Paper Overview
- Field: Computer Vision (CV)
- Authors: Yuan Wang, Yongchao Du, Mengting Chen, et al.
- arXiv: 2507.17086
- ExpertVerse benchmark: A capability-centric benchmark that evaluates generative models through a knowledge-intensive lens. It stratifies reasoning-based generation across an orthogonal taxonomy of 9 cognitive capabilities and 8 expert disciplines, yielding 58 sub-disciplines.
- 1,611 expert-annotated instances covering single-image editing, multi-image composition, and text-to-image generation.
- ExpertVerse-100K: An automated pipeline producing a large-scale dataset annotated with reasoning trajectories and knowledge-grounded rationales.
- KnowThinker: A VLM reasoning engine with world knowledge, trained via RL fine-tuning to jointly generate a thinking process and refined instructions.
- Bootstrapped Pareto Policy Optimization (BPPO): A tailored RL method addressing cross-modal credit-assignment misalignment and multi-objective gradient conflicts, combining Bootstrapped Reward Refinement (BRR) and Conflict-aware Pareto Advantage Fusion (CPAF).
- arXiv page: https://arxiv.org/abs/2507.17086
Summary
Recent advances in multimodal generative models have enabled instruction-based image generation to move beyond semantic manipulation toward knowledge-driven visual reasoning. However, existing methods focus on explicit commonsense reasoning, shallow causal understanding, and direct knowledge recall, and therefore fail on knowledge-intensive generation tasks.
Key Contributions
Findings
Extensive experiments on both open-source and proprietary models expose critical reasoning deficiencies, highlighting the necessity of knowledge-intensive benchmarks for next-generation visual generation.