Research map / Vision-language and multimodal / Multimodal LLMs

ViRC: Enhancing Visual Interleaved Mathematical CoT with Reason Chunking

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 VIRC, which structures multimodal math chain-of-thought into Critical Reasoning Units (Reason Chunking) with crop/scale/display visual tools, builds the 100K CRUX dataset, and trains Qwen2.5-VL-7B/3B through Instructional SFT, Practice SFT, and GRPO-based Strategic RL.

Problem

Existing multimodal LLMs reason from a single static image or interleave visual tokens at every step, which is redundant and ignores hierarchical decomposition of math problem solving.

Approach

Reason Chunking groups reasoning steps into CRUs, each a self-contained intermediate proposition; visual evidence is injected via tool calls only between CRUs, and training uses a progressive curriculum on a CRU-annotated dataset with CRU-level rewards.

Key results

• VIRC-7B reaches 77.79 average on GeoQA, MMStar-Math and MathVista-Math, versus 58.99 for Qwen2.5-VL-7B-Instruct (about 18.8% gain per the paper).
• VIRC-7B reaches 58.27 average on VisualProbe, V*, and HR-Bench generalization, versus 49.25 for Qwen2.5-VL-7B-Instruct.
• VIRC-7B outperforms its teacher Qwen2.5-VL-72B-Instruct on average (77.79 vs 75.12) and uses fewer tokens on average (879) than MINT-CoT-7B (1540).

Surprising findings

• Removing the multimodal coherence reward leads the model to bypass tool use and guess answers, a reward-hacking failure mode.
• Adding single reasoning patterns (backtracking, reflecting, or verifying) on top of planning sometimes underperforms the planning-only configuration.

Limitations

• Evaluation mostly on mathematical benchmarks with CRUX built from a 54K-problem math corpus.
• Training depends on a 72B VLM judge for multimodal coherence rewards and a 72B model for answer verification, which adds cost.
• Only crop, scale, and display tools are used; other visual operations are untested.

Experimental setup

Datasets: CRUX MINT-CoT 54K problems

Benchmarks: GeoQA MathVista-Math MMStar-Math VisualProbe V* HR-Bench MathVision We-Math MathVerse LogicVista

Metrics: accuracy average accuracy token usage latency

Baselines: GPT-4o Claude 3.5 Sonnet Qwen2.5-VL-7B-Instruct MINT-CoT-7B MM-Eureka-Qwen-7B Pixel Reasoner DeepEyes Mini-o3 R1-VL-7B Hint-GRPO

Builds on: MINT-CoT DeepEyes Mini-o3 GRPO Qwen2.5-VL

Closest papers by meaning

• MINT-CoT: Enabling Interleaved Visual Tokens in Mathematical Chain-of-Thought Reasoning (NeurIPS 2025 · similarity 0.91)
• Benchmarking Multimodal Mathematical Reasoning with Explicit Visual Dependency (NeurIPS 2026 · similarity 0.91)
• CogFlow: Bridging Perception and Reasoning through Knowledge Internalization for Visual Mathematical Problem Solving (ICLR 2026 · similarity 0.90)
• We-Math 2.0: A Versatile MathBook System for Incentivizing Visual Mathematical Reasoning (ICLR 2026 · similarity 0.90)
• Zebra-CoT: A Dataset for Interleaved Vision-Language Reasoning (ICLR 2026 · similarity 0.90)
• MV-MATH: Evaluating Multimodal Math Reasoning in Multi-Visual Contexts (CVPR 2025 · similarity 0.89)