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

# PhoenixNest-Video: Evidence-Grounded Multimodal Agent Framework for Automated Video Interview Assessment
Source: [https://arxiv.org/html/2609.02231](https://arxiv.org/html/2609.02231)
Miaojun HuangHaimei ZhangJingshen WuHao LiuAffiliation:The Hong Kong University of Science and Technology \(Guangzhou\)

###### Abstract

Interview assessment requires per\-criterion judgments grounded in behavioral evidence, yet surging applicant volumes have made human\-only evaluation costly and inconsistent, while existing AI approaches yield opaque scores without traceable rationale\. We introduce PhoenixNest\-Video, an evidence\-grounded multimodal agent framework for automated video interview assessment\. It builds a semantic video graph as structured working memory, performs rubric\-conditioned retrieval with cross\-modal verification across visual, audio, and textual streams, and produces per\-criterion scores anchored to the candidate’s materials\. A Scorer trained via Rubrics\-based Reinforcement Learning with dual rewards for rubric alignment and score\-level differentiation internalizes the discriminative structure of multi\-level rubrics\. PhoenixNest\-Video attains 91\.50% grade\-level accuracy on VInterview\-2025, outperforming substantially larger proprietary models\. A compact, rubric\-grounded agent therefore scores candidates in closer agreement with an expert panel than direct prompting of much larger models, and exposes the evidence behind each score for human review\.

## 1Introduction

Interviews have long served as a core step in selecting candidates across graduate admissions and professional hiring[Wiens \(1976\)](https://arxiv.org/html/2609.02231#bib.bib36);[Franklin \(2024\)](https://arxiv.org/html/2609.02231#bib.bib40);[Albaroudi et al\. \(2024\)](https://arxiv.org/html/2609.02231#bib.bib37)\. To reach fair and defensible decisions, institutions require evaluators to score each candidate along multiple criteria, justify every score, and cite specific moments and materials from the interview as supporting evidence[Maude and Kirby \(2022\)](https://arxiv.org/html/2609.02231#bib.bib41)\. Yet application volumes have surged sharply in recent years, with Common Application submissions rising over 37% between 2021 and 2025 to more than 7\.6 million[Song et al\. \(2026\)](https://arxiv.org/html/2609.02231#bib.bib38);[Altbach et al\. \(2019\)](https://arxiv.org/html/2609.02231#bib.bib39), placing unprecedented pressure on this evidence\-anchored protocol\. Sustaining such rigorous per\-candidate assessment with human experts alone has become increasingly impractical at this scale\. On one hand, recruiting, training, and compensating qualified interviewers across thousands of applicants imposes a substantial labor cost, and expert staffing has not kept pace with demand\. On the other hand, even when sufficient raters are available, prolonged large\-scale evaluation amplifies well\-documented human limitations such as fatigue, anchoring biases, and inconsistent criterion weighting[Conway et al\. \(1995\)](https://arxiv.org/html/2609.02231#bib.bib2);[Campion et al\. \(1997\)](https://arxiv.org/html/2609.02231#bib.bib3);[Landy and Farr \(1980\)](https://arxiv.org/html/2609.02231#bib.bib4), eroding the consistency that high\-stakes decisions depend on\. An automated assistant addresses this directly: it applies one rubric uniformly across a large candidate pool and makes the evidence behind each score inspectable\.

![Refer to caption](https://arxiv.org/html/2609.02231v1/fig1_v3.png)Figure 1:Three paradigms of interview assessment\. \(a\) Human\-based assessment is precise but unscalable\. \(b\) ML\-based assessment is scalable but opaque\. \(c\) PhoenixNest\-Video produces per\-criterion scores with verifiable evidence, achieving scalable and evidence\-traceable assessment\.As illustrated in Figure[1](https://arxiv.org/html/2609.02231#S1.F1), interview assessment has evolved through three paradigms\. Human panels remain the gold standard in accuracy but cannot scale to current applicant volumes\. ML\-based approaches[Naim et al\. \(2015\)](https://arxiv.org/html/2609.02231#bib.bib5);[Subramaniam et al\. \(2016\)](https://arxiv.org/html/2609.02231#bib.bib6);[Hemamou et al\. \(2019\)](https://arxiv.org/html/2609.02231#bib.bib7);[Agrawal et al\. \(2020\)](https://arxiv.org/html/2609.02231#bib.bib8)train supervised models on interview data and gain scalability, yet they yield opaque numeric scores without interpretable rationale, limiting their utility in high\-stakes decisions\. Multimodal Large Language Models \(MLLMs\) offer a more capable foundation by processing visual, audio, and textual streams within a unified architecture[Maaz et al\. \(2024\)](https://arxiv.org/html/2609.02231#bib.bib9);[Li et al\. \(2024b\)](https://arxiv.org/html/2609.02231#bib.bib42);[Cheng et al\. \(2024\)](https://arxiv.org/html/2609.02231#bib.bib10);[Zhang et al\. \(2025\)](https://arxiv.org/html/2609.02231#bib.bib11), yet directly applying them to interview assessment exposes structural challenges that standard training leaves unresolved\.

