Trust Guided Decision Transformer
TGDT uses the model’s own next-state prediction error to decide which rollout context is still trustworthy.
The paper says Decision Transformers degrade on long rollouts when their conditioning context drifts outside the training distribution. TGDT tests recent context suffixes against a conformal threshold from held-out offline data, then lets a frozen critic choose among actions from the trusted suffixes. The authors report gains over vanilla Decision Transformer, reset-based context control, and value-only context selection on D4RL navigation and locomotion tasks. To appear in NeurIPS 2026. ArXiv · AI/CL/LG's note
The paper says Decision Transformers degrade on long rollouts when their conditioning context drifts outside the training distribution. TGDT tests recent context suffixes against a conformal threshold from held-out offline data, then lets a frozen critic choose among actions from the trusted suffixes. The authors report gains over vanilla Decision Transformer, reset-based context control, and value-only context selection on D4RL navigation and locomotion tasks. To appear in NeurIPS 2026. ArXiv · AI/CL/LG's note
score 4