Research map / Vision-language and multimodal / Multimodal LLMs

Leveraging Latent Visual Reasoning in Silence

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.

Shows that latent tokens in latent visual reasoning models (Monet, LVR) can be replaced by noise or removed at inference with little loss, and proposes an attention-based latent reward for RL that improves performance even as latent generation vanishes.

Problem

Latent visual reasoning models may not actually rely on their generated continuous latent tokens at inference, and standard outcome-rewarded RL tends to suppress latent generation, so the value of latent visual reasoning is unclear.

Approach

Measure latent utilization rather than latent presence: a reward equal to the average attention from post-latent text tokens to latent tokens (first decoder layer) is added to VLPO RL, encouraging latents to influence later reasoning while letting the model fall back to pure text at inference.

Key results

• Adding the latent attention reward to VLPO improves Monet-SFT on V* overall from 80.1 to 85.3 and on MathVista from 60.5 to 67.9.
• For LVR-SFT*, attention reward raises V* overall from 82.7 (SFT) to 85.9 and MMVP from 75.7 to 77.7.
• Attention from text to latents at end of training is about 7x higher than vanilla VLPO, and latent usage ratio converges toward zero.

Surprising findings

• Replacing latent tokens with random Gaussian noise or removing them entirely causes little degradation, and sometimes improves accuracy, across spatial reasoning benchmarks.
• A latent necessity reward that directly targets latent usefulness makes the model skip latent generation even faster than vanilla RL, without improving reward.

Limitations

• Only two latent reasoning frameworks (Monet and LVR) are studied.
• Limited compute forced small batch sizes and single-run experiments without variance estimates.
• Attention mass is a proxy for latent utilization rather than a counterfactual measure of contribution.

Experimental setup

Datasets: Thyme-RL ViRL39K Visual-CoT

Benchmarks: MathVista MathVerse MMVP MMBench MME-RealWorld-Lite V* HRBench4K BLINK

Metrics: accuracy

Baselines: Monet-SFT + VLPO LVR-SFT* + VLPO Monet-SFT w/o latents + GRPO Deepeyes CoVT SkiLa GPT-4o Qwen2.5-VL-7B

Builds on: Monet LVR VLPO GRPO Thyme

Closest papers by meaning

• Latent Visual Reasoning (ICLR 2026 · similarity 0.93)
• SLVR: Structured Latent Visual Reasoning via Human-like Reasoning Flows (NeurIPS 2026 · similarity 0.93)
• Imagination Helps Visual Reasoning, But Not Yet in Latent Space (ICML 2026 · similarity 0.93)
• Monet: Reasoning in Latent Visual Space Beyond Image and Language (CVPR 2026 · similarity 0.93)
• Visual Enhanced Depth Scaling for Multimodal Latent Reasoning (NeurIPS 2026 · similarity 0.92)
• Fill the GAP: A Granular Alignment Paradigm for Visual Reasoning in Multimodal Large Language Models (NeurIPS 2026 · similarity 0.91)