Applying MLLMs to criterion\-level evaluative judgment exposes three fundamental limitations\. First, rubric scoring requires decomposing parallel visual, vocal, and verbal streams and attending to different signals per criterion[Kim et al\. \(2023\)](https://arxiv.org/html/2609.02231#bib.bib46);[Arakawa and Yakura \(2022\)](https://arxiv.org/html/2609.02231#bib.bib47);[Takeuchi and Koda \(2021\)](https://arxiv.org/html/2609.02231#bib.bib48), yet MLLMs treat video as an undifferentiated stream and cannot organize criterion\-relevant evidence across modalities and time[Wingate and Bourdage \(2024\)](https://arxiv.org/html/2609.02231#bib.bib49)\. Second, as Figure[2](https://arxiv.org/html/2609.02231#S1.F2)shows, general\-purpose MLLMs exhibit systematic scoring biases, some compressing into a narrow low band and others skewing high; optimized for fluent generation and safety\-aligned neutrality, they fail to differentiate candidates along rubric\-defined dimensions\. Third, defensible assessment requires every judgment to be traceable to specific behavioral observations, yet MLLMs produce assessment text without such anchoring and lack the temporal memory to cross\-verify what a candidate said, showed, and expressed, leaving their outputs unauditable[Fabeyo \(2025\)](https://arxiv.org/html/2609.02231#bib.bib12)\. Fine\-grained visual distinctions can also trigger hallucinations[Bai et al\. \(2026\)](https://arxiv.org/html/2609.02231#bib.bib63), while visual facts that are neither retained nor verbalized may become inaccessible in later interactions[Chen et al\. \(2026\)](https://arxiv.org/html/2609.02231#bib.bib65)\.

To address these challenges, we propose PhoenixNest\-Video, an evidence\-grounded multimodal agent framework for automated video interview assessment\. PhoenixNest\-Video constructs a semantic video graph as structured working memory, performs rubric\-conditioned retrieval with cross\-modal verification to locate and validate criterion\-relevant evidence, and produces per\-criterion scores anchored to the candidate’s materials\. A Rubrics\-based Reinforcement Learning procedure with dual reward signals for rubric alignment and score\-level differentiation trains the scoring module to internalize the discriminative structure of multi\-level rubrics\. Built on a Qwen3\-VL\-8B backbone, PhoenixNest\-Video achieves 91\.50% grade\-level accuracy on VInterview\-2025 with the lowest total\-score MAE and Wasserstein distance among all baselines, outperforming substantially larger proprietary models, and after retraining on RecruitView[Gupta et al\. \(2025\)](https://arxiv.org/html/2609.02231#bib.bib20)attains the best rank\-correlation and concordance scores macro\-averaged over its 12 regression targets\.

Figure 2:Score distributions of general\-purpose MLLMs versus human experts on VInterview\-2025\. GPT\-5\.4 systematically underscores within a narrow band, while VideoLLaMA2 overscores with a bimodal pattern\. These opposing deviations from the human expert distribution expose complementary failures in calibration and discrimination that motivate rubric\-grounded reinforcement learning\.The contributions of this paper are summarized as follows:

- •We propose PhoenixNest\-Video, an evidence\-grounded multimodal agent framework that produces criterion\-level scores with verifiable video evidence, enabling transparent and auditable automated interview assessment\.
- •We introduce Rubrics\-based Reinforcement Learning with dual reward signals for rubric alignment and score\-level differentiation, enabling MLLMs to overcome scoring biases and produce rubric\-faithful interview assessments\.
- •We evaluate PhoenixNest\-Video on VInterview\-2025 and RecruitView, where it achieves 91\.50% grade\-level accuracy and produces score distributions closely aligned with human experts, demonstrating the viability of compact, evidence\-grounded multimodal agents for automated interview assessment\.

