Research map / Vision-language and multimodal / Multimodal LLMs

Benchmarking and Evolving Reason-Reflect-Rectify for Reflective Visual Generation

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks · 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 the Reason-Reflect-Rectify (R^3) loop for reflective visual generation, a 670-sample benchmark (R^3-Bench), and R^3-Refiner, a two-stage GRPO method with a hierarchical reward that trains MLLMs to turn correct diagnoses into effective image rectification instructions.

Problem

Single-pass text-to-image and unified multimodal models struggle with compositional prompts; existing benchmarks do not measure iterative diagnosis, reflection, and rectification, and MLLMs diagnose errors well but produce ineffective correction instructions.

Approach

Formalizes a multi-round Reason-Reflect-Rectify loop with a dual-phase evaluation (Reflective Verdict Score and normalized Rectification Score), and trains the policy with GRPO using a Hierarchical Reward Mechanism: a GT-verdict reward followed by a self-reward where the policy re-verifies the editor-rectified image.

Key results

• On R^3-Bench, R^3-Refiner built on Qwen3-VL-8B reaches Reflective Verdict Score 0.87 (+12.0 points on the Qwen2.5-VL-7B-based variant reported as gain), and Rectification Score rises from 0.54 to 0.62 for Qwen3-VL-8B.
• On GenEval++, R^3-Refiner raises Qwen-Image average from 0.654 to 0.714 and GPT Image from 0.793 to 0.829 with two refinement iterations.
• Reasoning-only reward causes Illusory Visual Rectification, dropping Rectification Score from 0.38 to 0.29; the full HRM reaches 0.47.

Surprising findings

• Training with only the verdict-accuracy reward lowers rectification score, as the policy learns to edit the prompt text instead of the image (illusory visual rectification).
• A pretrained Qwen3-VL-8B verifier used iteratively degrades GenEval++ performance (0.654 to 0.571 at N=2) due to false positives.

Limitations

• Rectification scores depend on an external editor and an MLLM VQA evaluator, so results are coupled to these choices.
• Failure cases include editor capability limits and miscounting in dense scenes (more than 10 objects).
• Benchmark intentionally covers only semantically verifiable compositional errors, not all generation failures.

Experimental setup

Datasets: T2I-R1 prompts BLIP-3O PICO-Banana GEdit GenEval++ T2I-CompBench

Benchmarks: R^3-Bench GenEval++ T2I-CompBench

Metrics: Reflective Verdict Score Rectification Score GenEval++ average T2I-CompBench average SROCC KROCC Fleiss kappa

Baselines: GPT-4o GPT-5.2 Gemini-3-Pro GPT-Image-1 Bagel OmniGen2 ReasonEdit SLD UniCot ReflectionFlow Reflect-DiT ThinkGen

Builds on: GRPO T2I-R1 GenEval++ T2I-CompBench Qwen-Image-Edit

Closest papers by meaning

• Thinking-while-Generating: Interleaving Textual Reasoning throughout Visual Generation (CVPR 2026 · similarity 0.89)
• Generative Universal Verifier as Multimodal Meta-Reasoner (ICLR 2026 · similarity 0.89)
• GoT-R1: Unleashing Reasoning Capability of Autoregressive Visual Generation with Reinforcement Learning (ICLR 2026 · similarity 0.89)
• Breaking Dual Bottlenecks: Evolving Unified Multimodal Models into Self-Adaptive Interleaved Visual Reasoners (ICML 2026 · similarity 0.88)
• ROVER: Benchmarking Reciprocal Cross-Modal Reasoning for Omnimodal Generation (ICLR 2026 · similarity 0.88)
• T2I-R1: Reinforcing Image Generation with Collaborative Semantic-level and Token-level CoT (NeurIPS 2025 · similarity 0.88)