Study Finds Weight-Space Programming Boosts Accuracy by Up to 29.45 Points with 1,023 Parameter Edits

·2026.11.10 21:02

Key point

Systematic evaluation across 16 trajectories shows WSP enables substantial localized gains but suffers from inconsistent cross-task generalization and validation reliability.

Details

WSP achieves up to 29.45 percentage points accuracy improvement on MATH-500 by modifying only 1,023 parameters, demonstrating highly localized intervention potential.
Cross-task generalization is uneven; a Math-edited checkpoint improves GPQA Diamond, IFEval, and SciBench by 3.03, 12.18, and 20.10 points respectively, while leaving SciCode performance unchanged.
Editing reliability remains a challenge, with six of 16 trajectories retaining original weights due to inability to reliably validate candidate improvements.
Experiments conducted on Qwen3.5-0.8B and Qwen3.5-9B models across seven benchmarks reveal significant variance in outcomes across different editing trajectories.

ai-research arxiv model-editing qwen3.5 weight-space-programming

Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.