Research map / Vision-language and multimodal / Multimodal LLMs

Multi-step Visual Reasoning with Visual Tokens Scaling and Verification

NeurIPS 2025 · Vision-language and multimodal › Multimodal LLMs · approach: VLM capability analysis · 1 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 VTS-V, an inference-time visual token scaling framework where an MLLM reasoner invokes visual tools step by step and a multi-step DPO-trained verifier decides when to stop, supported by a new VTS-SFT/VTS-DPO dataset.

Problem

Most MLLMs encode an image into a fixed set of visual tokens once and reason statically, so they cannot iteratively revisit or refine visual evidence on fine-grained visual reasoning tasks.

Approach

Formulates visual reasoning as an MDP in which a tool-using reasoner proposes visual actions and a verifier, trained with multi-step DPO on reasoning trajectory preference pairs, scores each step and terminates reasoning when the reward gain falls below a threshold.

Key results

• GPT-4o + VTS-V raises BLINK 13-task average from 62.25% to 69.15%, above GPT-4o + CoT (64.96%) and GPT-4o + MMFactory (68.90%).
• Qwen2.5VL-7B fine-tuned on VTS-SFT and run with VTS-V reaches 56.65% BLINK average, up from 49.07% for the base model.
• Applying VTS-V to base Qwen2VL-7B without fine-tuning drops BLINK average from 48.01% to 38.54%.

Measured effects reported in the paper

Surprising findings

• Applying VTS-V without fine-tuning degrades open-source models (e.g., Qwen2.5VL-7B drops on V*Bench from 73.30 to 67.54), and LLaMA-3.2-11B produced no valid tool-use actions.
• On MathVista, VTS-V reduces Qwen2.5VL-7B from 22.60 to 7.80 and LLaMA-3.2-11B from 30.10 to 18.81 with the drop attributed to task bias.

Limitations

• Only ten fixed visual tools in the action space; authors suggest broader tool sets.
• Performance without fine-tuning is harmful for open-source models, and results on MathVista are mixed.
• Theoretical guarantees rely on assumptions (e.g., KL-ordering conditions) described as reasonable rather than verified.

Experimental setup

Datasets: LLaVA-OneVision single-image data BLINK V*Bench MMStar MathVista

Benchmarks: BLINK V*Bench MMStar MathVista

Metrics: accuracy

Baselines: GPT-4o + CoT GPT-4o + SoM GPT-4o + Visual Sketchpad GPT-4o + MMFactory Qwen2.5-VL-7B Qwen2-VL-7B LLaMA-3.2-11B-Vision-Instruct

Builds on: direct preference optimization chain-of-thought Visual Sketchpad V-STaR LLaVA-OneVision

Closest papers by meaning

• VLM-R³: Region Recognition, Reasoning, and Refinement for Enhanced Multimodal Chain-of-Thought (NeurIPS 2025 · similarity 0.93)
• POINTS-Long: Adaptive Dual-Mode Visual Reasoning in MLLMs (CVPR 2026 · similarity 0.93)
• VGR: Visual Grounded Reasoning (ICLR 2026 · similarity 0.92)
• VisionPulse: Dynamic Visual Sparsity for Efficient Multimodal Reasoning (ICML 2026 · similarity 0.92)
• VisuRiddles: Fine-grained Perception is a Primary Bottleneck for Multimodal Large Language Models in Abstract Visual Reasoning (ICLR 2026 · similarity 0.92)
• SLVR: Structured Latent Visual Reasoning via Human-like Reasoning Flows (NeurIPS 2026 · similarity 0.92)