FRAMES: Guarded and Dual-Objective Skill Evolution for Agents in Policy-Governed Enterprise Workflows
2026-08-03 • Artificial Intelligence
Artificial Intelligence
AI summaryⓘ
The authors created FRAMES, a system that helps large language model agents improve how they follow rules in tasks like document auditing. FRAMES starts with existing knowledge, then refines skills by comparing different versions and choosing the best balance between accuracy and cost. It also prevents new changes from causing mistakes in earlier cases while keeping everything easy to check. Testing in real settings showed FRAMES works better than other methods without extra computational cost.
large language modelsenterprise workflowsdocument auditingrule-based systemsclosed-loop frameworkPareto selectionconsensus mutationauditabilitycost-accuracy trade-offanti-regression
Authors
Xuhui Wang, Ruoqi Shu, Chen Dan, Tianhua Xu, Mengxi Luo, Yanming Mai, Bo Wan
Abstract
LLM agents increasingly run policy-bound enterprise workflows such as document auditing, where they must apply rules consistently, ground every value, and stay auditable. Improving these agents is hard: operational feedback is sparse and unlabeled, edits to one rule can regress unrelated cases, and accuracy must improve without inflating inference cost or losing auditability. We present FRAMES, a closed-loop framework that cold-starts deployable skills from existing assets and then evolves them through consensus-based mutation, Pareto selection over accuracy and cost, and an anti-regression guarantee, all while preserving auditability. Deployed on our internal production system, FRAMES attains the best accuracy-cost trade-off among baselines, with the same gains reproduced on tau-bench.