TraceGraph: Graph-Based Framework for Analyzing Agent Trajectories
A new framework called TraceGraph has been developed by researchers to convert multi-model agent interaction paths into collaborative decision landscapes. By constructing graphs from observable action-observation states derived from combined rollouts prior to model identification, TraceGraph pinpoints effective cores and trap areas. Each rollout is encapsulated by three key events: Access, Trap exposure, and Repair. An examination across five benchmark splits uncovers navigation variations obscured by overall scores, indicating that the splits differ in their incentives for avoiding traps versus recovering from them. Additionally, the framework supports a trap-aware recovery pipeline for SWE-bench, utilizing a runtime detector that activates on states corresponding to historical trap areas.
Key facts
- TraceGraph is a graph-based framework for analyzing agent trajectories.
- It builds graphs over action-observation states from pooled rollouts.
- It identifies productive cores and trap regions.
- Each rollout is summarized with three events: Access, Trap exposure, and Repair.
- Analysis across five benchmark splits reveals navigation differences.
- Splits differ in rewarding trap avoidance versus recovery.
- A trap-aware recovery pipeline for SWE-bench is motivated.
- The pipeline uses a runtime detector on historical trap regions.
Entities
—