Research map / Vision-language and multimodal / Multimodal LLMs

Thinking with Video: Video Generation as a Promising Multimodal Reasoning Paradigm

CVPR 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.

Proposes "Thinking with Video" and VideoThinkBench to evaluate video generation models (mainly Sora-2) on vision-centric and text-centric reasoning versus SOTA VLMs.

Problem

Existing reasoning paradigms (Thinking with Text and Thinking with Images) cannot represent dynamic processes and separate text from vision; it is unclear whether video generation models can serve as general multimodal reasoners.

Approach

Reframes video generation as a reasoning medium ("Thinking with Video") and builds a benchmark of verifiable vision-centric puzzles plus adapted text/multimodal reasoning tasks, systematically comparing Sora-2 against VLMs and analyzing sources of its text-centric ability (e.g., a prompt rewriter).

Key results

• Sora-2 (Major Frame) reaches 40.2% on eyeballing puzzles, above Claude Sonnet 4.5 (35.1%), GPT-5 high (29.7%) and Gemini 2.5 Pro (26.5%).
• Sora-2 audio accuracy is 98.9% on GSM8K and 92.0% on MATH-500, but only 1.3% on ARC-AGI-2.
• Majority vote over 5 Sora-2 videos raises Arc Connect accuracy from 68% (single major-frame) to 90%.

Surprising findings

• Wan 2.5 achieves nearly zero text reasoning accuracy without its prompt rewriter, suggesting Sora-2's text reasoning may come from an internal prompt rewriter.
• Sora-2 solves only 13.91% of text-centric solutions fully correctly in the written video process, despite high audio accuracy.

Limitations

• Only Sora-2 studied in depth; it is closed-source so internal mechanisms cannot be analyzed.
• Text-centric tasks use subsets and video text rendering is error-prone.
• Some results are manually evaluated and small sample sizes on mini-sets.

Experimental setup

Datasets: GSM8K MATH-500 MMLU MathVista MMMU ARC-AGI-2 PuzzleVQA BBH MMLU-Pro GPQA SuperGPQA MathVision

Benchmarks: VideoThinkBench ARC-AGI-2 GSM8K MATH-500 MMLU MathVista MMMU

Metrics: accuracy pixel accuracy

Baselines: GPT-5 high Claude Sonnet 4.5 Gemini 2.5 Pro Veo 3.1 Wan 2.5 Nano Banana Pro Qwen3-VL-235B-A22B

Builds on: Chain-of-Thought Thinking with Images PuzzleVQA ARC-AGI-2

Closest papers by meaning

• ThinkMorph: Emergent Properties in Multimodal Interleaved Chain-of-Thought Reasoning (ICLR 2026 · similarity 0.90)
• AStar: Boosting Multimodal Reasoning with Automated Structured Thinking (AAAI 2026 · similarity 0.90)
• When Visualizing is the First Step to Reasoning: MIRA, a Benchmark for Visual Chain-of-Thought (CVPR 2026 · similarity 0.89)
• Video-of-Thought: Step-by-Step Video Reasoning from Perception to Cognition (ICML 2024 · similarity 0.89)
• ROVER: Benchmarking Reciprocal Cross-Modal Reasoning for Omnimodal Generation (ICLR 2026 · similarity 0.89)
• SciVideoBench: Benchmarking Scientific Video Reasoning in Large Multimodal Models (ICML 2026 · similarity 0.89)