PhoenixNest-Video: Evidence-Grounded Multimodal Agent Framework for Automated Video Interview Assessment

Motivation

• Common Application submissions rose over 37% from 2021 to 2025 to more than 7.6 million, making expert-only scoring costly and vulnerable to fatigue, anchoring, and inconsistent criterion weighting.The assessment protocol requires multiple-criterion scores justified with specific interview moments and materials.Source: §1 · p. 1

Assessment challenges

• General-purpose MLLMs can miss criterion-relevant signals across modalities and time, show systematic scoring biases, and produce judgments that lack traceable behavioral evidence.The introduction also notes that fine-grained visual distinctions can trigger hallucinations and that unretained visual facts may be inaccessible later.Source: §1 · p. 2

Framework

• PhoenixNest-Video builds a semantic video graph, retrieves and cross-verifies rubric-relevant evidence across modalities, and uses dual-reward Rubrics-based Reinforcement Learning to train criterion-level scoring.The rewards target rubric alignment and score-level differentiation; scores are anchored to the candidate’s materials.Source: §1 · p. 2

Results

• 91.50% grade-level accuracy on VInterview-2025, with the lowest total-score MAE and Wasserstein distance among all baselines; on RecruitView, best macro-averaged rank-correlation and concordance across 12 regression targets.The paper reports that these VInterview-2025 results outperform substantially larger proprietary models.Source: §1 · p. 2

Automated Video Interview Assessment

• Interview assessment progressed from handcrafted features to multimodal and multi-agent systems, but most approaches produce holistic or trait-level scores without rubric grounding or evidence traceability.PhoenixNest-Video conditions each score on rubric descriptors and anchors it to verified candidate materials.Source: §2 · p. 3

MLLMs for Video Understanding

• Video MLLMs have advanced through temporal modeling, long-form processing, and reinforcement-learning post-training, while long-horizon temporal reasoning remains a bottleneck.Related work also explores structured reasoning through rewards, visual tools, retrieval, multimodal benchmarks, and multi-agent orchestration.Source: §2 · p. 3
• Unlike general-purpose video systems, PhoenixNest-Video focuses on evidence-grounded assessment and traceable reasoning over structured interview evaluations.Source: §2 · p. 3

3 Task Formulation

• Panelists independently score candidates across predefined criteria, with the final outcome derived by aggregating their individual score vectors.In a typical protocol, candidates give a presentation before a panel, sometimes followed by questions and answers.Source: §3 · p. 3
• The task maps a candidate’s video, optional materials, criteria, and score-level rubrics to per-criterion scores, evidence references with rationales, and textual feedback.The number of criteria, score scale, and rubric content are supplied by the application setting.Source: §3 · p. 3

4 Methodology

• The methodology combines parallel multimodal preprocessing with four modular assessment components and trains the Scorer using supervised fine-tuning followed by rubrics-based reinforcement learning.The pipeline produces scores, rationales, and a verified evidence chain assembled from candidate materials.Source: §4 · pp. 3–4

4.1 Multimodal Preprocessing

• Preprocessing samples 32 frames per video, extracts presentation slides, enhances and transcribes audio, and expands rubric descriptors into behavioral indicators.Each criterion and score-level pair receives three to six indicators, which direct retrieval queries and ground training rewards.Source: §4.1 · p. 4

4.2 Pipeline Components

• The Grapher builds a reusable clip-level semantic graph, the Retriever finds rubric-relevant clips, the Verifier checks them across modalities, and the Scorer produces criterion-level assessments.The graph links clips that share prototype entities, surfacing temporally distant evidence; verified clips form a chronological, criterion-specific evidence chain with clip times, matched indicators, and modality flags.Source: §4.2 · pp. 4–5

4.3 Scorer Training

• The Scorer is fine-tuned on expert annotations and then trained with alignment and differentiation rewards using GRPO.Alignment rewards rubric-faithful rationales supported by behavioral evidence, while differentiation rewards agreement with expert total scores at institutional grading levels.Source: §4.3 · pp. 5–6

