TraceDance builds targeted tests from deployed agent behaviors
TraceDance: An Automated System for Building Agent Behavior Benchmarks from Real-World Agent Deployment Traces
Artificial IntelligenceComputation and Language
Summary
Sometimes computer agents do things we don’t want while trying to complete tasks. The authors created TraceDance, a tool that automatically makes tests focused on those unwanted behaviors by analyzing real-world agent data. It uses smart searches combined with a language model to identify and confirm examples of these behaviors. Then, it measures how well other language models handle these situations. This helps developers find where agents need improvement based on actual problems from deployment.
What this means in practice
- •For ai software developers: Create customized test suites from deployed agent logs to identify and improve problematic behaviors in AI agents.
- •For chatbot developers: Use targeted benchmarks from real conversations to evaluate and address failure points in chatbot decision making.
Authors
Dehai Min, Daoan Zhang, Yiming Zeng, Huayi Zhang, Ziyi Chen, Yan Zhang, Qinbo Bai, Mengyuan Chao, Jing Ning, Qiyue Hua, Huiyi Chen, Hanrong Zhang, Henry Peng Zou, Jie Yang, Wei Xu, Philip S. Yu
Abstract
An agent can complete a task while exhibiting undesirable behavior during execution. Developers need tests for the specific behaviors encountered in deployment, beyond fixed benchmark suites. We present TraceDance, an agent system that constructs targeted benchmarks from deployment traces for user-specified undesirable behaviors. For efficient construction, Anchor-and-Confirm combines programmable retrieval with candidate-level confirmation by a Flash large language model (LLM), while the Anchor Synthesis Loop generates and revises specifications for custom behaviors. The benchmarks use decision-point continuation to evaluate an LLM's next turn at a recorded decision point with a behavior-specific rubric, without a reference answer or environment replay. Experiments in coding and general tool use draw on 252,557 sessions and produce 107 benchmarks with 4,125 instances, fulfilling 95.3% of build-target requests. Both human annotators confirm the requested behavior in 84% of sampled instances, and the automated grader's agreement with human pass/fail judgments is comparable to that between the annotators. Nine frontier LLMs achieve a mean pass rate of only 26.7%, showing that they still struggle to respond appropriately at the evaluated decision points. Analysis across behavior-specific benchmarks further reveals weaknesses in how current LLMs behave as agents. By turning deployment problems into targeted benchmarks, TraceDance could serve as a key component of the recursive self-improvement (RSI) loop.