Lightweight logit-level correction module fixes errors in frozen Gemma 4 E2B model without degrading base capabilities

Researchers propose CRN v2, a lightweight logit-level correction module designed to fix errors in a fully frozen language model without updating the base model. The correction module, consisting of approximately 34 million trainable parameters (0.73% of the 4.65B text module), is trained on 83,400 error-correction pairs using supervised fine-tuning followed by reference-free DPO.

On a 60-question domain exam (CEHRI: Certified Human-Robot Intelligence), CRN v2 corrects 53.3% of the base model’s errors (43.3% on a reworded variant) while showing no degradation on tested capability benchmarks, including MMLU and BoolQ (each with N=200) and car-wash (N=8). The preservation of base capabilities is attributed in part to a KL preservation term with lambda set to 0.1; reducing lambda to 0.01 degraded correction performance to 35.0%.

The study compares CRN v2 against several alternatives. A LoRA baseline with a matched parameter budget (6.6M parameters, rank 19) achieved higher correction (83.3%) but incurred substantial capability loss (30–75%) on the same benchmarks, illustrating a correction-capability tradeoff. Hidden-state injection variants at earlier layers (1.6M parameters, SFT-only) reached 50.0%/55.8% correction but did not surpass logit correction performance; performance dropped further with shallower injection (layer 4: 30.0%/28.3%). Multi-depth logit correction (~35M parameters) achieved only 40% correction, and longer training schedules (5,000 SFT + 2,000 DPO steps) did not improve beyond 53.3%.

The authors emphasize that this is a study of a design principle—frozen base model with logit correction and KL anchoring—rather than a claim of architectural novelty. All code, main-result weights, and evaluation scripts have been released, with a deep variant provided as code only (no trained deep checkpoints).