Tiny weight edits can fix models: 10 weights vs 130k in WSP blog benchmarks
simonguozirui · x · 2026-11-18
The alphaxiv WSP blog shows weight-space editing can fix model defects with tiny edits: GPT-6 Astra fixed a broken car policy by changing 6k weights and cut Qwen3.5-0.8B's math looping with 130k edits, while Inkling 2 Small did the same with just 10 and 8k edits — roughly 600x and 16x fewer — with no performance degradation on either task.
More from Models
- GPT-6.2 Astra models RC buggy part by part, generating animated assembly guide and parts catalog — Dr_Singularity · 2026-11-18
- Early test: user asks GPT-6.2 Astra to make a song and shares the result — Angaisb_ · 2026-11-18
- Microsoft launches MAI Cyber 2 Flash, beats GPT-6 Astra on CyberGymE2E at 2% of the cost — mustafasuleyman · 2026-11-18
- GitHub Copilot Fusion tops Terminal-Bench at 71.2%, costing 57-88% less than frontier models — mariorod1 · 2026-11-18
- Microsoft-Decision-1.1 hits GA with 2x speed at lower cost — marlene_zw · 2026-11-18
- OpenAI officially launches GPT-6.2 Astra and Sol models — kieranklaassen · 2026-11-18