Online Self-Weighted Fine-Tuning

Public
Authors:Haiquan Wen, Yiwei He, Bei Peng, Guangliang Cheng
More
Copy repository link

AI research summary

Standard supervised fine-tuning (SFT) treats all expert demonstrations identically, allocating gradient updates uniformly regardless of how well the model has already mastered each query. Online Self-Weighted Fine-Tuning (OSW-FT) addresses this limitation for binary-verifiable reasoning by dynamically rescaling the supervised fine-tuning loss using an online pass rate estimated from a small number of rollouts. By anchoring the optimization direction to high-quality expert trajectories while modulating update magnitude using an empirical advantage weight of one minus the estimated success rate, OSW-FT concentrates gradient steps on unsolved queries. On the Qwen3 model family across challenging benchmarks including AIME, AMC, MATH-500, and GPQA-Diamond, OSW-FT consistently improves over standard SFT with only two rollouts per query.

Generated by CiteArk.

What this means

Supervised fine-tuning of large language models on reasoning tasks traditionally applies equal gradient updates to every example, wasting capacity and injecting gradient noise on already-solved problems. OSW-FT bridges supervised fine-tuning and reinforcement learning by using lightweight online rollouts to estimate an empirical failure rate that dynamically scales the loss on expert trajectories. This allows post-training practitioners to focus gradient updates on unsolved capability frontiers without incurring the high sample complexity and training instability of full policy-gradient reinforcement learning. The primary limitation is its dependence on binary-verifiable reward verifiers and available high-quality expert demonstrations, restricting direct applicability in open-ended or non-verifiable domains.

Reproduction progress
RunMatchRepeat

Not reproduced yet

0/5

No verifiable claim has successful reproduction evidence yet.

Awaiting author or institution signature.

Research claims

The main empirical claims extracted in this research plan, with their plan coverage and live evidence state. Click a claim to expand it inline.

In rollout count ablation on Qwen3-0.6B-Base, increasing K from 1 to 2 yields a significant performance jump (from 15.15% to 20.20% on GPQA-Diamond, and from 6.67% to 10.00% on AIME 25 Pass@16), while further increasing K to 4 or 8 yields negligible marginal benefits.Experiment blocked; see the specific reasonReported 6.67 pp

Reported

6.67 pp

Observed

The rollout ablation K=1 vs K=2 tests the core theoretical inflection point on Qwen3-0.6B-Base and can be partially evaluated alongside the main experiment.

pass_at_16 · Not assessed

Reported: 6.67 pp

pass_at_16 · Not assessed

Reported: 6.67 pp

pass_at_1 · Not assessed

Reported: 22.8 pp

pass_at_1 · Not assessed

Reported: 15.15 pp

pass_at_16 · Not assessed

Reported: 6.67 pp

pass_at_16 · Not assessed

Reported: 10 pp

pass_at_1 · Not assessed

Reported: 24.2 pp

pass_at_1 · Not assessed

Reported: 20.2 pp

pass_at_1 · Not assessed

Reported: 20.71 pp

pass_at_1 · Not assessed

Reported: 20.71 pp

OSW-FT (K=2) on Qwen3-1.7B-Base achieves 5.42% Pass@1 and 26.67% Pass@16 on AIME 24, 3.96% Pass@1 and 23.34% Pass@16 on AIME 25, 22.24% Pass@1 and 55.81% Pass@16 on AMC 22, 27.45% Pass@1 and 65.22% Pass@16 on AMC 23, 17.78% Pass@1 and 48.89% Pass@16 on AMC 24, 57.40% Pass@1 on MATH-500, and 26.27% Pass@1 on GPQA-Diamond.Experiment blocked; see the specific reasonReported 5.42 pp

Reported

5.42 pp

Observed

Training Qwen3-1.7B full-parameter fine-tuning for 1 epoch with 2 online rollouts exceeds the cumulative compute and runtime budget when executed alongside the primary 0.6B verification experiments.

pass_at_1 · Not assessed

Reported: 5.42 pp

pass_at_16 · Not assessed

Reported: 26.67 pp

pass_at_1 · Not assessed

Reported: 3.96 pp

pass_at_16 · Not assessed

Reported: 23.34 pp

pass_at_1 · Not assessed

Reported: 57.4 pp

pass_at_1 · Not assessed

Reported: 26.27 pp

OSW-FT (K=2) on Qwen3-0.6B-Base achieves 0.83% Pass@1 and 6.67% Pass@16 on AIME 24, 1.04% Pass@1 and 10.00% Pass@16 on AIME 25, 8.72% Pass@1 and 41.86% Pass@16 on AMC 22, 13.04% Pass@1 and 56.52% Pass@16 on AMC 23, 5.14% Pass@1 and 28.89% Pass@16 on AMC 24, 24.20% Pass@1 on MATH-500, and 20.20% Pass@1 on GPQA-Diamond, matching or outperforming standard SFT across all evaluated benchmarks.Awaiting reproductionReported 0.83 pp

