Research map / Vision-language and multimodal / Multimodal LLMs

CodeV: Code with Images for Faithful Visual Reasoning via Tool-Aware Policy Optimization

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual 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.

Proposes CodeV, a code-based visual agent trained with SFT then Tool-Aware Policy Optimization (GRPO with dense judge-based rewards on tool outputs), and a faithfulness protocol showing high accuracy hides unfaithful crop use.

Problem

Agentic vision-language models that "think with images" by calling crop-style tools often reach high final-answer accuracy while their tool calls do not actually capture the queried evidence, and outcome-only RL does not penalize this.

Approach

TAPO augments GRPO with step-level rewards computed by a judge on non-model-generated tool outputs (the question plus the crop), asking whether the crop provides evidence for the question, combined with a final-answer reward; tools are represented as executable Python code in a sandbox.

Key results

• CodeV-7B-RL reaches 84.8 on V*, versus 75.0 for Qwen2.5-VL-7B and 64.4 for GPT-4o.
• On V*, faithful tool-use-and-correct rate is 68.0 for CodeV versus 49.7 for DeepEyes, 34.1 for Pixel-Reasoner, and 7.0 for Thyme-RL.
• CodeV-7B-RL scores 71.8 on MathVista, the best among the compared models, and 33.6 on MathVision-Mini versus 21.4 for Qwen2.5-VL-7B.

Surprising findings

• Thyme-RL-7B uses visual tools in fewer than 10% of rollouts, so its high accuracy is largely not backed by tool use.
• Zero-RL without cold-start SFT improves average performance over Qwen2.5-VL-7B but collapses to text-only reasoning with minimal tool use.

Limitations

• The judge is a static external model (Qwen2.5-VL-32B), adding deployment cost; the paper suggests learning a reward model as future work.
• The judge is tailored to image-centric operations like cropping and does not generalize to richer tool types.
• Faithfulness is evaluated with GPT-4o on crops for V* and HRBench, which only measures whether the target is visible in the crop.

Experimental setup

Datasets: Thyme-SFT Thinklite-70K DeepEyes-47K V* HRBench-4K HRBench-8K MathVista CharXiv MMMU MathVerse MathVision VLMBlinds

Benchmarks: V* HRBench-4K HRBench-8K VLMBlinds MME-Realworld-Lite MathVista CharXiv MMMU MathVerse-Mini MathVision-Mini

Metrics: accuracy faithful tool-use rate tool-output perturbation sensitivity

Baselines: Qwen2.5-VL-7B Pixel-Reasoner-7B Thyme-RL-7B DeepEyes-7B GPT-4o InternVL3-8B ThinkLite-7B

Builds on: GRPO DeepEyes Pixel-Reasoner Thyme Qwen2.5-VL

Closest papers by meaning

• CodeDance: A Dynamic Tool-integrated MLLM for Executable Visual Reasoning (CVPR 2026 · similarity 0.92)
• Diversity Over Frequency: Rethinking Tool Use in Visual Chain-of-Thought Agents (ICML 2026 · similarity 0.90)
• Thinking with Programming Vision: Towards a Unified View for Thinking with Images (CVPR 2026 · similarity 0.90)
• SenseSearch: Empowering Vision-Language Models with High-Resolution Agentic Search-Reasoning via Reinforcement Learning (CVPR 2026 · similarity 0.90)
• ATLAS: Agentic or Latent Visual Reasoning? One Word is Enough for Both (NeurIPS 2026 · similarity 0.90)
• Agent0-VL: Exploring Self-Evolving Agent for Tool-Integrated Vision-Language Reasoning (ICML 2026 · similarity 0.90)