Research map / Vision-language and multimodal / Multimodal LLMs

Diversity Over Frequency: Rethinking Tool Use in Visual Chain-of-Thought Agents

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Studies tool use in visual chain-of-thought agents on 3D spatial reasoning (3DSRBench) and finds tool-use collapse under vanilla RFT, then shows adaptive entropy regularization gives the best accuracy while tool use fades.

Problem

Visual agents that invoke zoom-in tools during reasoning are mostly studied on visual search; it is unclear how tool use behaves and helps in harder tasks like 3D spatial reasoning under RL fine-tuning, where tool use is optional rather than mandatory.

Approach

Vanilla RFT and tool-encouraging rewards collapse rollout diversity (textual n-gram and visual crop locations), so tools act as training-time scaffolding; adding an adaptive entropy regularization term to GRPO sustains exploration and yields better final accuracy even as tool use declines.

Key results

• Entropy-regularized RFT reaches 62.9% on 3DSRBench vs 59.2% for vanilla RFT, with tool use declining from ~20% to ~3%.
• Tool-banned training yields 58.1%, while forced tool use (tool bonus) reaches 59.9% with tool use rising to 100%.
• On VQA-RAD, entropy regularization gives 48.78% vs 46.34% for vanilla RFT and 47.23% for tool-encouraged RFT.

Surprising findings

• Vanilla RFT drives tool use from ~20% to near 2% while accuracy still improves, a tool-use collapse.
• Banning tools entirely from the start hurts accuracy, yet forcing tool use to 100% gives only marginal gains.

Limitations

• Experiments are limited to a single base model (Mini-o3 / Qwen2.5-VL-7B) and a single zoom-in tool in the main setting.
• Training runs are capped at 100 steps due to oscillation and DAPO dynamic sampling costs.
• Evaluation relies on a VLM-as-judge (Qwen2.5-VL-7B) with 4.4% human mismatch on a sample.

Experimental setup

Datasets: SpatialReasoner training subset (1.2k) SAT (6k) OpenImages VQA-RAD

Benchmarks: 3DSRBench CV-Bench-3D VQA-RAD

Metrics: Avg@8 accuracy tool use ratio n-gram distinct ratio mIoU between crop boxes CLIP score

Baselines: Qwen2.5-VL 7B DeepEyes Mini-o3 SpatialReasoner Tool-banned RFT Tool-Encourage reward (tool bonus) Curiosity reward

Builds on: Mini-o3 DAPO GRPO DeepEyes PixelReasoner

Closest papers by meaning

• CodeV: Code with Images for Faithful Visual Reasoning via Tool-Aware Policy Optimization (CVPR 2026 · similarity 0.90)
• AdaReasoner: Dynamic Tool Orchestration for Iterative Visual Reasoning (ICLR 2026 · similarity 0.89)
• Don't Show Pixels, Show Cues: Unlocking Visual Tool Reasoning in Language Models via Perception Programs (CVPR 2026 · similarity 0.89)
• CodeDance: A Dynamic Tool-integrated MLLM for Executable Visual Reasoning (CVPR 2026 · similarity 0.89)
• SenseSearch: Empowering Vision-Language Models with High-Resolution Agentic Search-Reasoning via Reinforcement Learning (CVPR 2026 · similarity 0.89)
• Leveraging Latent Visual Reasoning in Silence (NeurIPS 2026 · similarity 0.89)