Research map / Vision-language and multimodal / Multimodal LLMs

Reliable Thinking with Images

ICML 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.

Identifies "Noisy Thinking" in thinking-with-images (TWI) MLLMs and proposes RTWI, a test-time scaling method that scores mining and reasoning stage reliability via token entropy, filters unreliable traces, and performs reliability-weighted voting with early stopping.

Problem

Interleaved image-text chain-of-thought in thinking-with-images MLLMs assumes faultless visual cue mining and reasoning; errors in either stage accumulate and degrade final answers.

Approach

A text-centric reliability estimate built from top-k token entropies of the textual CoT (including tool-invocation turns) serves as a proxy for visual cue reliability; traces are filtered with percentile thresholds and voted with weights based on overall reliability and a reliability leap between mining and reasoning stages, enabling early stopping online.

Key results

• On V* Bench Attr with Qwen3-VL-8B-Thinking (online), RTWI reaches 85.2% accuracy with 61.1% token saving versus 80.0% for self-consistency.
• On TreeBench reasoning with Qwen3-VL-8B-Thinking, RTWI achieves 37.1 ACC and 59.4 TSR, versus 30.9 ACC for SC and 35.6 for Deepconf.
• On MathVision with Qwen3-VL-8B-Thinking, RTWI reaches 23.2% accuracy versus 19.9% for SC, and on LogicVista 61.7% versus 46.4%.

Surprising findings

• Existing TTS baselines (ASC, ESC, CISC, Deepconf, Self-Certainty) yield only marginal gains over vanilla self-consistency on TWI models, and some (e.g., CISC on TreeBench reasoning) drop below SC.

Limitations

• Method relies on hyperparameters (filtering ratio alpha, temperature tau, top-k selection) tuned per setting; sensitivity is shown only over limited ranges.
• Evaluated only on Qwen3-VL and DeepEyes backbones with a fixed tool set (zoom-in style), so generalization to other TWI systems is untested.

Experimental setup

Datasets: V* Bench HR-Bench TreeBench MathVision LogicVista VisualProbe

Benchmarks: V* Bench HR-Bench 4K HR-Bench 8K TreeBench MathVision LogicVista VisualProbe

Metrics: accuracy token saving ratio mIoU visual cue consistency

Baselines: Self-consistency (SC) ASC ESC CISC Deepconf Self-Certainty GPT-4o Thyme DeepEyes

Builds on: self-consistency DeepConf DeepEyes

Closest papers by meaning

• When Visualizing is the First Step to Reasoning: MIRA, a Benchmark for Visual Chain-of-Thought (CVPR 2026 · similarity 0.90)
• TVI-CoT: Text-Visual Interleaved Chain-of-Thought Reasoning for Multimodal Understanding (ICML 2026 · similarity 0.90)
• When to Think and When to Look: Uncertainty-Guided Lookback (CVPR 2026 · similarity 0.90)
• LENS: Multi-level Evaluation of Multimodal Reasoning with Large Language Models (ICLR 2026 · similarity 0.89)
• Visual Thoughts: A Unified Perspective of Understanding Multimodal Chain-of-Thought (NeurIPS 2025 · similarity 0.89)
• ThinkMorph: Emergent Properties in Multimodal Interleaved Chain-of-Thought Reasoning (ICLR 2026 · similarity 0.89)