCodeLanguageMod- els. In Fu, X.; and Fleisig, E., eds.,Proceedings of the 62nd Annual Meeting of the Association for Computational Lin- guistics(Volume4:StudentResearchWorkshop),ACL2024, Bangkok, Thailand, August 11-16, 2024, 531–542. Associa- tion for Computational Linguistics. Park, C.; Jiang, J.; Wang, F.; Paul, S.; Shen, J.; ...

    work page 2024

  2. [7]

    Rafailov, R.; Sharma, A.; Mitchell, E.; Manning, C

    CodeElo: Benchmarking Competition-level Code Genera- tion of LLMs with Human-comparable Elo Ratings.CoRR, abs/2501.01257. Rafailov, R.; Sharma, A.; Mitchell, E.; Manning, C. D.; Er- mon,S.;andFinn,C.2023. DirectPreferenceOptimization: Your Language Model is Secretly a Reward Model. In Oh, A.;Naumann,T.;Globerson,A.;Saenko,K.;Hardt,M.;and Levine, S., eds.,...

    Pith/arXiv arXiv 2023

  3. [8]

    Alignment with Fill-In-the-Middle for Enhancing Code Generation. In Christodoulopoulos,C.;Chakraborty,T.;Rose,C.;andPeng, V., eds.,Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, EMNLP 2025, Suzhou, China, November 4-9, 2025, 8304–8320. Associa- tion for Computational Linguistics. Rozière,B.;Gehring,J.;Gloeckle,F.;S...

    work page 2025

  4. [10]

    Shojaee, P.; Jain, A.; Tipirneni, S.; and Reddy, C

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models.CoRR, abs/2402.03300. Shojaee, P.; Jain, A.; Tipirneni, S.; and Reddy, C. K

    Pith/arXiv arXiv

  5. [11]

    Wang,A.;Yan,Y.;Zhou,N.;Lu,Z.;Lu,W.;Xiao,J.;Zhuang, Y.; and Shen, Y

    Qwen3 Technical Report.CoRR, abs/2505.09388. Wang,A.;Yan,Y.;Zhou,N.;Lu,Z.;Lu,W.;Xiao,J.;Zhuang, Y.; and Shen, Y. 2026a. Code-A1: Adversarial Evolving of Code LLM and Test LLM via Reinforcement Learning. CoRR, abs/2603.15611. Wang,J.;Zhang,Z.;He,Y.;Song,Y.;Shi,T.;Li,Y.;Xu,H.; Wu, K.; Qian, G.; Chen, Q.; and He, L

    Pith/arXiv arXiv

  6. [12]

    Wang, L.; Wu, X.; Hu, X.; Liu, Y.; Fan, Y.; Yu, K.; Weng, Q.; Xi, W.; and Li, X

    Enhancing Code LLMs with Reinforcement Learning in Code Genera- tion: A Survey.CoRR, abs/2412.20367. Wang, L.; Wu, X.; Hu, X.; Liu, Y.; Fan, Y.; Yu, K.; Weng, Q.; Xi, W.; and Li, X. 2026b. VeRPO: Verifiable Dense Reward Policy Optimization for Code Generation.CoRR, abs/2601.03525. Wang, Z.; Zhou, S.; Fried, D.; and Neubig, G

    Pith/arXiv arXiv

  7. [13]

    In Bouamor, H.; Pino, J.; and Bali, K., eds.,Findings of the Association for Computational Linguistics: EMNLP 2023, Singapore, December 6-10, 2023, Findings of ACL, 1271–1290

    Execution-BasedEvaluationforOpen-DomainCodeGener- ation. In Bouamor, H.; Pino, J.; and Bali, K., eds.,Findings of the Association for Computational Linguistics: EMNLP 2023, Singapore, December 6-10, 2023, Findings of ACL, 1271–1290. Association for Computational Linguistics. Williams, R. J

    work page 2023

  8. [14]

