Research map / Vision-language and multimodal / Multimodal LLMs

Vision-aligned Latent Reasoning for Multi-modal Large Language Model

ICML 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 VaLR, which inserts learnable latent tokens before each CoT step in an MLLM and aligns their hidden states with frozen vision-encoder features via REPA, improving long-context and perception VQA on Qwen2.5-VL-7B.

Problem

MLLMs lose visual signal during long multi-step chain-of-thought generation, so they fail to benefit from test-time scaling on tasks needing extended multi-step reasoning and precise perception.

Approach

Insert K latent tokens before each reasoning step, feed last hidden states back as input embeddings (latent mode), and supervise intermediate MLLM features with patch-wise cosine representation alignment to frozen vision encoders (DINOv3, SigLIPv2, pi^3, etc.), so visual checkpoints are re-injected at every step without using the encoder at test time.

Key results

• On VSI-Bench, VaLR-M (multi-encoder) reaches 52.9% average accuracy versus 33.0% for Qwen2.5-VL-7B and 34.0% for GPT-4o.
• VaLR-S (single DINOv3 alignment) reaches 41.5% on VSI-Bench and 63.1% on BLINK, 86.4% on V*.
• Ablation shows REPA at the middle layer (12th) outperforms front (4th) and last (27th) layers, e.g., V* 86.4 vs 83.8 and 85.3.

Surprising findings

• Latent reasoning baselines LVR, CoVT and Monet collapse on VSI-Bench (14-18.6% average), below the base model.
• Aligning with Qwen's own native vision encoder (no external encoder) still beats other baselines on VSI-Bench (39.6%).

Limitations

• Multi-view frame-to-step matching relies on GPT-4o to assign target images, adding a dependency on a proprietary model.
• Gains on perception benchmarks are reported against limited baselines and some comparisons use a single seed without variance reporting.

Experimental setup

Datasets: Zebra-CoT CogCoM ReFocus Visual-CoT OneThinker-SFT GCoT

Benchmarks: VSI-Bench BLINK MMVP MMStar V* CVBench MathVista MathVision MMHal

Metrics: accuracy (%) hallucination rate

Baselines: GPT-4o Claude-4-Sonnet R1-OneVision-7B Ocean-R1-7B LVR CoVT Monet vanilla SFT

Builds on: Chain-of-Thought latent reasoning (Coconut-style continuous thought) representation alignment (REPA) Qwen2.5-VL

Closest papers by meaning

• Latent Visual Reasoning (ICLR 2026 · similarity 0.93)
• Insight-V: Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models (CVPR 2025 · similarity 0.93)
• Dual-Latent Memory Routing for Vision-Language Reasoning (ICML 2026 · similarity 0.92)
• Fill the GAP: A Granular Alignment Paradigm for Visual Reasoning in Multimodal Large Language Models (NeurIPS 2026 · similarity 0.92)
• Reasoning-Aligned Perception Decoupling for Scalable Multi-modal Reasoning (ICLR 2026 · similarity 0.92)
• SLVR: Structured Latent Visual Reasoning via Human-like Reasoning Flows (NeurIPS 2026 · similarity 0.92)