Research map / Vision-language and multimodal / Multimodal LLMs

Think 360deg: Beyond Depth: Evaluating the Width-centric Reasoning Capability of MLLMs

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning

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.

Introduces Think360°, a 1,225-question multimodal benchmark targeting reasoning width (trial-and-error, branch-and-bound, divide-and-conquer, hypothesize-and-test) alongside depth, with a tree-of-thought evaluation protocol, and evaluates 30+ MLLMs.

Problem

Existing multimodal reasoning benchmarks mostly probe reasoning depth (long sequential chains), while reasoning width (broad exploration, pruning, backtracking under multiple constraints) is under-evaluated for MLLMs.

Approach

Distinguish reasoning width from depth, curate width-centric multimodal problems from competitions, textbooks, benchmarks and puzzle games, and score model responses by building a tree of reasoning steps and computing separate width and depth accuracies.

Key results

• Gemini-2.5-pro achieves the top overall accuracy of 46.0% on Think360, followed by o3 (42.3%) and o4-mini (42.1%).
• ToT-Width correlates more strongly with final accuracy (Spearman ρ=0.9218) than ToT-Depth (ρ=0.8492).
• MiMo-VL-RL-7B reaches 28.3% accuracy, surpassing Claude-4-Sonnet's 28.2%, while Qwen2.5-VL-Instruct-72B reaches only 16.2%.

Surprising findings

• Open-source MiMo-VL-RL-7B outperforms several closed-source models, including GPT-4.1 and Claude-4-Sonnet, on this benchmark.
• Qwen2.5-VL-Instruct-72B scores lower than the 32B variant (16.2% vs 17.6%), and smaller open models sometimes gain nothing or lose accuracy from CoT prompting.

Limitations

• Tree-of-thought scoring relies on GPT-4o for tree construction and node judging, so scores depend on judge reliability.
• Only models exposing intermediate reasoning can be evaluated with ToT; o3 and Gemini-2.5-pro excluded from ToT analysis.
• Benchmark is small (1,225 questions) with many free-form answers, and many questions were rewritten from proofs into verifiable forms.

Experimental setup

Datasets: ARML HMMT textbook exercises MathVision DynaMath MME-Reasoning MMMR RBench-V VisualSphinx online puzzle games

Benchmarks: Think360

Metrics: pass@1 accuracy ToT-Width ToT-Depth reasoning time token cost

Baselines: GPT-4o GPT-4.1 o1 o3 o4-mini Gemini-2.5-pro Claude-4-Opus Claude-3.7-Sonnet-Thinking MiMo-VL-RL-7B Qwen2.5-VL-Instruct

Builds on: MathVista MathVerse MathVision chain-of-thought

Closest papers by meaning

• MDK12-Bench: A Multi-Discipline Benchmark for Evaluating Reasoning in Multimodal Large Language Models (AAAI 2026 · similarity 0.90)
• MME-CoT: Benchmarking Chain-of-Thought in Large Multimodal Models for Reasoning Quality, Robustness, and Efficiency (ICML 2025 · similarity 0.88)
• Insight-V: Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models (CVPR 2025 · similarity 0.88)
• Think Deep, Not Just Long: Measuring LLM Reasoning Effort via Deep-Thinking Tokens (ICML 2026 · similarity 0.88)
• Hint Tuning: Less Data Makes Better Reasoners (NeurIPS 2026 · similarity 0.88)
• AStar: Boosting Multimodal Reasoning with Automated Structured Thinking (AAAI 2026 · similarity 0.88)