Research map / Vision-language and multimodal / Multimodal LLMs

What, Whether and How? Unveiling Process Reward Models for Thinking with Images Reasoning

AAAI 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.

Introduces ThinkWithImages-PRMBench, a 1,206-trajectory human-annotated benchmark with 7 error types for evaluating LVLMs as process reward models in thinking-with-images reasoning, finding current LVLMs fall well short of human performance.

Problem

Existing process reward model benchmarks are text-centric and do not assess PRMs' ability to detect step-level errors in interleaved text-image reasoning trajectories under the thinking-with-images paradigm.

Approach

A fine-grained error taxonomy (7 types) derived from manual analysis of thinking-with-images trajectories, plus a manually annotated step-level benchmark spanning 4 categories and 16 subcategories, used to evaluate LVLMs as PRMs, along with guided-search experiments showing PRM value.

Key results

• Ovis2-34B is the best open-source LVLM PRM with 41.74% ACC and 45.72% F1, slightly above Gemini-2.5-Flash (36.81% ACC).
• Human performance reaches 81.23% ACC, far above all evaluated LVLMs.
• Gemini-2.5-Flash reaches 78.89% accuracy on correct steps but only 38.81% on incorrect steps, indicating positive bias.

Surprising findings

• Most LVLMs score below random chance (50%) when identifying incorrect samples, showing a consistent positive-evaluation bias.
• PixelReasoner-7B shows minimal to no improvement from PRM guidance in guided search.

Limitations

• Trajectories come from only four thinking-with-images models that mainly use cropping and zooming, so error patterns may not generalize to more advanced tools.
• Benchmark covers only four source datasets (V*Bench, Blink, SAT, HRBench) with 1,206 samples, a modest size.
• Guided search experiments are preliminary and use only three surrogate PRMs.

Experimental setup

Datasets: V*Bench Blink SAT HRBench MME-RealWorld

Benchmarks: ThinkWithImages-PRMBench

Metrics: accuracy F1 per-error-type accuracy

Baselines: GPT-4o Gemini-2.5-Flash Qwen2.5-VL-72B InternVL2.5-78B Gemma3-27B Ovis2-34B LLaVA-OneVision-72B

Builds on: PRMBench VisualPRM ProcessBench

Closest papers by meaning

• From Seeing to Thinking: Decoupling Perception and Reasoning Improves Post-Training of Vision-Language Models (ICML 2026 · similarity 0.91)
• Thinking with Programming Vision: Towards a Unified View for Thinking with Images (CVPR 2026 · similarity 0.91)
• When to Think and When to Look: Uncertainty-Guided Lookback (CVPR 2026 · similarity 0.91)
• Unveiling Chain of Step Reasoning for Vision-Language Models with Fine-grained Rewards (NeurIPS 2025 · similarity 0.90)
• VISCO: Benchmarking Fine-Grained Critique and Correction Towards Self-Improvement in Visual Reasoning (CVPR 2025 · similarity 0.90)
• VisRes Bench: On Evaluating the Visual Reasoning Capabilities of VLMs (CVPR 2026 · similarity 0.90)