Research map / Vision-language and multimodal / Multimodal LLMs

UniVR: Thinking in Visual Space for Unified Visual Reasoning

NeurIPS 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual 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.

Introduces UniVR, which trains a unified autoregressive image-sequence model (Emu3.5-34B) to reason and plan directly in visual space via SFT followed by VR-GRPO, an RL method with global and step-focal rewards, and evaluates on a new VR-X benchmark.

Problem

Visual reasoning and long-horizon planning in unified models rely heavily on dense text chains and image-text pairs, and existing models struggle with physical consistency and logical coherence in multi-step visual sequences.

Approach

VR-GRPO combines a global VLM reward with a step-focal reward that uses CLIP-feature inter-trajectory variance to locate the most uncertain segment and scores it with pairwise VLM comparison, combined as R_g - λ|R_g - R_s| to discourage reasoning shortcuts, without text chains or task-specific rules.

Key results

• UniVR (34B) achieves 58.2 overall on VR-X vs 39.8 for Emu3.5 and 66.1 for Gemini-3-pro, with 59.5 long-term guidance and 68.0 robot planning scores.
• On understanding benchmarks, UniVR improves MMU from 0.292 to 0.337 and MathVista from 41.7 to 44.0 over Emu3.5.
• Ablation: adding step-focal reward to global reward improves VR-X long-term planning from 54.9 to 61.6 and reduces JEPA from 15.87 to 12.89 (table values).

Surprising findings

• Global VLM reward alone causes reward hacking, degrading long-term planning and JEPA scores despite improving short general reasoning.
• Text-only reasoning-chain training on the same data yields negligible gains on multimodal understanding, while visual-reasoning training gives larger gains.

Limitations

• Requires substantial compute for 34B models with long visual sequences.
• Reward relies on general-purpose VLMs with limited fine-grained physical knowledge.
• Judge-based evaluation (VLM score) and JEPA metric are proxies; no explicit release of benchmark stated.

Experimental setup

Datasets: EgoDex Action100M Epic-kitchen VideoCraftBench AgiBot Droid Bridge ZebraCoT ThinkMorph VisualCoT VRBench Zebra-Jigsaw

Benchmarks: VR-X MMMU MME MMBench MathVista MM-Vet WorldArena Uni-MMMU RBench

Metrics: VLM score (0-100) JEPA similarity (MMD on V-JEPA features, lower is better) accuracy

Baselines: Gemini 3 Pro + Nano Banana 2 GPT-5 + GPT-image-1.5 Qwen3-VL-235B + Qwen-Image-Edit Emu3.5 Bagel OmniGen2 Show-o2 Janus-pro

Builds on: GRPO Emu3.5 V-JEPA verl

Closest papers by meaning

• Zebra-CoT: A Dataset for Interleaved Vision-Language Reasoning (ICLR 2026 · similarity 0.89)
• Grounded Reinforcement Learning for Visual Reasoning (NeurIPS 2025 · similarity 0.89)
• Visual Planning: Let's Think Only with Images (ICLR 2026 · similarity 0.89)
• MindJourney: Test-Time Scaling with World Models for Spatial Reasoning (NeurIPS 2025 · similarity 0.89)
• CLiViS: Unleashing Cognitive Map through Linguistic-Visual Synergy for Embodied Visual Reasoning (CVPR 2026 · similarity 0.88)
• PhysVLM-AVR: Active Visual Reasoning for Multimodal Large Language Models in Physical Environments (NeurIPS 2025 · similarity 0.88)