PhoenixNest-Video is an evidence‑grounded multimodal agent framework for video interview assessment. The system first converts the candidate's video, audio, and transcript into a semantic video graph, which serves as a structured working memory. Under the condition of a given rubric, it performs retrieval and cross‑modal verification to ensure consistency among visual, acoustic, and textual streams, thereby extracting behavioral evidence for each assessment criterion.
The scoring component (Scorer) is trained with rubric‑based reinforcement learning. The reward function consists of two parts: alignment with the rubric’s structured requirements and differentiation across score levels. This dual‑reward design internalizes the discriminative structure of multi‑level rubrics.
On the VInterview-2025 benchmark, PhoenixNest-Video achieves 91.50% grade‑level accuracy, substantially outperforming much larger proprietary models. The compact rubric‑grounded agent aligns more closely with expert panels than direct prompting of larger models, and it exposes the evidence behind each score for human review.
Core Techniques
- Semantic video graph: maps multimodal inputs into a node‑edge structured representation.
- Rubric‑conditioned retrieval: fetches evidence guided by scoring rubrics.
- Cross‑modal verification: consistency check across visual, audio, and text streams.
- Rubrics‑RL dual‑reward Scorer: optimizes both rubric alignment and score differentiation.
Experimental Results
- Grade‑level accuracy on VInterview-2025: 91.50%
- Outperforms proprietary models with orders of magnitude more parameters
- Evidence visualization enables manual audit
Review