SkillCAT Introduces Topology-Aware Skill Self-Evolution for LLM Agents

The arXiv submission arXiv:2606.13317 (submitted 11 Jun 2026) presents SkillCAT, a training-free pipeline for converting LLM agent execution traces into reusable skills. The paper describes three named stages: Contrastive Causal Extraction (CCE), Assessment-Augmented Evolution (AAE), and Topology-Aware Task Execution (TTE), and evaluates the method on SpreadsheetBench, WikiTableQuestions, and DocVQA, per the submission. The authors report that SkillCAT raises the average score over baselines by up to 40.40%, and that the approach requires no additional model training, according to the arXiv paper.

Per the paper, CCE samples multiple success/failure trajectory pairs for the same task and extracts evidence that correlates with outcome differences. AAE replays candidate patches on source-task clones and retains only patches that improve or preserve outcomes before hierarchical merging. TTE compiles evolved skills into a routable sub-skill graph so inference loads only capability nodes relevant to a given task, as described in the submission.

Methods that contrast successful and failed trajectories to isolate causal behavior reduce reliance on single-shot traces and can produce higher-quality, evidence-backed skill patches. Replay-based validation of candidate patches, as described in the paper, aligns with broader reproducibility practices in agent training and can reduce propagated errors from noisy extractions. Topology-aware loading addresses a practical systems tradeoff between a large skill corpus and inference efficiency, a recurring concern in agent deployments.

For practitioners, the paper is notable because it proposes a training-free route to improve agent behavior and reusability, which can be attractive when retraining models is costly or infeasible. The reported 40.40% improvement, if replicated, would represent a substantive empirical gain on the evaluated benchmarks and merits follow-up replication and ablation studies to quantify where gains come from.

Observers should look for a public code release, replication across more tasks and LLM sizes, ablation of the CCE and AAE stages, and measurements of runtime and memory benefits from the topology-aware loader compared with full-corpus inference. The arXiv submission itself is the only source for these results at present.