frontier models
A Small Corrector Fixed Errors Without Rewriting the Base Model
A 34-million-parameter module corrected 53.3% of tested errors while the matched evaluations showed no base-capability loss.
Summary
A 34-million-parameter module corrected 53.3% of tested errors while the matched evaluations showed no base-capability loss.
CRN v2 leaves a 4.65-billion-parameter language model frozen and learns a roughly 34-million-parameter logit correction layer on 83,400 pairs. On a 60-question domain exam it corrected 53.3% of base-model errors; a reworded version reached 43.3%. A matched LoRA baseline corrected more errors but lost 30% to 75% on the same small capability checks. Lowering the preservation penalty sharply reduced correction. The experiment supports frozen-base adjustment as a tradeoff, but its 60-question domain exam and small benchmark samples are not evidence of broad capability preservation.
Why it matters
A 34-million-parameter module corrected 53.3% of tested errors while the matched evaluations showed no base-capability loss.
Limits and context
- The experiment supports frozen-base adjustment as a tradeoff, but its 60-question domain exam and small benchmark samples are not evidence of broad capability preservation.
Key claims
A 34-million-parameter module corrected 53.3% of tested errors while the matched evaluations showed no base-capability loss.
Qualification: The experiment supports frozen-base adjustment as a tradeoff, but its 60-question domain exam and small benchmark samples are not evidence of broad capability preservation.
Evidence: source-2026-09-16-004
Sources
- arXiv preprint 2609.16145arXiv · primary research
Corrections
No corrections have been recorded for this story.