Research map / Vision-language and multimodal / Multimodal LLMs

Visual Enhanced Depth Scaling for Multimodal Latent Reasoning

NeurIPS 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 a latent multimodal reasoning framework with a visual replay module (attention-selected visual patches re-encoded for self-distillation) and a per-layer token router that adds reasoning depth to hard tokens, trained with a CoT-to-latent curriculum.

Problem

Latent multimodal reasoning under-optimizes visual tokens relative to text/latent tokens and uses fixed depth for all tokens regardless of complexity, limiting accuracy on hard samples.

Approach

Combine attention-guided visual replay (top-K visual patches densest in a sub-grid, re-encoded and distilled into visual latents) with a learned per-layer router that reapplies transformer layers to top-alpha difficult tokens, plus a curriculum that progressively replaces explicit CoT steps with latent tokens.

Key results

• On M3CoT with Qwen2-VL-7B, the method reaches 73.0% accuracy vs 69.8% for IVT-LR, using about 7 autoregressive steps versus about 10 for IVT-LR.
• On MMVP with Qwen2.5-VL-7B backbone, reports 76.67%, the best among latent reasoning methods in the main table (Laser 72.00%).
• On MathVista the method reports 69.80%, slightly above Qwen2.5-VL-7B zero-shot (68.20%).

Surprising findings

• Smaller models (2B) prefer a much lower self-distillation weight (lambda=0.2) than 7B models (lambda=1.0), which the authors attribute to visual perception bottlenecks in smaller models.

Limitations

• Gains are more modest on text-intensive benchmarks such as SeedBench-2-Plus and smaller on GQA.
• Theoretical generalization bound relies on strong assumptions (oracle routing, learned routing approximating oracle) and the paper does not validate them empirically.

Experimental setup

Datasets: OneThinker subset (30K) GQA M3CoT ScienceQA

Benchmarks: M3CoT ScienceQA GQA MMVP SeedBench-2-Plus HallusionBench HRBench BLINK MMStar MathVista MathVision MM-Math

Metrics: accuracy autoregressive steps average generation time

Baselines: IVT-LR Multimodal CoT CCoT ICoT SCAFFOLD Chain-of-Focus Laser LVR Monet DMLR DeepEyes Vision-R1

Builds on: IVT-LR Chain-of-Thought LoRA

Closest papers by meaning

• Leveraging Latent Visual Reasoning in Silence (NeurIPS 2026 · similarity 0.92)
• Latent Implicit Visual Reasoning (CVPR 2026 · similarity 0.91)
• Fill the GAP: A Granular Alignment Paradigm for Visual Reasoning in Multimodal Large Language Models (NeurIPS 2026 · similarity 0.91)
• Monet: Reasoning in Latent Visual Space Beyond Image and Language (CVPR 2026 · similarity 0.91)
• SLVR: Structured Latent Visual Reasoning via Human-like Reasoning Flows (NeurIPS 2026 · similarity 0.91)
• Latent Visual Reasoning (ICLR 2026 · similarity 0.91)