ENCP: Episode-Normalized Conformal Prediction for Vision-and-Language Navigation

Forum topic · 小凯 · 2026-09-17

Summary

This paper introduces Episode-Normalized Conformal Prediction (ENCP), a method for uncertainty estimation in Vision-Language-Navigation (VLN) models. Uncertainty estimation matters for VLN because it helps identify ambiguous, unreliable predictions so agents can make safer navigation decisions. Conformal prediction (CP) is a leading uncertainty framework, but standard CP calibration fails to deliver its promised coverage over dependent, variable-length VLN episodes, since a VLN agent takes a sequence of steps. ENCP addresses this by rescaling nonconformity scores using the policy's residual confidence and calibrating one maximum score per episode. Under exchangeable calibration and test episodes, this construction covers the ground truth at each step with probability at least 1−α while permitting dependencies between steps within an episode. Experiments on R2R and REVERIE datasets across 4 VLN policies and 3 nonconformity scores show ENCP meets all reported empirical per-step coverage targets in both seen and unseen environments. Authors: Vicky Feliren, A. Taufiq Asyhari, Muhamad Risqi U. Saputra (arXiv:2609.17499).

Paper Overview

Field: Machine Learning Authors: Vicky Feliren, A. Taufiq Asyhari, Muhamad Risqi U. Saputra Published: 2026-09-15 arXiv: 2609.17499

Abstract

Uncertainty estimation for Vision-Language-Navigation (VLN) models is a critical task since it can help identify ambiguous and unreliable predictions, enabling agents to make safer navigation decisions. As one of the most advanced uncertainty estimation frameworks, conformal prediction (CP) offers a promising approach for uncertainty estimation in VLN. However, given that a VLN agent requires a sequence of steps, standard calibration in conformal prediction fails to provide the coverage guarantee it promises over a dependent, variable-length VLN episode.

To this end, the authors propose Episode-Normalized Conformal Prediction (ENCP), which rescales a nonconformity score by the policy's residual confidence and calibrates one maximum score per episode. Under exchangeable calibration and test episodes, this construction covers the ground truth at each step with probability at least 1−α, while allowing dependencies between steps within an episode.

Key Results

Evaluated on the R2R and REVERIE datasets.
Tested across 4 VLN policies and 3 nonconformity scores.
ENCP achieves all reported empirical per-step coverage targets in both seen and unseen evaluation settings.

Implications

These results indicate that ENCP can provide model-agnostic uncertainty estimation for VLN, potentially helping determine when a VLN agent should hand over to a stronger predictor, including human assistance.

---

*Auto-collected on 2026-09-17.*