Brief
Small correction module fixes some language model errors without degrading base skills
A new preprint describes a lightweight module that corrects errors in a frozen language model's outputs while preserving its base capabilities. It corrects fewer errors than a LoRA baseline that does degrade capabilities.
The study proposes CRN v2, a logit-level correction module with about 34 million trainable parameters, which is 0.73% of the 4.65B text module in a frozen Gemma 4 E2B model. The base model is never updated; only the correction module learns from 83,400 error-correction pairs. On a 60-question domain exam, CRN v2 corrects 53.3% of base-model errors (43.3% on a reworded variant) while showing no degradation on tested capability benchmarks (MMLU/BoolQ N=200; car-wash N=8). A LoRA baseline at a matched budget achieves 83.3% correction but suffers 30-75% capability loss. The authors note that the KL preservation term is critical: lowering it degrades correction. Code, main-result weights, and evaluation scripts are released, though the deep variant is code only with no trained checkpoints.
Source details and supporting facts
Each line is stated by the page named above it.
Stated by arXiv
- CRN v2 is a lightweight logit-level correction module (~34M trainable parameters, 0.73% of the 4.65B text module) that sits atop a fully frozen Gemma 4 E2B model.
- The base model is never updated; only the correction module learns, via supervised fine-tuning followed by reference-free DPO on 83,400 error-correction pairs.
- On a 60-question domain exam (CEHRI), CRN v2 corrects 53.3% of base-model errors (reworded variant: 43.3%) while showing no degradation on tested capability benchmarks (MMLU/BoolQ N=200; car-wash N=8).
- A LoRA baseline at the matched CRN v1 budget (6.6M params, rank 19) achieves 83.3% correction but suffers 30-75% capability loss on the same benchmarks.
- The KL preservation term (lambda=0.1) is critical: lowering it to 0.01 degrades correction to 35.0%.
- All code, main-result weights, and evaluation scripts are released (deep variant as code only -- no trained deep checkpoints).
Sources
- arXivText stored 16 September 2026
How this story was checked. Written from the 1 page listed above, stored 16 September 2026; claims checked against that stored text on 16 September 2026.
What that means
- 6 of 6 reported statements were confirmed against the page that carries them; the rest were removed rather than published.
- Figures in the text were required to appear in the stored source text: yes. Identifiers: yes.
- The check reads stored text only: no claim rests on a fresh look that did not happen.
- Where the reporting was silent, the text says so instead of filling the gap.