TRACER: Physics-Guided Causal Evidence Construction for Zero-Shot Traffic Anomaly Diagnosis

Identifying the root causes of non-recurrent traffic congestion is critical for maintaining urban mobility and resilience. However, accurate diagnosis in large-scale networks remains a formidable challenge. Existing deep learning paradigms often prioritize pattern reconstruction over causal identifiability, leading to localization errors due to spatial smoothing effects, while traditional causal discovery algorithms face prohibitive computational barriers and lack physical flow guarantees. Although Large Language Models (LLMs) offer promising neuro-symbolic reasoning capabilities, their direct application to physical traffic networks is hindered by spatial hallucinations, high token costs, and prohibitive inference latency. To tackle these challenges, we present TRACER (Traffic Root Anomaly Causal Evidence Reasoner), a training-free agent framework for zero-shot traffic diagnosis. Instead of feeding raw data directly to the model, TRACER employs a Physics-Grounded Evidence Construction mechanism, synthesizing kinematic shockwave tracing with statistical verification to isolate reliable causal chains. These chains are processed via a Parallel Reasoning and Consensus Fusion architecture, which decouples complex global inferences to accelerate diagnosis and ensure robust localization. Extensive experiments on SUMO simulations and the PeMS-BAY dataset demonstrate that TRACER achieves a 32.6% relative improvement in Hit@1 accuracy over state-of-the-art baselines, while reducing inference time by 85.7%. Furthermore, closed-loop evaluations confirm that interventions guided by our agent significantly accelerate traffic recovery.