Prune-Then-Plan: Step-Level Calibration for Stable Frontier Exploration in Embodied Question Answering

TL;DR

The paper introduces Prune-Then-Plan, a calibration framework that stabilizes step-level exploration in embodied question answering by pruning implausible options, leading to significant improvements in navigation efficiency and answer accuracy.

Contribution

It presents a novel pruning-based calibration method that reduces frontier oscillations in VLM-driven exploration, enhancing stability and performance in EQA tasks.

Findings

Achieves up to 49% improvement in SPL metric.

Improves scene coverage under fixed exploration budgets.

Reduces frontier oscillations and unstable movements.

Abstract

Large vision-language models (VLMs) have improved embodied question answering (EQA) agents by providing strong semantic priors for open-vocabulary reasoning. However, when used directly for step-level exploration, VLMs often exhibit frontier oscillations, unstable back-and-forth movements caused by overconfidence and miscalibration, leading to inefficient navigation and degraded answer quality. We propose Prune-Then-Plan, a simple and effective framework that stabilizes exploration through step-level calibration. Instead of trusting raw VLM scores, our method prunes implausible frontier choices using a Holm-Bonferroni inspired pruning procedure and then delegates final decisions to a coverage-based planner. This separation converts overconfident predictions into conservative, interpretable actions by relying on human-level judgments to calibrate the step-level behavior of VLMs. Integrated into the 3D-Mem EQA framework, our approach achieves relative improvements of up to 49% and 33% in visually grounded SPL and LLM-Match metrics respectively over baselines. Overall, our method achieves better scene coverage under equal exploration budgets on both OpenEQA and EXPRESS-Bench datasets.