RESEARCH · RESEARCH · #587
FINSKILLOPS: a self-evolving multi-agent system for SEC filing QA (arXiv:2609.19680v1)
The paper introduces FINSKILLOPS, a multi-agent system that frames post-deployment improvement of SEC-filing question answering as controlled behavioral maintenance: reusable, evidence-grounded skill patches are validated, regression-checked, and versioned. On six financial QA benchmarks a frozen skill registry yields the best verdict-weighted correctness and reference consistency among evaluated systems, with evolved skills improving correctness from 3.70 to 4.55; in a 12-round operational study 6 of 33 proposed skills were promoted while the monitored non-correct rate fell from 20.0% to 12.5%.
KEY POINTS
- The paper introduces FINSKILLOPS, a multi-agent system that frames post-deployment improvement of SEC-filing question answering as controlled behavioral maintenance: reusable, evidence-grounded skill patches are validated, regression-checked, and versioned.
- On six financial QA benchmarks a frozen skill registry yields the best verdict-weighted correctness and reference consistency among evaluated systems, with evolved skills improving correctness from 3.70 to 4.55; in a 12-round operational study 6 of 33 proposed skills were promoted while the monitored non-correct rate fell from 20.0% to 12.5%.
- It demonstrates a practical, controlled approach to post-deployment self-improvement for financial QA that reduces regressions by scoping, validating, and versioning skill patches.
WHY IT MATTERS
It demonstrates a practical, controlled approach to post-deployment self-improvement for financial QA that reduces regressions by scoping, validating, and versioning skill patches.