Research map / Vision-language and multimodal / Multimodal LLMs

VisRef: Visual Refocusing while Thinking Improves Test-Time Scaling in Multi-Modal Large Reasoning 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.

Proposes VisRef, a training-free test-time method that re-injects a DPP-selected coreset of relevant, diverse visual tokens at each reasoning step of multimodal reasoning models, with an entropy-based stopping rule.

Problem

In multimodal large reasoning models, longer textual thinking causes attention to visual tokens to fade and models to rely on textual priors, degrading vision-dependent reasoning; existing fixes need RL fine-tuning or are text-centric at test time.

Approach

Without any retraining, at each reasoning step select a fixed-size subset of visual tokens via greedy DPP-style log-determinant maximization over a kernel projected onto the current reasoning-text subspace, trading off relevance and diversity, and re-inject them; stop when answer entropy falls below a threshold.

Key results

• On MathVision with SAIL-VL2-8B, VisRef reaches 37.3% vs 29.8% standard thinking (+7.5) and 31.9% for textual self-reflection.
• With InternVL3.5-8B, VisRef improves MathVista from 68.1% to 79.3% (+11.2) over standard thinking.
• Combining Look-Back with VisRef reaches 83.1 on MathVista, 48.2 on MathVision and 66.0 on MM-Star, the best among compared configurations.

Surprising findings

• Textual self-reflection gives inconsistent gains and degrades MM-Star accuracy by 0.6% with Qwen3-VL-8B.
• Relevance-only token selection drops performance substantially relative to the full relevance-plus-diversity objective.

Limitations

• Adds inference latency from per-step DPP token selection (about 1.1 s over standard thinking on MathVista).
• Gains reported only on a limited set of 8B-scale models and benchmarks in the main text.
• Greedy approximation used for the NP-hard subset selection, with no optimality guarantee stated.

Experimental setup

Datasets: MathVista MathVision MM-Star

Benchmarks: MathVision MathVista MM-Star TallyQA RealWorldQA

Metrics: accuracy latency

Baselines: Standard Thinking Textual Self-Reflection Look-Back Parallel thinking

Builds on: Textual self-reflection (budget forcing style) Determinantal Point Processes Look-Back

Closest papers by meaning

• SPARC: Separating Perception And Reasoning Circuits for Test-time Scaling of VLMs (ICML 2026 · similarity 0.91)
• Multi-step Visual Reasoning with Visual Tokens Scaling and Verification (NeurIPS 2025 · similarity 0.90)
• More Thought, Less Accuracy? On the Dual Nature of Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.90)
• OpenVLThinker: Complex Vision-Language Reasoning via Iterative SFT-RL Cycles (NeurIPS 2025 · similarity 0.89)
• When to Think and When to Look: Uncertainty-Guided Lookback (CVPR 2026 · similarity 0.89)
• VisionPulse: Dynamic Visual Sparsity for Efficient Multimodal Reasoning (ICML 2026 · similarity 0.89)