Reported

0.83 pp

Observed

Experiment plan ready; no runs yet.

pass_at_1 · Not assessed

Reported: 0.83 pp

pass_at_16 · Not assessed

Reported: 6.67 pp

pass_at_1 · Not assessed

Reported: 1.04 pp

pass_at_16 · Not assessed

Reported: 10 pp

pass_at_1 · Not assessed

Reported: 8.72 pp

pass_at_16 · Not assessed

Reported: 41.86 pp

pass_at_1 · Not assessed

Reported: 13.04 pp

pass_at_16 · Not assessed

Reported: 56.52 pp

pass_at_1 · Not assessed

Reported: 5.14 pp

pass_at_16 · Not assessed

Reported: 28.89 pp

pass_at_1 · Not assessed

Reported: 24.2 pp

pass_at_1 · Not assessed

Reported: 20.2 pp

OSW-FT (K=2) on Qwen3-4B-Base achieves 13.33% Pass@1 and 36.67% Pass@16 on AIME 24, 16.25% Pass@1 and 36.67% Pass@16 on AIME 25, 41.42% Pass@1 and 74.42% Pass@16 on AMC 22, 45.11% Pass@1 and 78.26% Pass@16 on AMC 23, 35.00% Pass@1 and 75.56% Pass@16 on AMC 24, 79.00% Pass@1 on MATH-500, and 32.32% Pass@1 on GPQA-Diamond.Experiment blocked; see the specific reasonReported 13.33 pp

Reported

13.33 pp

Observed

Training Qwen3-4B full-parameter fine-tuning for 1 epoch with 2 online rollouts requires multi-GPU distributed setup and exceeds the single L4 GPU memory and budget limits.

pass_at_1 · Not assessed

Reported: 13.33 pp

pass_at_16 · Not assessed

Reported: 36.67 pp

pass_at_1 · Not assessed

Reported: 16.25 pp

pass_at_16 · Not assessed

Reported: 36.67 pp

pass_at_1 · Not assessed

Reported: 79 pp

pass_at_1 · Not assessed

Reported: 32.32 pp

Under K=4 rollouts on Qwen3-0.6B-Base, continuous online success-rate weighting in OSW-FT (24.00% MATH-500, 20.71% GPQA-Diamond) outperforms hard-thresholded binary rejection weighting (23.80% MATH-500, 15.15% GPQA-Diamond) and mean-matched random weighting (21.40% MATH-500, 15.17% GPQA-Diamond).Experiment blocked; see the specific reasonReported 23.8 pp

Reported

23.8 pp

Observed

Running additional training runs at K=4 across multiple alternative weighting heuristics exceeds the cumulative paper budget allocation.

pass_at_1 · Not assessed

Reported: 23.8 pp

pass_at_1 · Not assessed

Reported: 15.15 pp

pass_at_1 · Not assessed

Reported: 21.4 pp

pass_at_1 · Not assessed

Reported: 15.17 pp

pass_at_1 · Not assessed

Reported: 24 pp

pass_at_1 · Not assessed

Reported: 20.71 pp

OSW-FT currently requires a reliable binary outcome verifier and ground-truth expert trajectories, limiting direct application to non-binary or open-ended generative tasks.Stated by the authors

This is a qualitative scope limitation acknowledged by the authors rather than an empirical numerical benchmark claim.

Reproduction and technical details1
Implementation path
Official implementation0

The experiment runs a repository and commit explicitly declared by the paper and verified by CiteArk.

CiteArk reconstruction1

No author code is used. CiteArk independently rebuilds the paper-grounded protocol and records every generated file and assumption.

Information insufficient4

A falsifiable execution still needs a missing method detail, input, split, metric, or evaluation rule.

Signed research planVerifiedThe Research Plan CAP fixes paper sources, complete Claim coverage, planned Experiments, and source-declared values.Download research plan
Generated byCiteArk Official
Modelgoogle/gemini-3.8-flash
Completed
Artifact digest5dcba86dab
Attestation statusTrusted signature verified
Output
View independent verification JSON