TRACE: A Self-Evolving Skill Bank for Consistent, Limit-Aware LLM Agents
2026-08-24 • Computation and Language
Computation and LanguageArtificial Intelligence
AI summaryⓘ
The authors focus on making AI assistants used in cars more reliable by ensuring they act consistently and know their limits. They created CAR-bench, a test where an AI must handle unclear requests through conversation and tools while following strict rules. They found many models can solve tasks occasionally but struggle to do so reliably every time. To fix this, they developed TRACE, which improves how the AI uses different skills by learning from past successes and failures without changing the AI's core model. Using TRACE, the AI became much more consistent in completing tasks and ranked first on a challenging test set.
Large Language ModelsConsistencyMulti-turn DialogueTool UseDomain PoliciesSkill BankSelf-evolution LoopPass@kAgentic LearningCAR-bench
Authors
Wenhao Wu, Menghao Zhang, Xin Wang, Zhi Wang, Kun Shao, Jian Luan
Abstract
Reliable deployment of LLM agents in user-facing products depends not on raw task-solving ability but on consistency and limit-awareness: behaving the same way across repeated trials, and recognizing when a request cannot, or cannot yet, be safely fulfilled. CAR-bench exposes this reliability gap in the domain of in-car assistants: an LLM-simulated user issues incomplete or ambiguous requests, requiring the agent to resolve uncertainty through multi-turn dialogue and tool use while strictly adhering to domain policies. Even frontier models show a substantial gap between what they can solve at least once (Pass@3) and what they solve consistently across trials (Pass^k). We bridge this gap with TRACE (TRAjectory-Contrastive Evolution), which iteratively improves a skill-based agent's behavioral knowledge without modifying model weights. This knowledge is organized as a Skill Bank of modular, retrievable skills, each encoding a self-contained set of tool-use rules and behavioral guidelines. TRACE evolves this bank through an agentic self-evolution loop: after each evaluation round, it groups trajectories by the skills invoked and refines each skill by contrasting successful and failed behaviors. The updated bank then guides subsequent rounds, while during deployment the Actor performs state-conditioned skill orchestration at every turn. On GPT-5.5, TRACE improves consistency (Pass^3) by 34.6 points, from 59.9% to 94.5%, while shrinking the gap between potential and reliable performance to just 4.0 points. On the official hidden set, TRACE achieved first place using GPT-5.6-Sol, attaining a Pass^3 score of 70%-a 40% relative improvement over the baseline. These results show that TRACE converts high model potential into stable, consistent performance gain. Project homepage: https://darwin-agent.github.io/Car-bench-TRACE.