Research map / Vision-language and multimodal / Multimodal LLMs

SLVR: Structured Latent Visual Reasoning via Human-like Reasoning Flows

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

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 SLVR, a two-stage training framework that adds visual dependency induction (masked-caption MCQA with image-grounded distractors) and typed latent reasoning stages (plan, box, evidence, integration) supervised by plans, boxes, patch features, and rationales, built on Qwen2.5-VL-7B.

Problem

MLLMs answer visual reasoning questions using linguistic priors instead of visual evidence; explicit CoT adds decoding cost and latent reasoning methods lack functional structure and per-stage supervision.

Approach

Organize latent reasoning into typed latent segments (plan, grounding box, evidence patches, integration) each with its own supervision target, combined with a masked-caption contrastive MCQA stage that removes answer-revealing text to force visual dependence.

Key results

• SLVR reaches 76.1 on MMVP (+9.4 over Qwen2.5-VL-7B's 66.7) and 85.2 on V* overall (+6.7).
• On BLINK Relation SLVR scores 53.0 vs 38.8 for Qwen2.5-VL-7B (+14.2).
• On MathVista SLVR gets 73.4 (+9.7) and ChartQA 89.6 (+3.4).

Surprising findings

• The masked-text VDI stage alone lifts MMVP from 66.7 to 71.2, showing a large share of gains come from the first stage.

Limitations

• Two-stage pipeline increases data construction and training complexity.
• Only explored Qwen2.5-VL-7B; scaling and architecture transfer untested.
• Adaptive latent-state designs and alternative supervision signals not explored.

Experimental setup

Datasets: COVT Visual-CoT

Benchmarks: V* MMVP BLINK MathVista ChartQA

Metrics: accuracy

Baselines: Qwen2.5-VL-7B Monet LVR DeepEyes v2 COVT CodeVision LaViT

Builds on: Coconut Latent Visual Reasoning (LVR) chain-of-thought COVT Visual-CoT

Closest papers by meaning

• Latent Visual Reasoning (ICLR 2026 · similarity 0.95)
• Semantic-Enriched Latent Visual Reasoning (ICML 2026 · similarity 0.94)
• Leveraging Latent Visual Reasoning in Silence (NeurIPS 2026 · similarity 0.93)
• UniVLR: Unifying Text and Vision in Visual Latent Reasoning for Multimodal LLMs (NeurIPS 2026 · similarity 0.93)
• Monet: Reasoning in Latent Visual Space Beyond Image and Language (CVPR 2026 · similarity 0.93)
• VGR: Visual Grounded Reasoning (ICLR 2026 · similarity 0.92)