Reproduce “Recurrent Reasoning with Vision-Language Models for Estimating Long-Horizon Embodied Task Progress” on OpenResearch

About this paper

Yuelin Zhang, Sijie Cheng, Chen Li, Zongzhao Li, Yuxin Huang, Yang Liu, Wenbing Huang

R²VLM tracks progress through multistep tasks by maintaining an evolving account of completed and remaining steps. At each update, it revises that account using the previous written reasoning and a recent video snippet, then estimates progress. On Ego4D, its 7B model reports progress MAE of 19.25 versus 25.04 for GPT-5, though this benchmark result does not guarantee deployment accuracy.

Key insights

• Processing short video snippets with historical CoT significantly improves computational efficiency and inference time, preventing attention dispersion and maintaining accuracy for long-duration tasks.
• Explicit CoT generation and dynamic updating enable VLMs to perform structured reasoning about temporal dependencies, maintaining global task context and yielding more accurate and interpretable progress estimations.
• Reinforcement learning, particularly with an 'Improvement Reward' incentivizing error reduction from previous turns, is critical for refining the model's reasoning beyond supervised signals and learning generalizable adaptive strategies.

Summary by alphaXiv.