    In Koenig, S.; Jenkins, C.; and Taylor, M

    ReCode: Updating Code API Knowledge with Reinforcement Learn- ing. In Koenig, S.; Jenkins, C.; and Taylor, M. E., eds., Fortieth AAAI Conference on Artificial Intelligence, Thirty- Eighth Conference on Innovative Applications of Artificial Intelligence,SixteenthSymposiumonEducationalAdvances inArtificialIntelligence,AAAI2026,Singapore,January20- 27, 2026,...

    work page 2026

  9. [15]

    Process- Supervised Reinforcement Learning for Code Generation. In Christodoulopoulos, C.; Chakraborty, T.; Rose, C.; and Peng, V., eds.,Proceedings of the 2025 Conference on Em- pirical Methods in Natural Language Processing, EMNLP 2025, Suzhou, China, November 4-9, 2025, 14213–14226. Association for Computational Linguistics. Yu, Q.; Zhang, Z.; Zhu, R.;...

    work page 2025

  10. [16]

    Zeng, H.; Jiang, D.; Wang, H.; Nie, P.; Chen, X.; and Chen, W

    DAPO: An Open-Source LLM Reinforcement Learning System at Scale.CoRR, abs/2503.14476. Zeng, H.; Jiang, D.; Wang, H.; Nie, P.; Chen, X.; and Chen, W

    Pith/arXiv arXiv

  11. [17]

    In Che, W.; Nabende, J.; Shutova, E.; and Pilehvar, M

    ACECODER: Acing Coder RL via Automated Test-Case Synthesis. In Che, W.; Nabende, J.; Shutova, E.; and Pilehvar, M. T., eds.,Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2025, Vienna, Austria, July 27 - August 1, 2025, 12023–12040. Association for Compu- tational Linguistics. Zhang, D....

    work page 2025

  12. [18]

    Zhang, K.; Li, G.; Dong, Y.; Xu, J.; Zhang, J.; Su, J.; Liu, Y.; and Jin, Z

    PLUM: PreferenceLearningPlusTestCasesYieldsBetterCodeLan- guage Models.CoRR, abs/2406.06887. Zhang, K.; Li, G.; Dong, Y.; Xu, J.; Zhang, J.; Su, J.; Liu, Y.; and Jin, Z. 2025a. CodeDPO: Aligning Code Models with Self Generated and Verified Source Code. In Che, W.; Nabende,J.;Shutova,E.;andPilehvar,M.T.,eds.,Proceed- ingsofthe63rdAnnualMeetingoftheAssociat...

    Pith/arXiv arXiv 2025

  13. [19]

    Enhancing LLM Performance on Hardware Design Genera- tionTaskviaReinforcementLearning.InIEEEInternational Symposium on Circuits and Systems, ISCAS 2025, London, United Kingdom, May 25-28, 2025, 1–5. IEEE. Zheng, C.; Liu, S.; Li, M.; Chen, X.; Yu, B.; Gao, C.; Dang, K.; Liu, Y.; Men, R.; Yang, A.; Zhou, J.; and Lin, J

    work page 2025

  14. [20]

    Zhuo,T.Y.;Vu,M.C.;Chim,J.;Hu,H.;Yu,W.;Widyasari, R.; Yusuf, I

    Group Sequence Policy Optimization.CoRR, abs/2507.18071. Zhuo,T.Y.;Vu,M.C.;Chim,J.;Hu,H.;Yu,W.;Widyasari, R.; Yusuf, I. N. B.; Zhan, H.; He, J.; Paul, I.; Brunner, S.; Gong, C.; Hoang, J.; Zebaze, A. R.; Hong, X.; Li, W.; Kad- dour, J.; Xu, M.; Zhang, Z.; Yadav, P.; and et al

    Pith/arXiv arXiv

  15. [21]

    InTheThir- teenth International Conference on Learning Representa- tions, ICLR 2025, Singapore, April 24-28,

