What to Learn, What to Keep: Information-Guided Sampling and Text-Space Weight Decay for Self-Evolving Agent Skills
Desk rejected at ICLR 2027, so left out of the paper lists and the map.
Abstract
Large Language Model (LLM) agents fail on complex, domain-specific opera- tions that require heuristics absent from pretraining data. Skill-file self-evolution addresses this by iteratively rewriting a skill file from failure analysis, but exist- ing methods sample training instances uniformly at random, wasting budget on redundant or uninformative failures. We propose Information-Guided Active Sampling (IGAS), which casts instance selection as active learning over a non- parametric (text) skill: a multiplicative acquisition utility combining a failure- signal factor (failure similarity shaped by an entropy peak) with task represen- tativeness, and utility-weighted facility-location batch selection under a mono- tone submodular coverage objective with a (1 − 1/e) greedy guarantee. Be- cause a purely informative sampler drifts toward atypical instances whose im- provements do not transfer, we introduce ε-anchored mixing and a re-verification quota. Our evaluation protocol uses a four-way split (Train / gating ValA / final-selection ValB / Test) with strict gated acceptance (∆ > 0), a cross- round rejection memory, and a single post-hoc Test evaluation; it also exposes an inflation-accumulation phenomenon in single-validation-set selection, which our Best-of-N final selection counters through same-window dual evaluation and a paired-bootstrap discount. We evaluate on six datasets (LiveMathematician- Bench, SearchQA, DocVQA, SpreadsheetBench, ALFWorld, OfficeQA) under a preregistered controlled two-arm design—the original SkillOpt baseline ver- sus our full method—with 10 seeds per dataset per arm (120 runs in total).
- Primary area:
- foundation or frontier models, including LLMs