Research map / Vision-language and multimodal / Multimodal LLMs

Active Exploring like a Pigeon: Reinforcing Spatial Reasoning via Agentic Vision-Language Models

ICML 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 an agentic VLM pipeline that actively retrieves views, maintains a dynamic cognitive map, and emits Spatial Assertion Code (SAC) whose execution gives dense rewards for GRPO-based RL on spatial reasoning, reaching 80.5% on MindCube-Tiny.

Problem

VLMs act as passive observers for spatial reasoning over multi-view images, and RL for them relies on sparse final-answer rewards, which limits performance on complex spatial tasks.

Approach

A dynamic top-down cognitive map serves as persistent memory updated through iterative active view retrieval, and model-generated Python assertions (SAC) executed against the map provide dense, verifiable intermediate rewards for RL.

Key results

• Achieves 80.5% overall accuracy on MindCube-Tiny, 5.3 points above the prior best 3DThinker (75.2%).
• On the Rotation subset, reaches 85.0%, a 29.5-point gain over the best prior method (55.5%).
• Ablation: removing the SAC reward drops accuracy from 80.4 to 70.2; SFT alone reaches 67.2 pass@1 versus 80.5 for SFT+RL.

Surprising findings

• RL applied directly to the base model without SFT yields only 54.2 pass@1, showing SFT is needed as a cold start.
• Among samples with positional errors in the predicted cognitive map, more than half still answer correctly, and 15.7% of correct predictions still have map errors.

Limitations

• Performs poorly on EmbodiedBench Common and Complex settings without an added language model (0.0 on both), attributed to task mismatch.
• Relies on dataset metadata for part of the retrieval reward; removing the ground-truth answer from it drops accuracy by 7.6 points.
• SFT data for SAC reasoning is generated by a proprietary VLM, adding dependence on closed models.

Experimental setup

Datasets: MindCube

Benchmarks: MindCube-Tiny EmbodiedBench

Metrics: accuracy macro precision macro recall F1 pass@k

Baselines: MindCube 3DThinker Spatial-MLLM Space-Qwen Qwen2.5-VL-7B-Instruct Qwen2.5-VL-3B-Instruct GPT-4o Claude-4-Sonnet-20250514

Builds on: GRPO MindCube veRL LLaMA-Factory

Closest papers by meaning

• SpatialClaw: Rethinking Action Interface for Agentic Spatial Reasoning (NeurIPS 2026 · similarity 0.92)
• Thinking with Imagination: Agentic Visual Spatial Reasoning with World Simulators (NeurIPS 2026 · similarity 0.90)
• Perceive, Interact, Reason: Building Tool-Augmented Visual Agents for Spatial Reasoning (NeurIPS 2026 · similarity 0.90)
• EvoEmpirBench: Dynamic Spatial Reasoning with Agent-ExpVer (AAAI 2026 · similarity 0.90)
• Visual Agentic AI for Spatial Reasoning with a Dynamic API (CVPR 2025 · similarity 0.90)
• SpatialLadder: Progressive Training for Spatial Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.89)