SkillCAT: Contrastive Assessment and Topology-Aware Skill Self-Evolution for LLM Agents

A new framework called SkillCAT allows large language model agents to improve their own capabilities without additional training, according to a paper submitted to arXiv on 11 Jun 2026 [1]. The system organizes learned skills into a routable topology, boosting benchmark scores by up to 40.40% over baseline methods [2]. SkillCAT separates the skill self-evolution process into three stages: Contrastive Causal Extraction, Assessment-Augmented Evolution, and Topology-Aware Task Execution [1]. The framework is designed to address limitations in current pipelines, which typically learn from a single trajectory per task, merge candidate skill patches before verification, and load an entire skill corpus prior to inference [2]. In the first stage, Contrastive Causal Extraction samples multiple trajectories for each task and compares successful and failed attempts on the same task to isolate evidence that explains the difference in outcomes [2]. The second stage, Assessment-Augmented Evolution, replays each candidate skill patch on clones of the source task and retains only those patches that improve or preserve task results before performing a hierarchical merge [2]. The final stage, Topology-Aware Task Execution, compiles the evolved skills into a routable sub-skill topology so that inference loads only the capability nodes relevant to the current task [2]. The researchers evaluated SkillCAT on common agent benchmarks including SpreadsheetBench, WikiTableQuestions, and DocVQA, and also tested cross-model and out-of-distribution generalization [2]. Across these settings, SkillCAT raised the average score over baselines by up to 40.40% [2]. The paper was posted on arXiv, an open-access repository of electronic preprints that, as of November 2024, receives about 24,000 submissions per month and hosts over two million articles [7]. The work appears under the Computation and Language category [1]. Large language models, the underlying technology for these agents, are neural networks trained on vast text corpora for tasks such as generation, summarization, and translation [9]. Benchmark evaluations for such models typically measure reasoning, factual accuracy, and safety [9]. The SkillCAT framework does not require model training, instead relying on trajectory replay and topological routing to achieve reliable skill evolution [2].