IWIBench tests gradient-free intra-weight injection by editor models
Original title: IWIBench: Evaluating Gradient-Free Model Driven Intra-weight Injection
Astra and GPT-6.1 Sol lead editors of Gemma 3 4B IT; structured-response edits resisted every attempt.
IWIBench measures whether a large editor can change a smaller target, typically Gemma 3 4B IT, by editing weights without gradients toward one of ten text-editing objectives while preserving other capabilities. Editors get four weight read/write tools, bounded tensor access, and a fixed rollout budget; a validator applies each patch in a clean environment. Frontier closed- and open-weight models diverge sharply. Astra and GPT-6.1 Sol lead; little of the ability appears in GPT-5.6 Sol. Structured response resisted every attempt, and answer-length control had one success, while alias acquisition and recoding were easiest.
Unconfirmed: The claim that the capability was acquired through recent training is the authors' inference from recency and qualitative analysis.
- Typical target is Gemma 3 4B IT across ten objectives.
- Editors are limited to four tools and a fixed rollout budget.
- GPT-5.6 Sol shows little of the capability seen in GPT-6.1 Sol.
- Success in one task family correlated poorly with others.
From the original article
1 min read at huggingface.co
Intra-weight injection lets one model change another model's behaviour by editing its weights directly, without gradients. Such edits can range from altering lexical associations to enforcing output formats or updating factual knowledge. Despite growing interest in these approaches, there has been no systematic way to evaluate how well models perform them.
By Chris von Csefalvay, Pranav Doma, Achyuthan Unni Krishnan · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.