Online Self-Weighted Fine-Tuning · CiteArk

Online Self-Weighted Fine-Tuning

公开

作者：Haiquan Wen、Yiwei He、Bei Peng、Guangliang Cheng

arXiv 2026 机器学习 (cs.LG)人工智能

复现结果

全部结论 ↗

尚无复现结果

论文中的结论可在「研究结论」中查看。

完整研究主张目录

0 / 6 条结论已通过验证

其余仍在验证中

关键指标论文报告值 → 复现观测值 · 点击行展开详情

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.整组结论：实验受阻，详见具体原因0/10 项指标获得支持 · 0 项已有评估10项测量结果展开结果表

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.

测量结果

已评估 0/10 项

获得支持 0有争议或冲突 0尚未评估 10

保留同一结论下的完整结果组。已有评估排在前面；单项判断不替代整组结论。展开行查看条件、来源与历次证据。

每行是一项测量，分别显示论文报告、最近可用实测和累计证据状态。
论文报告	最近可用实测	证据判断
6.67 percentage_points	—	尚未评估
6.67 percentage_points	—	尚未评估
22.8 percentage_points	—	尚未评估
15.15 percentage_points	—	尚未评估
6.67 percentage_points	—	尚未评估
10 percentage_points	—	尚未评估
24.2 percentage_points	—	尚未评估
20.2 percentage_points	—	尚未评估
20.71 percentage_points	—	尚未评估
20.71 percentage_points	—	尚未评估

打开 Claim 详情

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.整组结论：实验受阻，详见具体原因0/6 项指标获得支持 · 0 项已有评估6项测量结果展开结果表

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.

测量结果

已评估 0/6 项

获得支持 0有争议或冲突 0尚未评估 6

保留同一结论下的完整结果组。已有评估排在前面；单项判断不替代整组结论。展开行查看条件、来源与历次证据。

每行是一项测量，分别显示论文报告、最近可用实测和累计证据状态。
论文报告	最近可用实测	证据判断
5.42 percentage_points	—	尚未评估
26.67 percentage_points	—	尚未评估
3.96 percentage_points	—	尚未评估
23.34 percentage_points	—	尚未评估
57.4 percentage_points	—	尚未评估
26.27 percentage_points	—	尚未评估

打开 Claim 详情

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.整组结论：等待复现0/12 项指标获得支持 · 0 项已有评估12项测量结果展开结果表

实验方案已生成，还没有运行记录

测量结果

已评估 0/12 项

获得支持 0有争议或冲突 0尚未评估 12

保留同一结论下的完整结果组。已有评估排在前面；单项判断不替代整组结论。展开行查看条件、来源与历次证据。

每行是一项测量，分别显示论文报告、最近可用实测和累计证据状态。
论文报告	最近可用实测	证据判断
0.83 percentage_points	—	尚未评估
6.67 percentage_points	—	尚未评估
1.04 percentage_points	—	尚未评估
10 percentage_points	—	尚未评估
8.72 percentage_points	—	尚未评估
41.86 percentage_points	—	尚未评估
13.04 percentage_points	—	尚未评估
56.52 percentage_points	—	尚未评估
5.14 percentage_points	—	尚未评估
28.89 percentage_points	—	尚未评估
24.2 percentage_points	—	尚未评估
20.2 percentage_points	—	尚未评估

打开 Claim 详情

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.整组结论：实验受阻，详见具体原因0/6 项指标获得支持 · 0 项已有评估6项测量结果展开结果表

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.

测量结果

已评估 0/6 项

获得支持 0有争议或冲突 0尚未评估 6

保留同一结论下的完整结果组。已有评估排在前面；单项判断不替代整组结论。展开行查看条件、来源与历次证据。

每行是一项测量，分别显示论文报告、最近可用实测和累计证据状态。
论文报告	最近可用实测	证据判断
13.33 percentage_points	—	尚未评估
36.67 percentage_points	—	尚未评估
16.25 percentage_points	—	尚未评估
36.67 percentage_points	—	尚未评估
79 percentage_points	—	尚未评估
32.32 percentage_points	—	尚未评估

打开 Claim 详情

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).整组结论：实验受阻，详见具体原因0/6 项指标获得支持 · 0 项已有评估6项测量结果展开结果表

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

测量结果

已评估 0/6 项

获得支持 0有争议或冲突 0尚未评估 6

保留同一结论下的完整结果组。已有评估排在前面；单项判断不替代整组结论。展开行查看条件、来源与历次证据。

每行是一项测量，分别显示论文报告、最近可用实测和累计证据状态。
论文报告	最近可用实测	证据判断
23.8 percentage_points	—	尚未评估
15.15 percentage_points	—	尚未评估
21.4 percentage_points	—	尚未评估
15.17 percentage_points	—	尚未评估
24 percentage_points	—	尚未评估
20.71 percentage_points	—	尚未评估

打开 Claim 详情

其余结论1

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.论文自述边界

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

打开 Claim 详情