Search papers, labs, and topics across Lattice.
The paper introduces DIVERT, a novel evaluation framework for LLM-based agents that uses a snapshot-based, coverage-guided user simulation to efficiently explore agent-user interactions. DIVERT captures agent-environment states at decision points, resumes execution from these snapshots, and branches using diversity-inducing user responses to explore alternative interaction paths. Experiments demonstrate that DIVERT discovers more failures per token and expands the set of tasks on which failures are identified compared to standard linear rollout protocols.
Uncover more LLM agent failures, faster: DIVERT's diversity-guided user simulation finds more bugs per token than standard rollout methods.
Large language models (LLMs) are increasingly deployed as customer-facing agents, yet evaluating their reliability remains challenging due to stochastic, multi-turn interactions. Current evaluation protocols rely on linear Monte Carlo rollouts of complete agent-user conversations to estimate success. However, this approach is computationally inefficient, repeatedly regenerating identical early prefixes, and often fails to uncover deep failure modes that arise from rare user behaviors. We introduce DIVERT (Diversity-Induced Evaluation via Branching of Trajectories), an efficient, snapshot-based, coverage-guided user simulation framework for systematic exploration of agent-user interactions. DIVERT captures the full agent-environment state at critical decision points and resumes execution from these snapshots, enabling reuse of shared conversation prefixes and reducing redundant computation. From each junction, the framework branches using targeted, diversity-inducing user responses, allowing directed exploration of alternative interaction paths. By focusing evaluation on semantically diverse and underexplored trajectories, DIVERT improves both efficiency and coverage. Empirical results show that it discovers more failures per token compared to standard linear rollout protocols, while expanding the set of tasks on which failures are identified.
