Shanghai University

Agentic reinforcement learning shifts the reasoning trajectories of large language models, and trajectory-based confidence estimators trained on a base model lose much of their ranking ability after that shift. StaR-TTA treats trajectory stability as a label-free reliability signal: it learns the link between stability and correctness on labeled base trajectories, then uses stability rankings from unlabeled RL trajectories to correct confidence at test time. Across four QA benchmarks it reaches an average AUROC of 0.7208, with gains on unseen datasets, different model families and scales, and both PPO and GRPO.

Agentic reinforcement learning shifts the reasoning trajectories of large language models, and trajectory-based confidence estimators trained on a base model lose much of their ranking ability after that shift. StaR-TTA treats trajectory stability as a label-free reliability signal: it learns the link between stability and correctness on labeled base trajectories, then uses stability rankings from unlabeled RL trajectories to correct confidence at test time. Across four QA benchmarks it reaches an average AUROC of 0.7208, with gains on unseen datasets, different model families and scales, and both PPO and GRPO.

Subway fire risk assessment has long depended on experts manually choosing Bayesian network nodes, which is subjective and hard to adapt. This pipeline extracts 215 fire keywords and six risk labels from accident records and fire codes with natural language processing, builds a 666-node knowledge graph, classifies nodes with a three-layer graph convolutional network, and combines confidence weighting with K-means clustering to select 22 representative nodes for the Bayesian network. Under coupled risks such as malicious passenger behavior, smoke exhaust failure and delayed alarm, the high-casualty probability rises from a 6.58% baseline to 53.2%.

Subway fire risk assessment has long depended on experts manually choosing Bayesian network nodes, which is subjective and hard to adapt. This pipeline extracts 215 fire keywords and six risk labels from accident records and fire codes with natural language processing, builds a 666-node knowledge graph, classifies nodes with a three-layer graph convolutional network, and combines confidence weighting with K-means clustering to select 22 representative nodes for the Bayesian network. Under coupled risks such as malicious passenger behavior, smoke exhaust failure and delayed alarm, the high-casualty probability rises from a 6.58% baseline to 53.2%.