Research map / Vision-language and multimodal / Multimodal LLMs

Latent Visual Reasoning

ICLR 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.

Proposes Latent Visual Reasoning (LVR), which lets Qwen2.5-VL reason by regenerating latent hidden states that reconstruct query-relevant visual tokens, trained with SFT (MSE + NTP) and a GRPO variant for latent reasoning.

Problem

MLLM reasoning is confined to text space or relies on external visual-editing tools, leaving a gap between visual inputs and the final answer on perception-intensive VQA.

Approach

Interleave autoregressive latent steps in which the LLM's last hidden states are trained to reconstruct ground-truth ROI visual tokens (via MSE) with normal text generation, and adapt GRPO to latent reasoning by replaying recorded latent hidden states during importance-ratio computation.

Key results

• LVR (7B, 4-16 steps) reaches 71.7-72.0% on MMVP vs 66.7% for Qwen2.5-VL-7B, and 84.4% on V*_D.A. vs 81.7%.
• Ablation: LVR with MLP or GLU latent heads and Latent End Token variant underperform the standard design (e.g., LatentEnd drops MMVP to 19.0%).
• GRPO_latent on the 3B model raises MMVP from 56.0% (SFT) to 58.0% at 16 steps and V* average gains are modest.

Surprising findings

• Mode Switching Loss collapsed to zero latent steps, failing to learn a stopping condition.
• Latent End Token decoding was highly unstable, often failing to terminate and hitting max generation steps.

Limitations

• Trained only on single-image data, so underperforms on multi-image Relative Reflectance.
• Variable-length latent reasoning decoding is unstable; the fixed-step budget is used in practice.
• RL was applied only to the 3B variant due to limited compute.

Experimental setup

Datasets: Visual CoT ViRL V* MMVP BLINK

Benchmarks: V* MMVP BLINK Counting BLINK IQ-Test BLINK JigSaw BLINK Relative Reflectance BLINK Spatial Relation

Metrics: accuracy

Baselines: Qwen2.5-VL PAPO Vision-R1 PixelReasoner Argus-X3 SFT GPT-4o Gemini2.5-Pro

Builds on: Chain-of-Thought GRPO Coconut latent reasoning Qwen2.5-VL

Closest papers by meaning

• SLVR: Structured Latent Visual Reasoning via Human-like Reasoning Flows (NeurIPS 2026 · similarity 0.95)
• Semantic-Enriched Latent Visual Reasoning (ICML 2026 · similarity 0.94)
• Monet: Reasoning in Latent Visual Space Beyond Image and Language (CVPR 2026 · similarity 0.93)
• Leveraging Latent Visual Reasoning in Silence (NeurIPS 2026 · similarity 0.93)
• Vision-aligned Latent Reasoning for Multi-modal Large Language Model (ICML 2026 · similarity 0.93)
• Fill the GAP: A Granular Alignment Paradigm for Visual Reasoning in Multimodal Large Language Models (NeurIPS 2026 · similarity 0.93)