From Noisy Traces to Root Causes: Structural Trajectory Analysis and Causal Extraction for Agent Optimization
STRACE tries to make agent self-optimization less dependent on noisy execution logs.
The paper proposes filtering failures at two levels: first across batches of traces, then inside individual trajectories. It keeps representative failure cases, builds a textual dependency graph, and removes steps judged non-causal. The goal is to point the optimizer at the module that actually caused the failure, rather than feeding it redundant or misleading context. In tests, the authors report STRACE beating standard context-filtering baselines, including a VeruSAGE-Bench result that raised success rate from 42.5% to 58.5%. HF Daily Papers' note
The paper proposes filtering failures at two levels: first across batches of traces, then inside individual trajectories. It keeps representative failure cases, builds a textual dependency graph, and removes steps judged non-causal. The goal is to point the optimizer at the module that actually caused the failure, rather than feeding it redundant or misleading context. In tests, the authors report STRACE beating standard context-filtering baselines, including a VeruSAGE-Bench result that raised success rate from 42.5% to 58.5%. HF Daily Papers' note
score 4