Research map / Vision-language and multimodal / Multimodal LLMs

Machine Mental Imagery: Empower Multimodal Reasoning with Latent Visual Tokens

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.

Mirage interleaves latent visual tokens (hidden states reused as visual embeddings) with text in VLM reasoning, trained via joint latent-visual distillation, text-only latent relaxation, and GRPO, outperforming text-only and unified-model baselines on spatial reasoning benchmarks.

Problem

VLMs reason only in text, forcing visual reasoning to be verbalized; prior unified image-generating models are costly to pretrain and can degrade reasoning, so multimodal reasoning on spatial tasks remains limited.

Approach

Reuse the model's own last-layer hidden states as compact latent visual tokens interleaved in the text trajectory, supervised first by cosine alignment to average-pooled helper-image features, then relaxed to text-only loss so latent tokens become task-adaptive, followed by GRPO; no pixel-level image generation.

Key results

• On VSP spatial planning, Mirage (CoT) reaches 0.58 average accuracy vs 0.51 for CoT SFT + GRPO; with GRPO it reaches 0.60.
• On VSP spatial reasoning, Mirage + GRPO achieves 0.89 average vs 0.85 for CoT SFT + GRPO.
• On SAT synthetic average, Mirage reaches 0.98 vs 0.89 for SFT + GRPO on Qwen2.5-VL-7B; Jigsaw 0.88 vs 0.86.

Surprising findings

• Direct answer-label SFT outperforms synthesized-thought SFT on VSP spatial planning, both with and without latent tokens.
• Latent tokens trained only with text loss (no stage 1) perform only slightly better than text-only baseline, unlike findings for LLM latent reasoning such as Coconut.

Limitations

• Performance depends on quality of synthesized helper images and reasoning traces from a 32B model.
• Evaluation limited to spatial reasoning benchmarks.
• Synthetic SAT helper images from a video generator may add noise; VSP spatial planning underperforms direct SFT.

Experimental setup

Datasets: VSP BLINK Jigsaw SAT COMT

Benchmarks: VSP BLINK Jigsaw SAT COMT

Metrics: accuracy

Baselines: Zero-Shot Direct SFT CoT SFT GRPO Anole MVoT

Builds on: Coconut GRPO Visual CoT MVoT LMM-R1

Closest papers by meaning

• Latent Implicit Visual Reasoning (CVPR 2026 · similarity 0.92)
• MentisOculi: Revealing the Limits of Reasoning with Mental Imagery (ICML 2026 · similarity 0.92)
• See It, Say It, Sorted: An Iterative Training-Free Framework for Visually-Grounded Multimodal Reasoning in LVLMs (CVPR 2026 · similarity 0.92)
• ThinkMorph: Emergent Properties in Multimodal Interleaved Chain-of-Thought Reasoning (ICLR 2026 · similarity 0.92)
• Open Vision Reasoner: Transferring Linguistic Cognitive Behavior for Visual Reasoning (NeurIPS 2025 · similarity 0.91)
• Toward Cognitive Supersensing in Multimodal Large Language Models (NeurIPS 2026 · similarity 0.91)