Source record

Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2609.19680.

Pipeline notes

The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.

FINSKILLOPS is a self-evolving multi-agent system for SEC filing QA that treats post-deployment improvement as controlled behavioral maintenance: typed failure diagnoses produce scoped skill patches, each admitted only after targeted validation, regression checks, negative controls, and versioned replacement. Across six benchmarks, a frozen skill registry achieves the highest verdict-weighted correctness, with evolved skills raising correctness from 3.70 to 4.55 and cutting the monitoring non-correct rate from 20.0% to 12.5%.

Relevant to anyone building agents that update after deployment: the skill admission, scope, and retirement discipline is a concrete pattern for avoiding regressions in self-improving systems.

Cited by

No citations on record.