Research map / Vision-language and multimodal / Multimodal LLMs

Thinking Diffusion: Penalize and Guide Visual-Grounded Reasoning in Diffusion Multimodal Language Models

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.

Training-free inference methods (Position & Step Penalty and Visual Reasoning Guidance) that delay premature answer generation and boost visual grounding in diffusion multimodal LLMs during chain-of-thought reasoning.

Problem

Diffusion multimodal LLMs (dMLLMs) with chain-of-thought generate final answers too early in the denoising process and rely weakly on visual inputs in early timesteps, degrading reasoning quality.

Approach

PSP penalizes confidence of later-position tokens during early timesteps to delay answer tokens, and VRG applies classifier-free-guidance-style amplification of visual-conditioned logits to strengthen visual grounding; both are plug-and-play at inference without training.

Key results

• On M3CoT with L/T=64/32, LaViDa with PSP and VRG reaches 48.4% vs 45.8% for Low-conf baseline.
• LaViDa at L/T=64/32 with PSP and VRG (74.9% MMBench) surpasses Low-conf at L/T=256/128 (74.3%), with about 3x faster inference.
• MMaDa at 64/32 with PSP and VRG improves MMBench accuracy by up to 7.5% over Low-conf at 256/128 per the paper's claim.

Surprising findings

• Existing VLM CoT methods DDCoT and CCoT give limited or no gains on dMLLMs, sometimes underperforming the Low-confidence baseline.

Limitations

• Excessively large visual guidance scale causes the model to ignore textual information.
• VRG roughly doubles logit computation, adding time overhead (e.g., 4.73s vs 4.01s at 64/32).

Experimental setup

Datasets: M3CoT ScienceQA MMBench V* Bench LLaVA-Bench COCO MME

Benchmarks: M3CoT MMBench SQA-IMG V* Bench MME LLaVA-Bench

Metrics: accuracy time per sample (s) PDM

Baselines: DDCoT CCoT Entropy remasking Margin remasking Low-confidence remasking L2R decoding

Builds on: classifier-free guidance LLaDA LaViDa MMaDa chain-of-thought

Closest papers by meaning

• VGR: Visual Grounded Reasoning (ICLR 2026 · similarity 0.92)
• DiffThinker: Towards Generative Multimodal Reasoning with Diffusion Models (ICML 2026 · similarity 0.92)
• AStar: Boosting Multimodal Reasoning with Automated Structured Thinking (AAAI 2026 · similarity 0.91)
• Insight-V: Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models (CVPR 2025 · similarity 0.91)
• LENS: Multi-level Evaluation of Multimodal Reasoning with Large Language Models (ICLR 2026 · similarity 0.91)
• LLaDA-V: Large Language Diffusion Models with Visual Instruction Tuning (CVPR 2026 · similarity 0.91)