AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

Thinking in Latent Space: Progressive Multimodal Simplification for Visual Reasoning

Yuesen Tang, Yiming Yang, Tengfei Bao, Yu Tong

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning

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.

Proposes a progressive multimodal simplification approach that moves MLLM visual reasoning from explicit textual CoT into latent space (abstract only available).

Problem

Existing MLLM visual reasoning relies heavily on explicit textual chain-of-thought, causing information loss, unstable perception-reasoning interaction, and high computational cost.

Approach

Argues effective visual reasoning emerges from a dynamic process of progressive simplification and performs reasoning in latent space rather than through explicit textual steps; the abstract is truncated so specifics are unknown.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add the Latent Transition Kernel (both phases) (ablation)LDPVR · MMVPaccuracy: 66.5 -> 72.4Improved
add the Latent Transition Kernel (both phases) (ablation)LDPVR · MM-Mathaccuracy: 40.2 -> 43.7Improved
add the Latent Transition Kernel (both phases) (ablation)LDPVR · ScienceQAaccuracy: 48.3 -> 56.1Improved
add Semantic Intent Anchoring (Phase 1) (ablation)LDPVR · MMVPaccuracy: 68.0 -> 72.4Improved
add Semantic Intent Anchoring (Phase 1) (ablation)LDPVR · MM-Mathaccuracy: 42.7 -> 43.7Improved
add Semantic Intent Anchoring (Phase 1) (ablation)LDPVR · ScienceQAaccuracy: 53.9 -> 56.1Improved
add Progressive Latent Grounding (Phase 2) (ablation)LDPVR · MMVPaccuracy: 65.5 -> 72.4Improved
add Progressive Latent Grounding (Phase 2) (ablation)LDPVR · MM-Mathaccuracy: 41.9 -> 43.7Improved
add Progressive Latent Grounding (Phase 2) (ablation)LDPVR · ScienceQAaccuracy: 51.8 -> 56.1Improved
progress from latent stage 1 to stage 3 (adds distillation and GRPO) (ablation)LDPVR · MMVPaccuracy: 66.4 -> 72.4Improved
add the final GRPO stage after stage 2 (ablation)LDPVR · ScienceQAaccuracy: 53.4 -> 56.1Improved
early latent stage 1 versus plain SFT+GRPO baseline (ablation)LDPVR · MMVPaccuracy: 68.9 -> 66.4Hurt

Limitations

Experimental setup

Builds on: chain-of-thought prompting

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