Editing 0.13% of weights, no training, fixes LightOnOCR-2-1B blank-page hallucination
vanstriendaniel · x · 2026-11-04
A developer shows that a no-training edit to just 0.13% of LightOnOCR-2-1B's weights stops the model from hallucinating long "there is no content in the provided image" responses on blank pages, which would move it from 14th to 8th of 26 on the BHL OCR leaderboard with text pages unchanged.
A side finding: the behavior only appears when served with vLLM — 40/40 blank pages triggered under vLLM vs 1/40 under transformers — hinting the inference stack itself influences model behavior. The author cautions against using the edit in real work.
Related event: Editing just 0.13% of weights fixes OCR model's blank-page hallucination(2 posts)→
More from Models
- Kimi K3's compute load too heavy to serve economically, analysis argues — teortaxesTex · 2026-11-04
- Kimi K3.1 Pro leaked doc shows Gated DeepSeek Sparse Attention for long context — zephyr_z9 · 2026-11-04
- Nathan Lambert: Chinese open-model labs have never missed a weights release promise — natolambert · 2026-11-04
- MIT: GPT-6 Astra rewrites other models' weights directly, lifting Qwen math accuracy from 7.9% to 48.1% — The Decoder · 2026-11-04
- Moonshot releases Kimi K3.1-Pro as open weights on Hugging Face — ApprehensiveAd3629 · 2026-11-04
- Moonshot open-weights Kimi K3.1 Pro: 2.8T-param MoE with 1M context — scaling01 · 2026-11-04