LearnActCoder: Role-Aware Error Memory for Adaptive Clinical Coding Agents
3.00T1 sourcearXiv cs.MA
Source record
Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2609.19721.
Pipeline notes
The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.
SummaryLearnActCoder is a Coder-Judge clinical coding agent pipeline that uses a Learn-Then-Act framework converting errors from a small labeled batch into a structured Mistake Knowledge Database. On 150 MIMIC-III notes, structured memory lifts CPT F1 by 5.9 points; ICD-9 and ICD-10 F1 gains are not significant. Absolute CPT performance remains low; evaluation is retrospective.
Why it mattersTests whether structured, feedback-derived error memory helps LLM agents adapt without weight updates. Worth noting for the mixed results: structured memory helps CPT but not ICD F1.
Cited by
No citations on record.