Safe Error Correction for Language Models: Frozen-Base Adjustment with Capability Preservation

Safe Error Correction for Language Models. Frozen-Base Adjustment with Capability Preservation. By Gautam Kishore, from Eulogik. Published on the fourteenth of September. Its code repository is worked out: about eighty source files, with a readme and a licence.

The question here is practical. Can a small correction module fix the errors a frozen language model makes, without degrading what the model could already do? The paper's answer is a design principle rather than a new architecture, and it says so plainly. Keep the base model frozen. Correct its outputs at the level of the logits. And anchor the correction with a KL preservation term. The vehicle is CRN v2, a lightweight module of about thirty-four million trainable parameters, which is zero point seven three percent of the four point six five billion parameter text module it sits on top of. That base is a fully frozen Gemma 4 E2B model, and it is never updated.

Only the correction module learns. It is trained with supervised fine-tuning and then reference-free DPO, on eighty-three thousand four hundred error-correction pairs. The test is a sixty-question domain exam called CEHRI, short for Certified Human-Robot Intelligence, which covers facts, arithmetic and implicit-goal reasoning. CRN v2 corrects fifty-three point three percent of the base model's errors, and forty-three point three percent on a reworded variant of the exam. On the capability benchmarks tested, MMLU, BoolQ and a small car-wash set, it shows no degradation.

The contrast is a LoRA baseline given the budget of the earlier CRN v1, six point six million parameters at rank nineteen. It corrects more, eighty-three point three percent, but loses between thirty and seventy-five percent of its capability on the same benchmarks. That gap is what the paper calls the correction-capability tradeoff. An ablation shows the KL term matters: with its weight at zero point one things work, and lowering it to zero point zero one drops correction to thirty-five percent.

The author is careful about what the fifty-three percent means. Every alternative tried fell short of it. Injecting into hidden states at earlier layers, with one point six million parameters and supervised training only, reported fifty percent and fifty-five point eight percent, but did not beat logit correction. Injecting shallower, at layer four, fell to thirty percent and twenty-eight point three percent. Correcting logits at multiple depths reached only forty percent, and longer training stayed at fifty-three point three. Since none of these exceeded the rank one hundred twenty-eight logit result, the paper frames that figure as the best it achieved, not as a floor. The capability checks are also small: two hundred items for MMLU and BoolQ, and eight for the car-wash set. Code, the main-result weights and evaluation scripts are released, with the deeper variant shared as code only.

Stripped down, the claim is about where to intervene. Leave the model alone, adjust only what comes out of it, and tether that adjustment to the original behaviour. On this exam, that bought roughly half the errors back with nothing measurable lost, while a comparable LoRA bought more corrections by giving up the model's broader competence.