Home | Publications | ZDW+26

ReEXplore: Improving MLLMs for Embodied Exploration With Contextualized Retrospective Experience Replay

MCML Authors

Link to Profile Volker Tresp

Volker Tresp

Prof. Dr.

Core PI

Database Systems, Data Mining and AI

Abstract

Embodied exploration requires MLLM agents to navigate unfamiliar environments through fine-grained perception and knowledgedriven decision-making. While MLLMs offer strong perceptual and reasoning capabilities, MLLM-based agents remain suboptimal due to three limitations: (i) reliance on static pre-trained knowledge that is never updated through interaction, (ii) the prohibitive cost of training-based approaches for long-horizon tasks with sparse rewards, and (iii) the large, visually similar action spaces produced by frontier-based exploration that make reliable decision-making difficult. We propose ReEXplore, a training-free framework that addresses these limitations through two tightly coupled components: Retrospective Experience Abstraction and Replay, which distills completed trajectories into transferable strategic priors injected at inference time, and Hierarchical Frontier Selection, which decomposes frontier ranking into coarse-to-fine decisions for stable and efficient exploration. Experiments on OpenEQA and GOAT-Bench show that ReEXplore improves answer quality and navigation efficiency on A-EQA under both an open-source and a commercial backbone, and substantially improves navigation efficiency with an open-source backbone on GOAT-Bench. Our code will be made publicly available.

inproceedings ZDW+26

EMR @ECCV 2026

Workshop on Embodied Multimodal Reasoning in Physical Environments at the 19th European Conference on Computer Vision. Malmö, Sweden, Sep 08-12, 2026.

Authors

G. Zhang • M. Ding • J. Wu • R. Liao • V. Tresp

Links

URL

Research Areas

 A3 | Computational Models

 B3 | Multimodal Perception

BibTeXKey: ZDW+26

#p-kasneci-enkelejda #p-tresp
Back to Top