A team of researchers has built an automated video-interview scoring system that shows its work instead of just spitting out a number.
E-AVI pulls timestamped evidence from three channels: what a candidate says, how they say it, and how they look and move on camera. It then applies what the researchers call dimension-conditioned evidence attention, weighing that evidence against specific interview criteria before scoring. A shared evidence pool doubles as the basis for natural-language feedback and for answering follow-up questions about how a score was reached. Tested against fine-tuned multimodal baseline models on RecruitView, a public interview dataset, and a private hospitality-industry dataset, E-AVI came out ahead on rank correlation, a measure of how closely its rankings of candidates matched real outcomes; the team also ran ablation, evidence-deletion, bootstrap, and human-audit tests to check the evidence pathway was doing real work rather than window dressing.
Automated interview screening has a trust problem: a bare score is hard for a candidate to contest and hard for a recruiter to defend. Grounding a score in inspectable evidence does not fix bias baked into the underlying model, but it does make the model's reasoning auditable, which matters as regulators start asking automated hiring tools to explain themselves.
For now this is a research paper, not a shipped hiring product, but it reads like a preview of the paperwork every HR tech vendor will eventually need to produce.