Using Grounded Theory for Agent Behavior Analysis at Scale
AutoTraceGT turns agent traces into task-specific failure taxonomies without starting from fixed labels.
The paper applies grounded theory to large sets of agent trajectories, using iterative open, axial, and theoretical coding until saturation. Across six corpora, its generated codebooks recovered 73-91 percent of failure modes from human taxonomies and found additional patterns those missed. The authors also report that the codebooks beat zero-shot and few-shot LLM baselines for downstream failure prediction. HF Daily Papers' note
The paper applies grounded theory to large sets of agent trajectories, using iterative open, axial, and theoretical coding until saturation. Across six corpora, its generated codebooks recovered 73-91 percent of failure modes from human taxonomies and found additional patterns those missed. The authors also report that the codebooks beat zero-shot and few-shot LLM baselines for downstream failure prediction. HF Daily Papers' note
score 5