Research map / Large language models / Reasoning and chain of thought

VisionLeaf: Entropy-Guided Leaf-First Reasoning for Efficient and Accurate Think-with-Image

CVPR 2026 · Large language models › Reasoning and chain of thought · approach: Long CoT dynamics and efficiency · 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 VisionLeaf, an entropy-guided tree-based GRPO variant that branches from the highest-entropy leaf nodes during rollouts to train think-with-image agents more stably and efficiently.

Problem

Think-with-image agents trained with vanilla GRPO suffer from training instability and redundant, excessive tool calls at inference, because terminal-only rewards give early multi-turn steps high-variance supervision and rollouts remain single-path.

Approach

Instead of expanding every rollout from the root with one branch per leaf, grow a rollout tree from leaf nodes: repeatedly pick the active node with the highest next-token entropy and spawn multiple continuation branches beneath it, with tool-return tokens masked from the loss.

Key results

• VisionLeaf reaches 84.8% overall on V* versus 80.6% for DeepEyes, a 4.2-point gain with the same data and base model.
• On HR-Bench 8K, VisionLeaf scores 72.1% versus 68.8% for DeepEyes.
• VisionLeaf uses roughly half as many inference tool-call rounds as DeepEyes across benchmarks (Fig. 3).

Surprising findings

• Splitting on low-entropy nodes performs worse than random splitting, indicating high-entropy node selection matters for exploration.

Limitations

• Entropy is approximated by min-entropy on the first token of each turn rather than full-distribution entropy.
• Evaluation relies on an LLM-as-a-judge protocol inherited from DeepEyes.
• Results are reported only for a single 7B base model and one training dataset.

Experimental setup

Datasets: DeepEyes-Datasets-47k

Benchmarks: V*Bench MME-RealWorld-lite HR-Bench 4K HR-Bench 8K

Metrics: accuracy number of tool calls / interaction turns pass@16

Baselines: DeepEyes Pixel Reasoner SEAL DyFo PaLI-X-VPD LLaVA-OV Qwen2.5-VL GPT-4o o3

Builds on: GRPO DeepEyes VeRL

Closest papers by meaning

• MiCo: Multi-image Contrast for Reinforcement Visual Reasoning (NeurIPS 2025 · similarity 0.88)
• TVI-CoT: Text-Visual Interleaved Chain-of-Thought Reasoning for Multimodal Understanding (ICML 2026 · similarity 0.88)
• Reflection Anchors for Propagation-Aware Visual Retention in Long-Chain Multimodal Reasoning (NeurIPS 2026 · similarity 0.88)
• VLM-R³: Region Recognition, Reasoning, and Refinement for Enhanced Multimodal Chain-of-Thought (NeurIPS 2025 · similarity 0.88)
• VGR: Visual Grounded Reasoning (ICLR 2026 · similarity 0.88)
• Multi-step Visual Reasoning with Visual Tokens Scaling and Verification (NeurIPS 2025 · similarity 0.88)