Brief
Small correction module fixes 53% of frozen model errors in limited study
A new study tests whether a small correction module can fix errors in a frozen language model without degrading its base abilities. It reports 53.3% error correction on a 60-question exam with no measured degradation on small tested benchmarks, but the evidence is from a preprint and limited evaluations.
The study describes CRN v2, a logit-level correction module with about 34 million trainable parameters—0.73% of a 4.65B text module—sitting atop a fully frozen Gemma 4 E2B model. Only the correction module learns, using supervised fine-tuning followed by reference-free DPO on 83,400 error-correction pairs.
On a 60-question domain exam covering facts, arithmetic, and implicit-goal reasoning, it corrects 53.3% of base-model errors (43.3% on a reworded variant) while showing no degradation on tested capability benchmarks. A LoRA baseline at a matched smaller budget corrects 83.3% but loses 30–75% capability. The authors call this a design-principle study, not architectural novelty, and release code, main-result weights, and evaluation scripts.
Our reading
Our reading is that this is a modest, carefully bounded demonstration of a correction-capability tradeoff, not evidence that frozen-model error correction is solved.
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 with ~34M trainable parameters, 0.73% of the 4.65B text module.
- 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.
- An ablation shows that 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.