Research map / Vision-language and multimodal / Multimodal LLMs

GoT-R1: Unleashing Reasoning Capability of Autoregressive Visual Generation with Reinforcement Learning

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

Applies GRPO reinforcement learning with a dual-stage, MLLM-based multi-dimensional reward to Generation Chain-of-Thought in Janus-Pro, improving compositional text-to-image generation on T2I-CompBench and GenEval.

Problem

Autoregressive visual generation models struggle with complex prompts requiring precise spatial layout and attribute binding; GoT reasoning trained by SFT on fixed templates is limited and can be unfaithful to the prompt.

Approach

Use GRPO to let an autoregressive unified MLLM (Janus-Pro) explore free-form generation chains of thought, rewarded by a product of four MLLM-based signals: prompt-image, prompt-reasoning semantic and spatial (spatial scored on rendered bounding boxes), reasoning-image IoU, plus HPS v2.1.

Key results

• GoT-R1-7B reaches 0.75 overall on GenEval, with Two Obj. rising from 0.69 to 0.94 and Attr. Binding from 0.43 to 0.68 over Janus-Pro-7B-GoT.
• GoT-R1-7B obtains 0.8139 Color, 0.7339 Texture, 0.3306 2D-Spatial and 0.3944 Complex on T2I-CompBench, the highest in five of six categories.
• On COCO 2014 val, GoT-R1-7B gets CLIP 31.83 and Aesthetic 5.41, and is preferred by humans in 77% of 300 comparisons.

Surprising findings

• Using only the prompt-image reward R_PI gives the best single-reward variant, nearly matching combined rewards on some categories, suggesting the reasoning chain is largely bypassed.
• Using only the reasoning-image reward R_RI without prompt-reasoning reward degrades performance substantially (e.g., 2D-Spatial 0.0076), indicating harmful guidance from misaligned reasoning.

Limitations

• Accuracy of bounding boxes degrades when prompts contain more than ten objects.
• Inference latency increases by about 30% due to explicit reasoning chains.
• Performance depends on the capability of the MLLM used as reward model.

Experimental setup

Datasets: LAHR-GoT JourneyDB-GoT FLUX-GoT T2I-CompBench training prompts LAION-Aesthetics

Benchmarks: T2I-CompBench GenEval COCO 2014 val

Metrics: T2I-CompBench category scores GenEval overall and sub-scores CLIP score Aesthetic score human preference GPT-4o vote Kendall's tau Spearman's rho

Baselines: Janus-Pro-1B Janus-Pro-7B Janus-Pro-GoT GoT Emu3 SD-XL DALLE-3 FLUX.1 SD3 JanusFlow T2I-R1

Builds on: Generation Chain-of-Thought GRPO Janus-Pro DeepSeek-R1 HPS v2.1

Closest papers by meaning

• T2I-R1: Reinforcing Image Generation with Collaborative Semantic-level and Token-level CoT (NeurIPS 2025 · similarity 0.92)
• Thinking-while-Generating: Interleaving Textual Reasoning throughout Visual Generation (CVPR 2026 · similarity 0.92)
• Janus-Pro-R1: Advancing Collaborative Visual Comprehension and Generation via Reinforcement Learning (NeurIPS 2025 · similarity 0.90)
• Delving into RL for Image Generation with CoT: A Study on DPO vs. GRPO (NeurIPS 2025 · similarity 0.90)
• Grounded Reinforcement Learning for Visual Reasoning (NeurIPS 2025 · similarity 0.89)
• ThinkGen: Generalized Thinking for Visual Generation (CVPR 2026 · similarity 0.89)