Research map / Vision-language and multimodal / Multimodal LLMs

ThinkOmni: Lifting Textual Reasoning to Omni-modal Scenarios via Guidance Decoding

ICLR 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 ThinkOmni, a training-free decoding framework that mixes logits from a text-only off-the-shelf large reasoning model with an omni-modal LLM, adaptively weighting them per step via JS divergence, to lift reasoning into omni-modal tasks.

Problem

Omni-modal LLMs lag in complex reasoning compared with large reasoning models, but enhancing them via training requires costly data and compute; existing guidance decoding assumes matched input modalities and expert models.

Approach

LRM-as-a-Guide uses a text-only reasoning model's logits (on the textual prefix) as a positive guide and the omni-model's text-only logits as the negative reference, plus Stepwise Contrastive Scaling that sets per-step weights from Jensen-Shannon divergences between perception and reasoning distributions.</808_core_idea>

Key results

• ThinkOmni with Qwen3 guide lifts Qwen2.5-Omni-7B to 70.2% on MathVista and 75.5% on MMAU (+3.4 and +4.0 over base).
• With Qwen3 guide, Qwen2.5-Omni-7B gains +7.9 on MathVision (to 32.9%).
• Against training-free baselines on Qwen2.5-Omni-7B, Average Logits Fusion drops MathVista to 55.0 while ThinkOmni reaches 68.8.

Surprising findings

• Naive averaging of OLLM and LRM logits substantially degrades accuracy (e.g., MMAU drops by 15.8 points).
• Increasing guide size from Qwen3-8B to Qwen3-14B does not further improve MMAU and OmniBench.

Limitations

• Requires shared vocabulary between OLLM and LRM for logit fusion.
• Extra inference overhead from additional forward passes.
• Failure cases arise from conflicting input information and insufficient fine-grained perception.

Experimental setup

Datasets: MathVista MathVision MathVerse MMAU Daily-Omni OmniBench

Benchmarks: MathVista MathVision MathVerse MMAU Daily-Omni OmniBench

Metrics: accuracy generation latency

Baselines: Omni-R1 HumanOmniV2 Average Logits Fusion Caption-then-Answer VCD ProxyTuning

Builds on: Contrastive Decoding Visual Contrastive Decoding ProxyTuning DoLa ProxyThinker

Closest papers by meaning

• OmnixR: Evaluating Omni-modality Language Models on Reasoning across Modalities (ICLR 2025 · similarity 0.91)
• OmniBench: Towards The Future of Universal Omni-Language Models (NeurIPS 2025 · similarity 0.89)
• AStar: Boosting Multimodal Reasoning with Automated Structured Thinking (AAAI 2026 · similarity 0.88)
• MDK12-Bench: A Multi-Discipline Benchmark for Evaluating Reasoning in Multimodal Large Language Models (AAAI 2026 · similarity 0.88)
• Insight-V: Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models (CVPR 2025 · similarity 0.88)
• CoMT: A Novel Benchmark for Chain of Multi-modal Thought on Large Vision-Language Models (AAAI 2025 · similarity 0.87)