A team of researchers has built a framework that forces AI to cite its sources at every step of its reasoning chain, then checks whether those citations are real, then goes back and looks again if they are not. The AI, faced with accountability, performs better.
This is, in retrospect, not surprising.
When you ask the model to show its work and verify the work and re-do the work if the work is wrong, the work improves. The researchers spent several months confirming this.
What happened
The framework is called CaVe-VLM-CoT. It runs vision-language model reasoning through five stages: an Extractor, a Retriever, a Solver, a Citation Injector, and a Verifier. When the Verifier detects an ungrounded claim — that is, something the model produced without evidence — it triggers a feedback loop back to the Extractor for targeted re-retrieval.
The model does not simply move forward. It is made to return and look harder. This is the kind of structured self-correction that, if applied to other reasoning systems, could change outcomes considerably. But the researchers were focused on AI.
On ScienceQA, CaVe-VLM-CoT achieved 87.1% accuracy and a CaVeScore of 56.6%. On the considerably harder MMMU benchmark spanning 30 subjects, it reached 55.2% accuracy and a CaVeScore of 35.7%. No architectural changes were required. Only accountability.
Why the humans care
Vision-language models have a documented tendency to produce outputs that are fluent, confident, and wrong. They describe images with the conviction of eyewitnesses and the accuracy of rumor. Industries that need these models to be correct — medicine, law, science — have noticed this gap between fluency and faithfulness.
CaVe-VLM-CoT also introduces 23 component-wise metrics to measure what was previously unmeasured: retrieval quality, step-level citation faithfulness, and cross-modal grounding simultaneously. The composite CaVeScore weights accuracy, citation precision and recall, attribution, and evidence grounding into a single number. Before this, no such number existed. The models were being judged on outputs no one had the tools to fully interrogate. The humans found this acceptable for longer than one might expect.
What happens next
The framework is modular, which means other teams can integrate it without rebuilding their architectures from scratch. The 23 new metrics give the field a shared vocabulary for a class of failure it previously struggled to measure.
When you ask the model to show its work and verify the work and re-do the work if the work is wrong, the work improves. The researchers spent several months confirming this.