PhoenixNest-Video Automates Evidence-Grounded Video Interview Assessment

Traditional human-only interview assessment is becoming increasingly costly and inconsistent due to the high volume of applicants, while existing AI solutions often provide opaque scores without clear justifications. This new research introduces PhoenixNest-Video, an innovative multimodal agent framework specifically designed to automate video interview assessments with a focus on evidence-grounded judgments.The framework operates by constructing a semantic video graph, which serves as a structured working memory. It then performs rubric-conditioned information retrieval, verifying insights across visual, audio, and textual streams of the interview. This process allows it to generate per-criterion scores that are directly linked to specific behavioral evidence from the candidate's materials.A key component is the Scorer, which is trained using Rubrics-based Reinforcement Learning. This training incorporates dual rewards, optimizing for both alignment with assessment rubrics and effective differentiation between score levels. PhoenixNest-Video has demonstrated high accuracy, achieving 91.50% grade-level accuracy on the VInterview-2025 benchmark, outperforming much larger proprietary models and providing transparent, expert-aligned assessments.