Research map / Vision-language and multimodal / Robot action and navigation
Recurrent Reasoning with Vision-Language Models for Estimating Long-Horizon Embodied Task Progress
Yuelin Zhang, Sijie Cheng, Chen Li, Zongzhao Li, Yuxin Huang, Yang Liu, Wenbing Huang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add CoT SFT before RL (instead of direct RL) (ablation) | R2VLM · ALFRED progress estimation | progress MAE: 7.21 -> 2.19 | Improved |
| continue with RL after SFT (ablation) | R2VLM · ALFRED progress estimation | task completion accuracy: 0.961 -> 0.988 | Improved |
| continue with RL after SFT (ablation) | R2VLM · Ego4D progress estimation | progress MAE: 23.03 -> 19.25 | Improved |
| add CoT SFT training to Qwen2.5-VL-7B | Qwen2.5-VL-7B · ALFRED progress estimation | progress MAE: 27.87 -> 2.77 | Improved |
| add improvement reward (ablation) | R2VLM RL · ALFRED progress estimation | progress MAE: 2.42 -> 2.19 | Improved |
| add MAE reward (ablation) | R2VLM RL · ALFRED progress estimation | progress MAE / bin: bin 0.924 -> 0.917; Delta pmae 2.23 -> 2.17 | Mixed |
| add bin reward (ablation) | R2VLM RL · ALFRED progress estimation | progress MAE: 2.22 -> 2.19 | Improved |
| snippet plus historical CoT instead of full video input (ablation) | R2VLM · Ego4D progress estimation | MAE / inference time | Improved |
| use R2VLM progress instead of time-based progress | Seq2Seq · ALFRED valid seen policy learning | goal-condition completion (%): 8.7 -> 10.9 | Improved |
| use R2VLM progress on unseen split | Seq2Seq · ALFRED valid unseen policy learning | goal-condition completion (%): 6.8 -> 7.3 | No clear effect |
| use time-based progress supervision | Seq2Seq · ALFRED valid seen policy learning | goal-condition completion (%): 9.4 -> 8.7 | Hurt |
| use R2VLM dense reward instead of sparse step reward | SPRINT online RL · EVALINSTRUCT online RL (overall) | task completion rate: 0.42 -> 0.50 | Improved |
Closest papers by meaning
- ROVER: Recursive Reasoning Over Videos with Vision-Language Models for Embodied Tasks (NeurIPS 2025 · similarity 0.91)
- SOLE-R1: Video-Language Reasoning as the Sole Reward for On-Robot Reinforcement Learning (NeurIPS 2026 · similarity 0.90)
- ReWatch-R1: Boosting Complex Video Reasoning in Large Vision-Language Models through Agentic Data Synthesis (ICLR 2026 · similarity 0.90)
- ThinkAct: Vision-Language-Action Reasoning via Reinforced Visual Latent Planning (NeurIPS 2025 · similarity 0.89)
- Unveiling Chain of Step Reasoning for Vision-Language Models with Fine-grained Rewards (NeurIPS 2025 · similarity 0.89)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
