SciGen-Verifier: A Multimodal Reasoner for Explainable Verification in Scientific Image Generation
The paper targets a narrow failure point: grading whether generated scientific diagrams are actually correct.
The authors introduce SciGen-Verify, a benchmark for checking scientific image generation across instruction following, reasoning, and world-knowledge tasks. Its labels go beyond yes/no, adding explanations and corrective edit instructions. They also train SciGen-Verifier with supervised fine-tuning and a two-stage reinforcement learning pipeline. On the benchmark, they say it is competitive with much larger proprietary models and can act as an online critic for iterative fixes. HF Daily Papers' note
The authors introduce SciGen-Verify, a benchmark for checking scientific image generation across instruction following, reasoning, and world-knowledge tasks. Its labels go beyond yes/no, adding explanations and corrective edit instructions. They also train SciGen-Verifier with supervised fine-tuning and a two-stage reinforcement learning pipeline. On the benchmark, they say it is competitive with much larger proprietary models and can act as an online critic for iterative fixes. HF Daily Papers' note
score 4