DAWE edits 5,064 weights out of 4B to fix harness mismatch, no retraining
lateinteraction · x · 2026-11-18
llmluthor describes DAWE (Data-Attributed Weight-Editing): when their model broke on a new harness, instead of retraining, they directly edited 5,064 weights out of 4 billion — lifting accuracy from 20.1% to 88.2% with no gradient descent, while old-harness accuracy stayed unchanged. The method is fully open-sourced.
More from Research
- Author clarifies "nothing looks interesting" remark: it's about reviewing, not condescension — rishabh16_ · 2026-11-18
- Researcher laments reviewer fatigue as coding agents flood academia with incremental papers — rishabh16_ · 2026-11-18
- Stanford team predicts viral evolution and designs antibodies that neutralize future escape variants — BrianHie · 2026-11-18
- GPBam: New German Public Law Benchmark Finds Retrieval Gains Limited, Essay Quality Tracks Statutory Knowledge — aaire · 2026-11-18
- AI agents uncover first known genome-wide frameshifting in bacteria — GaryMarcus · 2026-11-18
- Google launches Antigravity Science plugin, gives 10,000 scientists free top-tier access — DynamicWebPaige · 2026-11-18