    BigCodeBench: Benchmarking Code Generation with Di- verseFunctionCallsandComplexInstructions. InTheThir- teenth International Conference on Learning Representa- tions, ICLR 2025, Singapore, April 24-28,

    work page 2025

  16. [2021]

    Dai,N.;Wu,Z.;Zheng,R.;Wei,Z.;Shi,W.;Jin,X.;Liu,G.; Dun,C.;Huang,L.;andYan,L.2024

    Evaluating Large Language Models Trained on Code.CoRR, abs/2107.03374. Dai,N.;Wu,Z.;Zheng,R.;Wei,Z.;Shi,W.;Jin,X.;Liu,G.; Dun,C.;Huang,L.;andYan,L.2024. ProcessSupervision- Guided Policy Optimization for Code Generation.CoRR, abs/2410.17621. DeepSeek-AI.2025.DeepSeek-R1:IncentivizingReasoning Capability in LLMs via Reinforcement Learning.CoRR, abs/2501.12...

    Pith/arXiv arXiv 2024

  17. [2022]

    CodeRL: Mastering Code Generation through Pretrained Models and Deep Reinforcement Learning. In Koyejo, S.; Mohamed, S.; Agarwal, A.; Belgrave, D.; Cho, K.; and Oh, A., eds.,Advances in Neural Information Pro- cessing Systems 35: Annual Conference on Neural Informa- tion Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9,

    work page 2022

  18. [2023]

    Schulman, J.; Wolski, F.; Dhariwal, P.; Radford, A.; and Klimov,O.2017

    Code Llama: Open Foundation Models for Code.CoRR, abs/2308.12950. Schulman, J.; Wolski, F.; Dhariwal, P.; Radford, A.; and Klimov,O.2017. ProximalPolicyOptimizationAlgorithms. CoRR, abs/1707.06347. Shao,Z.;Wang,P.;Zhu,Q.;Xu,R.;Song,J.;Zhang,M.;Li, Y. K.; Wu, Y.; and Guo, D

    Pith/arXiv arXiv 2017

  19. [2024]

    Gehring, J.; Zheng, K.; Copet, J.; Mella, V.; Cohen, T.; and Synnaeve, G

    StepCoder: Improve Code Generation with Reinforcement Learning from Compiler Feedback.CoRR, abs/2402.01391. Gehring, J.; Zheng, K.; Copet, J.; Mella, V.; Cohen, T.; and Synnaeve, G

    Pith/arXiv arXiv

  20. [2025]

    RLEF: Grounding Code LLMs in Ex- ecution Feedback with Reinforcement Learning. In Singh, A.; Fazel, M.; Hsu, D.; Lacoste-Julien, S.; Berkenkamp, F.; Maharaj,T.;Wagstaff,K.;andZhu,J.,eds.,Forty-secondIn- ternational Conference on Machine Learning, ICML 2025, Vancouver, BC, Canada, July 13-19, 2025, Proceedings of Machine Learning Research. PMLR / OpenRevie...

    work page 2025

  21. [2026]

    Quan, S.; Yang, J.; Yu, B.; Zheng, B.; Liu, D.; Yang, A.; Ren, X.; Gao, B.; Miao, Y.; Feng, Y.; Wang, Z.; Yang, J.; Cui, Z.; Fan, Y.; Zhang, Y.; Hui, B.; and Lin, J

    TAROT: Test-driven and Capability-adaptive CurriculumReinforcementFine-tuningforCodeGeneration with Large Language Models.CoRR, abs/2602.15449. Quan, S.; Yang, J.; Yu, B.; Zheng, B.; Liu, D.; Yang, A.; Ren, X.; Gao, B.; Miao, Y.; Feng, Y.; Wang, Z.; Yang, J.; Cui, Z.; Fan, Y.; Zhang, Y.; Hui, B.; and Lin, J

    arXiv