Research map / Reinforcement learning / Visual and representation RL
What Does Vision Tool-Use Reinforcement Learning Really Learn? Disentangling Tool-Induced and Intrinsic Effects for Crop-and-Zoom
Yan Ma, Weiyu Zhang, Tianle Li, Du Linge, Xuyang Shen, Pengfei Liu
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 |
|---|---|---|---|
| show the tool schema but forbid execution (vs plain tool-free prompt) | Qwen2.5-VL-7B · Average of 6 benchmarks | accuracy: 48.4 -> 42.6 (-5.8) | Hurt |
| show the tool schema but forbid execution (vs plain tool-free prompt) | Qwen3-VL-8B · Average of 6 benchmarks | accuracy: 53.0 -> 40.0 (-13.0) | Hurt |
| show the tool schema but forbid execution (vs plain tool-free prompt) | Qwen3-VL-8B · VStar | accuracy: 82.7 -> 57.1 | Hurt |
| allow tool execution (vs tool-free inference) | Qwen2.5-VL-7B · Average of 6 benchmarks | accuracy: 48.4 -> 42.2 | Hurt |
| allow tool execution (vs tool-free inference) | Qwen3-VL-8B · Average of 6 benchmarks | accuracy: 53.0 -> 61.2 | Improved |
| allow tool execution (vs tool-free inference) | Qwen2.5-VL-7B · VisualProbe Easy | accuracy: 39.0 -> 21.3 | Hurt |
| allow tool execution (vs tool-free inference) | Qwen3-VL-8B · VisualProbe Easy | accuracy: 41.8 -> 56.7 | Improved |
| train with tools removed during RL (no-tool RL vs tool-RL), tool-free accuracy (ablation) | Qwen2.5-VL-7B · final Acc_wo | accuracy: 0.644 vs 0.636 | No clear effect |
| train with tools removed during RL (no-tool RL vs tool-RL), tool-free accuracy (ablation) | Qwen3-VL-8B · final Acc_wo | accuracy: 0.678 vs 0.681 | No clear effect |
| add +0.1 tool-call bonus to the reward (ablation) | Qwen2.5-VL-7B · MED terms | Call Gain and Call Harm: +0.1 bonus for correct responses that call the tool | Mixed |
| remove tool access in RL training (ablation) | Qwen2.5-VL-7B · MED terms | Call Gain and Call Harm: both call-side terms close to zero | Hurt |
Closest papers by meaning
- Learning to Select Visual Tools from Experience (CVPR 2026 · similarity 0.88)
- ParaVT: Taming the Tool Prior Paradox for Parallel Tool Use in Agentic Video Reinforcement Learning (NeurIPS 2026 · similarity 0.87)
- SenseSearch: Empowering Vision-Language Models with High-Resolution Agentic Search-Reasoning via Reinforcement Learning (CVPR 2026 · similarity 0.87)
- ViPER: Empowering the Self-Evolution of Visual Perception Abilities in Vision-Language Models (ICLR 2026 · similarity 0.86)
- Diversity Over Frequency: Rethinking Tool Use in Visual Chain-of-Thought Agents (ICML 2026 · similarity 0.86)
- GUI-Eyes: Tool-Augmented Perception for Visual Grounding in GUI Agents (AAAI 2026 · similarity 0.86)
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.
