SciGen-Verifier: A Multimodal Reasoner for Explainable Verification in Scientific Image Generation

SciGen-Verifier. A Multimodal Reasoner for Explainable Verification in Scientific Image Generation. By Jiali Chen and eight others, from the South China University of Technology and the Hong Kong Polytechnic University. Published on the twenty-seventh of September. No code repository is linked to it.

In real teaching, a solution is often a drawing as much as words: a circuit, a geometric construction, a plot of a function. A teacher has to grade the drawing as carefully as the text. Unified multimodal models can now generate scientific images, but checking whether those images are correct has become a critical bottleneck. The errors rarely sit on the surface. They come from intricate domain knowledge, from structural reasoning, and from instructions with several steps. Existing verifiers mostly target natural images and collapse their judgement into a single scalar score. Scientific coverage, and explanations that could guide a correction, are left underexplored.

The response has two parts. SciGen-Verify is a benchmark dedicated to explainable verification of scientific image generation. SciGen-Verifier is a multimodal verifier driven by reasoning, trained to do that verification. The names differ by only their last few letters, so it helps to hold on to which is which: the shorter name is the benchmark, the longer one is the model. The third contribution is the evaluation that ties them together.

The benchmark spans instruction following, multidisciplinary reasoning, and world knowledge. Its protocol has three tiers layered on the verdict: the binary judgement itself, an explanation supporting it, and an instruction for corrective editing. The verifier is first trained with cold-start supervised fine-tuning, then with a two-stage reinforcement learning pipeline organised as a curriculum. In the first stage, process rewards guided by rubrics strengthen exploration of scientific reasoning. In the second, outcome rewards align the output with the ground-truth annotation.

On SciGen-Verify, SciGen-Verifier achieves competitive performance against much larger proprietary models. It also works as a practical online critic, feeding into iterative rectification of generated images. That second use connects back to the opening concern. A verifier that explains its verdict and proposes an edit can sit inside a generation loop, with each round of feedback steering the next attempt at the image. No figures are given in the abstract, so the comparison with proprietary models rests on the word competitive.

A score alone can say a diagram is wrong. It cannot say where, or how to fix it. The three-tier protocol asks for all of that, and the verifier is trained to supply it, first by reasoning its way through the science and then by matching the annotated answers. Used as a critic in a loop, it turns grading into a route to a corrected image.