Research map / Vision-language and multimodal / Multimodal LLMs

Imagination Helps Visual Reasoning, But Not Yet in Latent Space

ICML 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.

Uses causal mediation analysis to show latent visual reasoning tokens in MLLMs are homogeneous, weakly affect answers, and carry little visual info, then proposes CapImagine, a text-space imagination method trained on rewritten Monet-SFT data that beats latent baselines.

Problem

Latent visual reasoning in MLLMs is popular, but it is unclear whether latent tokens actually carry visual reasoning that causally affects the answer.

Approach

Model latent reasoning as a causal chain input→latent tokens→answer and perturb each link; then replace latent imagination with explicit text captions describing auxiliary-image manipulations, trained on filtered and rewritten Monet-SFT-125K data.

Key results

• CapImagine reaches 85.9 on V* overall, versus 83.3 for Monet and 81.7 for LVR.
• CapImagine surpasses Monet by 4.0% on HR-Bench-8K and 4.9% on MME-RealWorld-Lite.
• Replacing latent tokens with a shared tensor in Monet changes V* accuracy by only about 0.5%, while intervening on CapImagine text imagination drops V* to 22.5.

Surprising findings

• Direct SFT on Monet-SFT-125K without latent-related mismatch matches the full Monet pipeline with policy optimization, questioning the role of latent tokens.
• Latent tokens of Monet, LVR and Mirage are highly similar across instances and tasks, and intervening on them barely changes answers.

Limitations

• Latent analysis covers only a few latent reasoning methods (Monet, LVR, Mirage) on small samples (100 instances for probing).
• Text-space CapImagine relies on long text sequences, with efficiency close to latent methods rather than faster.
• Filtering the Visual-CoT data relies on MLLM judgment and manual checks.

Experimental setup

Datasets: Monet-SFT-125K V* HR-Bench MME-RealWorld-Lite BLINK Hyperphantasia STARE TableVQA

Benchmarks: V* HR-Bench-4K HR-Bench-8K MME-RealWorld-Lite BLINK TableVQA Hyperphantasia STARE

Metrics: accuracy cosine similarity

Baselines: Monet LVR Mirage DeepEyes PixelReasoner InternVL3-8B Qwen2.5-VL-7B GPT-4o

Builds on: Monet Latent Visual Reasoning (LVR) Mirage Monet-SFT-125K

Closest papers by meaning

• Leveraging Latent Visual Reasoning in Silence (NeurIPS 2026 · similarity 0.93)
• Machine Mental Imagery: Empower Multimodal Reasoning with Latent Visual Tokens (CVPR 2026 · similarity 0.90)
• When Visualizing is the First Step to Reasoning: MIRA, a Benchmark for Visual Chain-of-Thought (CVPR 2026 · similarity 0.90)
• On the Faithfulness of Visual Thinking: Measurement and Enhancement (NeurIPS 2026 · similarity 0.90)
• Visual Enhanced Depth Scaling for Multimodal Latent Reasoning (NeurIPS 2026 · similarity 0.89)
• Monet: Reasoning in Latent Visual Space Beyond Image and Language (CVPR 2026 · similarity 0.89)