HyperSkill: Training-Free Omnimodal GRPO via Hypergraph-Indexed Skill-Library Evolution
Summary
- Researchers introduced HyperSkill, a training-free omnimodal GRPO framework that keeps backbone models frozen.
- It replaces gradient updates with auditable edits to a Hypergraph-Indexed Skill-Library using advantage-bearing nodes and hyperedges.
- The framework consists of training-time distillation, gated hypergraph retrieval, and an execution protocol for answer commitment.
- HyperSkill outperforms the best baseline by an average of 13.0 F1 points across 11 text, vision-language, and omnimodal benchmarks.
Why it matters
HyperSkill offers a novel training-free approach to enhance LLM and omnimodal reasoning without expensive parameter updates, using a structured skill-library instead.