Frozen-base 34M logit correction module fixes 53.3% of Gemma errors losslessly
eulogik · hf · 2026-09-29
Can a small correction module fix a frozen language model's errors without degrading base capabilities? CRN v2 answers partially yes.
Key findings:
- CRN v2 is a 34M-parameter (0.73% of the 4.65B text module) logit-level correction module atop a fully frozen Gemma 4 E2B, trained via SFT plus reference-free DPO on 83,400 error-correction pairs;
- It corrects 53.3% of base-model errors on a 60-question domain exam with no degradation on MMLU/BoolQ benchmarks;
- A matched-budget LoRA baseline achieves 83.3% correction but loses 30–75% capability, illustrating the correction-capability tradeoff;
- Ablations show the KL preservation term (λ=0.1) is critical (dropping it to 0.01 cuts correction to 35.0%); hidden-state injection, multi-depth, and longer training variants all underperform;
- Code, weights, and eval scripts released.
More from Research
- Why chirality matters for AI for Science: enantiomers behave differently in biology — madeofAjala · 2026-11-16
- RSS 2027 submissions open with FAQ clarifying two-stage review and negative-result bar — GeorgiaChal · 2026-11-16
- Mobile AI Factory: agents auto-build on-device SLMs with 19x less GPU time — pmttyji · 2026-11-16
- Berkeley paper: LLMs answer with stale values even when the update is in context — ombre-ai · 2026-11-16
- LatamGPT releases 178.4M-document open Latin American corpus on Hugging Face — OmarUFlorez · 2026-11-16
- Open-source TotalSegmentator spots a 42mm dilated aorta on a routine chest CT — MaziyarPanahi · 2026-11-16