HyperSkill：无需训练的全模态 GRPO 框架，基于超图索引技能库演化
HyperSkill: Training-Free Omnimodal GRPO via Hypergraph-Indexed Skill-Library Evolution
AI 导读
HyperSkill 是一个 Training-Free 的全模态 GRPO 框架，冻结主干模型，用可审计的超图索引技能库编辑替代梯度更新，技能为带优势值的节点、组合为超边。包含训练时蒸馏、门控超图检索和执行协议三个组件。在 11 个文本、视觉语言和全模态基准上全部最优或持平，平均 F1 超最佳基线 13.0 分，软件与数据已公开。
正文 · 原文
Abstract:Reinforcement learning sharpens LLM reasoning, but parameter updates are expensive, forgetful, and opaque; training-free libraries swap gradients for a textual store but still use trajectory-level, correlation-based admission and flat structure. To address both, we propose HyperSkill, a Training-Free Omnimodal GRPO framework that keeps the backbone frozen and replaces gradient updates with auditable edits to a Hypergraph-Indexed Skill-Library, where skills are advantage-bearing nodes and compositions are hyperedges. It has three components: training-time distillation admits counterfactually checked skills and learns a protocol skill from failures; gated hypergraph retrieval injects advantage-ranked skills where a per-dataset gate admits them; and an execution protocol makes the frozen backbone commit to an answer. On 11 text, vision-language, and omnimodal benchmarks, HyperSkill is best or tied on every dataset and beats the best baseline by 13.0 average F1 points while keeping the backbone frozen. Our software and data are publicly available at this https URL.
| Comments: | Accepted by NeurIPS 2026 main conference |
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2611.15178 [cs.AI] |
| (or arXiv:2611.15178v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2611.15178 arXiv-issued DOI via DataCite (pending registration) |
|
| Journal reference: | NeurIPS 2026 |
Submission history
From: Xinliang Zhou [view email]
[v1]
Wed, 11 Nov 2026 08:20:21 UTC (1,218 KB)
来源：arXiv cs.AI · arxiv.org