AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

Mental Imagery That Matters: Curing Latent Collapse in Multimodal Reasoning

Romy Luo, Rohan Dugad, Zihui (Sherry) Xue, Alex Dimakis, Kristen Grauman

NeurIPS 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 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 latent collapse in latent multimodal reasoning, where interleaved continuous visual latent blocks lose their visual content, and proposes a fix to keep internal visual workspaces meaningful.

Problem

Latent multimodal reasoning, which interleaves text with continuous visual latent blocks to simulate mental imagery, fails in practice because the visual latents collapse and do not carry meaningful visual state.

Approach

Diagnoses the collapse of continuous visual latents in interleaved multimodal reasoning and addresses it so that latent blocks act as a real internal visual workspace without decoding to pixels. Details beyond the abstract are not available.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add AnchorSim Stage 2 (contrastive anchoring + EMA distillation) on top of the interleaved Stage 1 modelQwen-2.5-VL-7B (interleaved) · Average of 9 benchmarksaverage accuracy: 59.7 -> 62.4Improved
train AnchorSim vs plain base modelQwen2.5-VL-7B · Average of 9 benchmarksaverage accuracy: 48.9 -> 62.4Improved
train AnchorSim vs interleaved Stage 1Qwen-2.5-VL-7B (interleaved) · VSPaccuracy: 74.8 -> 78.0Improved
train AnchorSim vs interleaved Stage 1Qwen-2.5-VL-7B (interleaved) · V*accuracy: 79.1 -> 82.7Improved
add AnchorSim objectiveQwen-2.5-VL-7B (interleaved) · CV-Bench 3Daccuracy: 77.3 -> 86.4Improved
add anchor loss on top of distillation-only training (ablation)AnchorSim · V*accuracy: +4.2 pointsImproved
add L_distill loss to base model (ablation)Qwen2.5-VL-7B · BLINK, V*, CV-BenchaccuracyImproved
increase latent block size from 8 to 16 tokens (ablation)AnchorSim · BLINK and CV-BenchaccuracyHurt
increase latent block size from 8 to 16 tokens (ablation)AnchorSim · V*accuracy: small improvementImproved
add latent tokens vs no-latent Qwen base (ablation)Qwen2.5-VL-7B · BLINK, V*, CV-BenchaccuracyImproved
perturb visual latents (Drop/Cross/Shuffle) in existing latent-reasoning MLLMs (ablation)Mirage, SkiLa, CoVT, Monet · pivot study, 100 samples per modelfinal-answer log-probability: Mirage Drop slope vision 0.87 vs text 0.59No clear effect
apply AnchorSim to make latents consequential in closed-loop perturbation (ablation)Qwen-interleaved vs AnchorSim · closed-loop flip rateflip rate of final answer: 15-17% -> 29-37%Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy