Editing 0.13% of weights, no training, fixes OCR blank-page hallucination and climbs leaderboard
vanstriendaniel · x · 2026-11-04
Daniel vanstrien demonstrates a training-free weight-editing experiment: LightOnOCR-2-1B hallucinates "There is no content in the provided image to convert into Markdown format…" on blank pages. Editing just 0.13% of its weights — with no training — eliminates this failure mode while leaving text pages unchanged. On the BHL OCR leaderboard, the edited model would jump from 14th to 8th out of 26.
He半-jokingly suggests Hugging Face needs a new basemodel metadata relationship type: "weights directly edited by a model", and has published the edited repo (LightOnOCR-2-1B-no-blank-comm...), whose model tree lists lightonai/LightOnOCR-2-1B as base.
Related event: Editing just 0.13% of weights fixes OCR model's blank-page hallucination(2 posts)→
More from Research
- NYU team open-sources MABEL, an $8K mobile bimanual robot built in 2 weeks — RemiCadene · 2026-11-05
- Retweet: Google's recurrent transformer could make sub-10B models viable soon — rickasaurus · 2026-11-05
- New paper proposes replication certification checkable by anyone in seconds — RexDouglass · 2026-11-05
- Tsinghua, NYU and Northwestern unveil RoboDoctor, a training-free framework for fixing robot actions — jiqizhixin · 2026-11-05
- ipSAE extends to ligands: fixing poor ipTM scoring in AF3 and Boltz-2 cofolding models — sokrypton · 2026-11-05
- CMU's Visual Token Routing cuts VLM agent visual tokens up to 5.8x with no accuracy loss — wellecks · 2026-11-05