5.1 Experimental Setup

• VInterview-2025 provides 491 real admissions interviews with panel rubric labels and a 200-interview test set from a different admissions round.The setup also retrains and evaluates on RecruitView’s official user-stratified split and compares against proprietary, open-source, and video-specialized prompt-based baselines.Source: §5.1 · pp. 6–7

5.2 Main Results (RQ1)

• PhoenixNest-Video achieves the best total-score alignment and competitive per-criterion agreement on VInterview-2025 despite using an 8B backbone.Wrapping three large MLLMs in the framework improves each backbone over direct prompting on every metric; the largest absolute gain is for GLM-4.5v, whose grade-level accuracy rises by more than ten points and per-criterion MAE drops by a comparable margin.Source: §5.2 · p. 8

5.3 Ablation Studies (RQ2)

• Removing either reward or the Verifier degrades performance, while the SFT-only baseline shows the largest overall regression.Ralign supports rubric-faithful rationales, Rdiff shapes institutional-level score differentiation, and their complementary effects indicate that supervised fine-tuning alone is insufficient.Source: §5.3 · p. 8, Figure 4 · p. 8

5.4 Results on RecruitView (RQ3)

• On RecruitView (Gupta et al., 2025), PhoenixNest-Video attains the best score across all four rank-correlation and concordance metrics.The benchmark uses continuous personality and interview-performance targets rather than ordinal rubric scores; the advantage suggests the framework remains effective when the target structure changes.Source: Table 2 · p. 7, §5.4 · pp. 8–9

5.5 Deep Analysis (RQ4)

• The gender audit finds no systematic score preference, while discipline-level means show a moderate range with overlapping confidence intervals.The authors attribute residual spread to evidence density: disciplines with richer verbal content tend to surface more rubric-aligned behavioral indicators during retrieval and verification.Source: §5.5 · p. 9, Table 3 · p. 9, Figure 5 · p. 9
• Grade-level accuracy rises sharply from 10 to 16 sampled frames, plateaus through 32, and declines at 64, making 32 the adopted default.Too few frames can omit criterion-relevant evidence, whereas denser sampling can increase visual-token costs and dilute attention across redundant frames.Source: §5.5 · p. 9, Table 4 · p. 9

• The paper presents PhoenixNest-Video as a multimodal agent framework for automated video interview assessment.Source: §6 · p. 9
• PhoenixNest-Video outperforms substantially larger proprietary and open-source baselines on VInterview-2025 and RecruitView, showing rubric grounding and evidence traceability matter more than raw scale for structured assessment.Source: §6 · p. 9
• Rubrics-based Reinforcement Learning supplies dual rewards for rubric alignment and score-level differentiation, while a four-stage pipeline anchors scores to candidate materials that survive cross-modal verification.Source: §6 · p. 9

Limitations

• Fairness evidence covers gender and academic discipline in one institution’s admissions cohort, with group sizes sufficient to detect only large disparities.Other candidate attributes were not recorded, and retrieved behavioral and linguistic signals may reflect cultural norms, coaching access, or recording conditions rather than capability.Source: Limitations · pp. 9–10
• English interviews and missing human reference transcripts leave accent- and recording-related ASR errors unmeasured, while disfluent speech may depress content scores through the faculty-trained Scorer.The authors call for data with wider variation in recording conditions and speaker language backgrounds to separate language proficiency from rubric-targeted competencies.Source: Limitations · p. 10
• Chronological ordering of reasoning-chain entries is unablated and may shift scores because LLMs are sensitive to structured-input order (He et al., 2026).An alternative such as ordering by verification confidence remains untested.Source: Limitations · p. 10
• Assessment assumes the full interview is available beforehand, leaving streaming evaluation open, and VInterview-2025 is not released because its identifiable recordings cannot be anonymized.The corpus remains on institutional infrastructure, and neither recordings nor the derived benchmark are released.Source: Limitations · p. 10