Research map / Vision-language and multimodal / Multimodal LLMs

UniVLR: Unifying Text and Vision in Visual Latent Reasoning for Multimodal LLMs

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 UniVLR, which renders text reasoning and auxiliary images into one visual canvas and trains Qwen2.5-VL-7B to reason via a short sequence of visual latent tokens before a direct answer, with no explicit text CoT at inference.

Problem

Existing visual latent reasoning methods interleave explicit text chain-of-thought with latent visual tokens, which limits efficiency and leaves auxiliary images underused in reasoning.

Approach

Render textual reasoning traces together with auxiliary images into a single canvas, pool its vision-encoder features into K latent targets, and train the MLLM to regress these targets in a two-stage curriculum (grounding on auxiliary images, then unified canvases), so inference uses only latent visual tokens before the answer.

Key results

• UniVLR reaches 82.7 overall on V* versus 77.4 for Qwen2.5-VL-7B and 80.6 for LVR.
• UniVLR averages 68.9 across benchmarks, above LVR, Monet, and SkiLa, while using 12 latent tokens and no generated text tokens versus 190 to 270 for interleaved methods.
• Reported average accuracy gain of 5.4% over prior visual latent reasoning methods with 15.2x fewer generated reasoning tokens.

Surprising findings

• Vanilla SFT of Qwen2.5-VL-7B does not consistently improve over the base model (e.g., V* drops from 77.4 to 75.4).
• Aligning latent tokens to middle-layer (layer 14 of 28) hidden states works better than aligning to the final layer.

Limitations

• Effectiveness depends on the vision encoder's OCR and layout understanding, so weaker visual priors may benefit less.
• Latent tokens are less inspectable than natural-language rationales, reducing transparency.
• Fixed latent-token budget regardless of task difficulty.

Experimental setup

Datasets: Zebra-CoT VisCoT UniVLR-SFT-140K

Benchmarks: V* HRBench4K HRBench8K MME-RealWorld-Lite

Metrics: accuracy generated reasoning tokens

Baselines: GPT-4o Qwen2.5-VL-7B PixelReasoner DeepEyes LVR Monet SkiLa CoVT

Builds on: LVR Monet SkiLa CoVT

Closest papers by meaning

• SLVR: Structured Latent Visual Reasoning via Human-like Reasoning Flows (NeurIPS 2026 · similarity 0.93)
• Latent Visual Reasoning (ICLR 2026 · similarity 0.93)
• TVI-CoT: Text-Visual Interleaved Chain-of-Thought Reasoning for Multimodal Understanding (ICML 2026 · similarity 0.92)
• AStar: Boosting Multimodal Reasoning with Automated Structured Thinking (AAAI 2026 · similarity 0.92)
• VLM-R³: Region Recognition, Reasoning, and Refinement for Enhanced Multimodal Chain-of-Thought (NeurIPS 2025 · similarity 0.92)
• ThinkMorph: Emergent Properties in Multimodal Interleaved Chain-of-Thought Reasoning (ICLR 2026 · similarity 0.92)