Research map / Vision-language and multimodal / Multimodal LLMs

Small Drafts, Big Verdict: Information-Intensive Visual Reasoning via Speculation

ICLR 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.

Proposes Speculative Verdict, a training-free framework where several small VLM draft experts produce reasoning paths, selected by a consensus score, and a large verdict VLM synthesizes them into a final answer for information-intensive visual QA.

Problem

Large VLMs struggle on information-intensive images that densely interleave text annotations with fine-grained graphics, requiring precise localization of dispersed cues and multi-hop reasoning; existing zoom-in tool methods need costly training or rely on weak attention signals.

Approach

Adapts speculative decoding's draft-then-verify idea to multimodal reasoning at the reasoning-path level: small VLMs draft CoT reasoning paths, a consensus score (cross-peer NLL differences) selects the most agreed-upon drafts, and a single strong verdict call synthesizes them with the image, correcting minority-correct errors that majority voting discards.

Key results

• SV with GPT-4o verdict reaches 88.4 ANLS on InfographicVQA, 11.9 points above GPT-4o (76.5) and 3.6 above the best draft expert.
• SV with GPT-4o verdict reaches 64.0% on ChartQAPro vs 52.6% for GPT-4o, and 49.3% on ChartMuseum vs 42.7%.
• With Qwen2.5-VL-72B verdict, SV reaches 75.6% on HR-Bench 4K, 2.5 points above the verdict alone (73.1%).

Surprising findings

• Removing NLL normalization from consensus selection collapses selection to a near-fixed subset of drafts, yet on InfographicVQA SV without normalization scores 88.9 vs 88.4 with it.
• Small reasoning verdicts (GLM-4.1V-9B-Thinking, MiMo-VL-7B-RL) generate 60-200x more output tokens than GPT-4o yet underperform SV.

Limitations

• Verdict depends on a large proprietary or strong model, so per-sample cost is nonzero and reliance on GPT-4o limits openness.
• Gains over the best draft expert are smaller on some benchmarks (e.g., ChartMuseum +1.3%) and the verdict choice matters task by task.
• Draft pool and verdict choices were tuned on the evaluated benchmarks, and most results are on a fixed set of information-intensive VQA benchmarks.

Experimental setup

Datasets: InfographicVQA ChartMuseum ChartQAPro HR-Bench 4K TallyQA MathVista

Benchmarks: InfographicVQA ChartMuseum ChartQAPro HR-Bench 4K TallyQA-Complex MathVista

Metrics: ANLS accuracy API cost per sample output tokens

Baselines: GPT-4o GPT-4o-mini Qwen2.5-VL-72B-Instruct DeepEyes Pixel-Reasoner LLaVA-Critic-72B o1 majority voting

Builds on: speculative decoding chain-of-thought prompting universal self-consistency

Closest papers by meaning

• SLVR: Structured Latent Visual Reasoning via Human-like Reasoning Flows (NeurIPS 2026 · similarity 0.89)
• DeepScan: A Training-Free Framework for Visually Grounded Reasoning in Large Vision-Language Models (CVPR 2026 · similarity 0.89)
• Multi-step Visual Reasoning with Visual Tokens Scaling and Verification (NeurIPS 2025 · similarity 0.89)
• VGR: Visual Grounded Reasoning (ICLR 2026 · similarity 0.89)
• VLM-R³: Region Recognition, Reasoning, and Refinement for Enhanced Multimodal Chain-of-Thought (NeurIPS 2025 · similarity 0.88)
• Evidence-RL: Towards Evidence-intensive Visual Reasoning (NeurIPS 2026 · similarity 0.88)