IWIBench: Evaluating Gradient-Free Model Driven Intra-weight Injection
Summary
- IWIBench is introduced as a benchmark for evaluating gradient-free intra-weight injection, where an editor model modifies another model's weights directly.
- The benchmark measures an editor model's ability to edit a target model, typically Gemma 3 4B IT, across ten text-editing objectives while preserving capabilities.
- Editors use four bounded tensor reading and writing tools, with validators applying patches and comparing edited models against unmodified targets.
- Evaluations show diverging capabilities, with models like Astra and GPT-6.1 Sol outperforming others, and tasks like alias acquisition proving more tractable than structured response.
Why it matters
Intra-weight injection represents an emerging gradient-free approach for behavior modification in models, and IWIBench provides a systematic framework to evaluate these capabilities.