Qin Li, Minghe Gao, Zhiqi Ge, Juncheng Li, Richang Hong · ACM International Conference on Multimedia (ACM MM) 2026 · 2026
DOI: 10.1145/3767308.3836163
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Vision-Language Navigation in Continuous Environments (VLN-CE) requires agents to follow natural language instructions in unseen 3D spaces. The dominant training paradigm relies on imitation learning from static expert trajectories, which optimizes action-sequence imitation without modeling the step-wise local intent behind each decision. This leaves agents unable to understand why each action is taken, limiting generalization to unseen environments and hindering purposeful exploration and self-correction. Existing solutions remain limited by unconstrained reasoning representations, expert-confined exploration, and the wholesale discard of failed self-generated trajectories. To address these, we propose INTACT, an intent-centered self-improvement framework that treats step-wise local intent as an explicit, structured, and optimizable variable throughout self-training. At its core, PLIA-CoT decomposes local intent into four schema-constrained dimensions, transforming free-form reasoning into a controllable semantic interface. Building on this representation, our framework forms a closed self-training loop that uses local intent to guide exploration, filter self-generated trajectories, and optimize policy learning from both successful and informative failed experiences. Experiments on R2R-CE and RxR-CE demonstrate that INTACT raises Success Rate from 32.5% to 58.8% over three self-training rounds, outperforming methods with large-scale external data using only 40.5% of their training samples, without any additional human annotation.
No comments yet — start the discussion below.