Coalition-Aware Skill Reliability for Self-Evolving Agents
Abstract Overview
This paper investigates whether skills stored in self-evolving LLM agents' skill banks make positive mechanistic contributions rather than merely improving aggregate outcome metrics. Through a Skill Mechanistic Reliability Audit (SMRA), the authors identify two failure modes: coalition pollution, where bank-level gains mask individual skills with zero or negative coalition-level contributions, and cross-domain utility reversal, where skills helpful in a source domain degrade performance after transfer. To address these issues, the authors propose Coalition-Aware Skill Selection (CASS) for skill accumulation and the Unsupervised Skill-Masked Coalition Optimizer (u-SMCO) for post-transfer masking. The work demonstrates that skill reliability is a joint property of the individual skill, the surrounding skill bank, and the target deployment domain.
Novelty
The paper reframes agent skill reliability as a coalition- and domain-dependent property rather than an intrinsic attribute of individual skills. It introduces two targeted interventions: CASS, which uses Monte Carlo sampled Shapley marginals to gate skill accumulation during evolution, and u-SMCO, which uses label-free target-domain retrieval quality knockouts to greedily mask utility-reversing skills post-transfer.
Results
Evaluations across LoCoMo, LongMemEval, HotpotQA, and ALFWorld show that CASS and u-SMCO consistently outperform strong self-evolving agent baselines such as MemSkill. CASS admitted significantly fewer candidate skills than standard outcome gates while avoiding net-negative additions, and u-SMCO improved transfer performance on all tested trained banks without requiring target-domain supervision.
Key Points
- SMRA reveals that outcome-only gates are structurally blind to coalition pollution during skill evolution and cross-domain utility reversal during transfer.
- CASS incorporates sampled coalition knockouts and Shapley marginal contributions into the gating mechanism to admit only skills with verified coalition-conditioned reliability.
- u-SMCO evaluates retrieval-quality knockouts on unlabeled target-domain queries to identify and mask transferred skills whose presence degrades target-domain behavior.