← sources

Can open-weight LLMs can predict the effects of weight edits without running the model (9B–30B)

source: reddit (reddit:LocalLLaMA)published 2026-11-14 09:07ingested 2026-11-15 02:36

retained text

I’ve been experimenting with weight-space reading using Qwen and Muse models (9B–30B).

The goal: can an LLM inspect raw model weights and predict the behavioral effects of weight edits without running the target model?

A few findings:  
🔹Qwen 27B reached 73.0% balanced accuracy on subtle edits with weight access, vs. 57.8% from text descriptions.

🔹We couldn’t establish a conclusive advantage for models reading their own weights.

🔹On Muse-30B, 34 standard edits and 6 no-op controls produced no detectable behavioral changes, but weight-reading agents frequently predicted changes.

I’ve released the paper’s evaluation harness and experimental code, plus an open-source toolkit for LLM-based weight-space programming and reading.

Code: https://github.com/suayptalha/wsr

Paper: https://www.alphaxiv.org/abs/2611.weight-space-reading-language-models

I’d be interested in feedback on the evaluation design, especially from people experimenting with open-weight models.

document fetched 2026-11-15 00:36 · canonical https://reddit.com/r/LocalLLaMA/comments/1xvxyhj/can_openweight_llms_can_predict_the_effects_of