Research map / Vision-language and multimodal / Multimodal LLMs

Lavida-R1: Advancing Reasoning for Unified Multimodal Diffusion Language Models

ICML 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.

Proposes LaViDa-R1, a unified SFT + multi-task RL post-training recipe for the multimodal diffusion LLM LaViDa-O, with answer-forcing, tree search, and complementary-masking likelihood estimation.

Problem

Building general-purpose reasoning for multimodal diffusion LLMs with RL is unstable (collapse with KL regularization), suffers vanishing training signal on hard prompts, and lacks reliable sequence likelihoods for dLLMs.

Approach

Unify SFT, online GRPO and best-of-N self-distillation as one weighted policy-gradient objective (SFT replaces the KL term), add answer-forcing (inpainting reasoning conditioned on the ground-truth answer) and diffusion tree search for rollouts, and estimate likelihood with complementary masking at uniform weight w(t)=1.

Key results

• On Lisa-Grounding, LaViDa-R1 reaches 66.7 P@0.5 and 60.0 mIoU versus 29.2/26.1 for LaViDa-O and 62.4/56.6 for Qwen3-VL-8B.
• On ImgEdit, LaViDa-R1 scores 3.90 overall versus 3.71 for LaViDa-O and 3.81 for its SFT variant.
• On GSM8K and MATH-500, LaViDa-R1 reports 81.5 and 38.6 versus 47.4 and 23.4 for LaViDa-O.

Surprising findings

• Answer-forcing with 100% injection collapses training (near-zero benchmark scores), while 10% injection gives the best results.
• GRPO diverges even with KL regularization beta=0.1 on high-entropy image distributions, whereas the SFT-regularized unified loss is stable.

Limitations

• Ablation tables contain TODO entries, and the Math500 result for some likelihood-estimator rows is missing.
• Image editing evaluation relies on a GPT-4-based judge (ImgEdit), and the paper does not report human evaluation.
• Some reported baselines (e.g., MMaDa RL checkpoint) are not open-sourced, limiting direct comparison.

Experimental setup

Datasets: ImgEdit Lisa-Grounding MathVista MathVerse ChartQA AI2D MMMU-Pro GSM8K MATH-500 MMMU MMBench MME

Benchmarks: MathVista MathVerse ChartQA AI2D MMMU-Pro GSM8K MATH-500 MMMU MMBench MME ImgEdit Lisa-Grounding

Metrics: accuracy P@0.5 mIoU overall score

Baselines: LaViDa-O LLaDA-8B-Instruct DiffuGRPO Dream-7B LaViDa-L Dimple MMaDa-8B UniGRPO GPT-4o Qwen2.5-VL-7B Qwen3-VL-8B LISA-7B

Builds on: LaViDa-O GRPO d1 UniGRPO BOND LaViDa

Closest papers by meaning

• d1: Scaling Reasoning in Diffusion Large Language Models via Reinforcement Learning (NeurIPS 2025 · similarity 0.91)
• MMaDA: Multimodal Large Diffusion Language Models (NeurIPS 2025 · similarity 0.91)
• Thinking Diffusion: Penalize and Guide Visual-Grounded Reasoning in Diffusion Multimodal Language Models (CVPR 2026 · similarity 0.91)
• Revisual-R1: Advancing Multimodal Reasoning From Optimized Cold Start to Staged Reinforcement Learning (ICLR 2026 · similarity 0.91)
• Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models (ICLR 2026 · similarity 0.90)
• SRPO: Enhancing Multimodal LLM Reasoning via Reflection-Aware Reinforcement Learning (NeurIPS 2025 · similarity 0.89)