Research map / Reinforcement learning / RL from feedback
GTR: Guided Thought Reinforcement Prevents Thought Collapse in RL-based VLM Agent Training
Tong Wei, Yijun Yang, Junliang Xing, Yuanchun Shi, Zongqing Lu, Deheng Ye
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 process guidance via VLM corrector (GTR) over outcome-only RL | RL4VLM (PPO) · gym_cards EZPoints | success rate (%): 48.0 -> 94.5 | Improved |
| add GTR over RL4VLM | LLaVA-v1.6-mistral-7B · gym_cards Numberline | success rate (%): 90.5 -> 100.0 | Improved |
| add GTR over RL4VLM | LLaVA-v1.6-mistral-7B · gym_cards Blackjack | success rate (%): 40.1 -> 41.3 | Improved |
| add GTR over RL4VLM | LLaVA-v1.6-mistral-7B · gym_cards EZPoints | episode return: 4.19 -> 9.43 | Improved |
| add GTR over RL4VLM, no text observation | LLaVA-v1.6-mistral-7B · ALFWorld (without text) | success rate (%): 5.4 -> 17.8 | Improved |
| add text description of observation to RL4VLM | RL4VLM · ALFWorld | success rate (%): 5.4 -> 21.7 | Improved |
| add text description of observation to GTR | GTR · ALFWorld | success rate (%): 17.8 -> 21.0 | Improved |
| use GPT-4o instead of Qwen2.5-VL-72B as corrector (ablation) | GTR · Points24 | performance (%): 6.5 -> 17.5 | Improved |
| use GPT-4o instead of Qwen2.5-VL-7B as corrector (ablation) | GTR · Points24 | performance: N/A -> 17.5 | Improved |
| keep thought-cloning loss throughout training instead of cosine annealing it (ablation) | GTR · Points24 | training curve | Improved |
| add tool function calling to GPT-4o corrector (ablation) | GTR · Points24 | training curve | Improved |
| add DAgger aggregation of thought dataset (ablation) | GTR · Points24 | training curve | Improved |
Closest papers by meaning
- Fine-Tuning Large Vision-Language Models as Decision-Making Agents via Reinforcement Learning (NeurIPS 2024 · similarity 0.91)
- Think or Not? Selective Reasoning via Reinforcement Learning for Vision-Language Models (NeurIPS 2025 · similarity 0.90)
- GFlowVLM: Enhancing Multi-step Reasoning in Vision-Language Models with Generative Flow Networks (CVPR 2025 · similarity 0.90)
- All Roads Lead to Rome: Incentivizing Divergent Thinking in Vision-Language Models (CVPR 2026 · similarity 0.90)
- On Robustness and Chain-of-Thought Consistency of RL-Finetuned VLMs (ICML 2026 · similarity 0.90)
- GoalLadder: Incremental Goal Discovery with Vision-Language Models (NeurIPS 2025 · 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.
