Pith. sign in

REVIEW 8 cited by

SRFT: A Single-Stage Method with Supervised and Reinforcement Fine-Tuning for Reasoning

Not yet reviewed by Pith; the record is open.

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 2506.19767 v1 pith:2LVCV5IP submitted 2025-06-24 cs.CL cs.AIcs.LG

classification cs.CLcs.AIcs.LG
keywords fine-tuningreasoningreinforcementsrftsupervisedbenchmarksdistributionsintegration
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
read the original abstract

Large language models (LLMs) have achieved remarkable progress in reasoning tasks, yet the optimal integration of Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) remains a fundamental challenge. Through comprehensive analysis of token distributions, learning dynamics, and integration mechanisms from entropy-based perspectives, we reveal key differences between these paradigms: SFT induces coarse-grained global changes to LLM policy distributions, while RL performs fine-grained selective optimizations, with entropy serving as a critical indicator of training effectiveness. Building on these observations, we propose Supervised Reinforcement Fine-Tuning (SRFT), a single-stage method that unifies both fine-tuning paradigms through entropy-aware weighting mechanisms. Our approach simultaneously applies SFT and RL to directly optimize the LLM using demonstrations and self-exploration rollouts rather than through two-stage sequential methods. Extensive experiments show that SRFT achieves 59.1% average accuracy, outperforming zero-RL methods by 9.0% on five mathematical reasoning benchmarks and 10.9% on three out-of-distribution benchmarks.

Discussion (0). Sign in to comment.

Forward citations

Cited by 8 Pith papers

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

  1. Deferred Exposure of Future Trajectories for Verifiable Reasoning in Autonomous Driving VLMs

    cs.AI 2026-08 conditional novelty 7.0 of 10

    Hiding future trajectory information until after a driving model forms its decision reduces rationalization and improves verifiable autonomous-driving reasoning in the proposed AD-MCQ and DEFT-RLVR framework.

  2. Entropy-Preserving Supervised Fine-Tuning via Adaptive Self-Distillation for Large Reasoning Models

    cs.LG 2026-02 conditional novelty 6.0 of 10

    CurioSFT preserves exploration during supervised fine-tuning by distilling toward the model's own temperature-scaled distribution and adaptively increasing entropy at high-entropy tokens, improving SFT accuracy by ~2....

  3. Beyond Reasoning Gains: Mitigating General-Capability Forgetting in Large Reasoning Models

    cs.LG 2025-10 conditional novelty 6.0 of 10

    A dynamic replay and reweighting scheduler (RECAP) preserves general capabilities during RLVR while keeping reasoning performance at least as good as reasoning-only finetuning.

  4. Don't Tell the Answer, Truly Guide the Reasoning During RL Rollouts

    cs.LG 2025-10 conditional novelty 6.0 of 10

    HINT boosts LLM reasoning RL by injecting teacher-generated heuristic hints only on all-failed rollouts, keeping hints out of the policy-optimization prompt, and monitoring guidance quality with a new Affinity metric.

  5. CORRECT: COndensed eRror RECognition via knowledge Transfer in multi-agent systems

    cs.MA 2025-09 conditional novelty 6.0 of 10

    CORRECT distills recurring multi-agent failure patterns into reusable error schemas and retrieves them at inference time to localize the decisive error step more accurately than judging or fine-tuning baselines.

  6. ARMOR: Stabilizing On-Policy LLM RL with Off-Policy Anchor Samples

    cs.LG 2026-07 conditional novelty 5.5 of 10

    Injecting correct reference-policy samples and optimizing a mixed importance-sampling ratio prevents validation collapse and raises asymptotic math-reasoning scores beyond reverse-KL baselines.

  7. RoboGPT-R1: Enhancing Robot Task Planning with Reinforcement Learning

    cs.AI 2025-10 conditional novelty 5.0 of 10

    A 3B VLM trained with SFT plus GRPO and an LCS-based reward reaches 55.3% on EmbodiedBench's EB-ALFRED, beating GPT-4o-mini and the 7B REBP planner.

  8. Beyond Scaling Law: A Data-Efficient Distillation Framework for Reasoning

    cs.LG 2025-08 reject novelty 5.0 of 10

    DED trains a 32B reasoning model to state-of-the-art levels on AIME and LiveCodeBench using roughly 0.8k curated examples, via teacher selection, hard-example compression, and trajectory diversity.

Pith tools