## 2Related Work

Automated Video Interview Assessment\.Automated interview assessment has progressed from handcrafted prosodic and facial features on mock interviews[Naim et al\. \(2015\)](https://arxiv.org/html/2609.02231#bib.bib5);[Chen et al\. \(2016\)](https://arxiv.org/html/2609.02231#bib.bib55);[Chen et al\. \(2017\)](https://arxiv.org/html/2609.02231#bib.bib58)through bimodal personality prediction on short clips[Subramaniam et al\. \(2016\)](https://arxiv.org/html/2609.02231#bib.bib6);[Escalante et al\. \(2018\)](https://arxiv.org/html/2609.02231#bib.bib30)to hierarchical neural models for asynchronous screening[Hemamou et al\. \(2019\)](https://arxiv.org/html/2609.02231#bib.bib7);[Agrawal et al\. \(2020\)](https://arxiv.org/html/2609.02231#bib.bib8);[Singhania et al\. \(2020\)](https://arxiv.org/html/2609.02231#bib.bib31)\. Recent work has broadened to audio\-visual personality benchmarks[Liao et al\. \(2024\)](https://arxiv.org/html/2609.02231#bib.bib32), pose\-based analysis[Tang et al\. \(2025\)](https://arxiv.org/html/2609.02231#bib.bib33), multimodal performance assessment[Li et al\. \(2025a\)](https://arxiv.org/html/2609.02231#bib.bib34);[Inam et al\. \(2026\)](https://arxiv.org/html/2609.02231#bib.bib53), naturalistic interview datasets[Gupta et al\. \(2025\)](https://arxiv.org/html/2609.02231#bib.bib20), fairness audits[Mujtaba and Mahapatra \(2025\)](https://arxiv.org/html/2609.02231#bib.bib35);[Leong et al\. \(2019\)](https://arxiv.org/html/2609.02231#bib.bib57);[Putra et al\. \(2024\)](https://arxiv.org/html/2609.02231#bib.bib51), psychometric validation[Liff et al\. \(2024\)](https://arxiv.org/html/2609.02231#bib.bib54), and multi\-agent evaluation frameworks[Sun et al\. \(2026\)](https://arxiv.org/html/2609.02231#bib.bib56)\. These efforts predominantly output holistic or trait\-level scores without rubric grounding or evidence traceability\. PhoenixNest\-Video addresses this gap by conditioning every score on rubric descriptors and anchoring it to verified candidates’ materials\.

MLLMs for Video Understanding\.Multimodal Large Language Models \(MLLMs\) typically bridge a vision encoder with an LLM via instruction tuning[Maaz et al\. \(2024\)](https://arxiv.org/html/2609.02231#bib.bib9);[Zhang et al\. \(2023\)](https://arxiv.org/html/2609.02231#bib.bib13), with subsequent work improving temporal modeling[Cheng et al\. \(2024\)](https://arxiv.org/html/2609.02231#bib.bib10);[Zhang et al\. \(2025\)](https://arxiv.org/html/2609.02231#bib.bib11);[Shao et al\. \(2025\)](https://arxiv.org/html/2609.02231#bib.bib59)and long\-form processing through time\-aware querying[Ren et al\. \(2024\)](https://arxiv.org/html/2609.02231#bib.bib14), dual\-rate visual streams[Xu et al\. \(2024\)](https://arxiv.org/html/2609.02231#bib.bib15), and chunk\-level compression[Shu et al\. \(2025\)](https://arxiv.org/html/2609.02231#bib.bib16)\. Reasoning\-oriented post\-training via reinforcement learning has further improved spatio\-temporal reasoning[Feng et al\. \(2025\)](https://arxiv.org/html/2609.02231#bib.bib17);[Li et al\. \(2025b\)](https://arxiv.org/html/2609.02231#bib.bib18);[Tao et al\. \(2025\)](https://arxiv.org/html/2609.02231#bib.bib60)\. Outside video, domain\-specific MLLMs have used multidimensional reasoning rewards and iterative visual tools to improve structured reasoning[Hao et al\. \(2026b\)](https://arxiv.org/html/2609.02231#bib.bib62);[Fan et al\. \(2026a\)](https://arxiv.org/html/2609.02231#bib.bib64), while multimodal agents combine tool use with knowledge\-grounded retrieval[Hao et al\. \(2026a\)](https://arxiv.org/html/2609.02231#bib.bib66)\. Recent benchmarks have also begun to test what MLLMs miss beyond surface recognition, including intent\-level audiovisual understanding[Fan et al\. \(2026b\)](https://arxiv.org/html/2609.02231#bib.bib61), cross\-modal ambiguity resolution[Wang et al\. \(2025b\)](https://arxiv.org/html/2609.02231#bib.bib69), and competence in humanities and social\-science domains where judgment rather than fact retrieval is at stake[Kang et al\. \(2026b\)](https://arxiv.org/html/2609.02231#bib.bib68)\. Multi\-agent orchestration provides another route to structured reasoning, both over video[Kugo et al\. \(2025\)](https://arxiv.org/html/2609.02231#bib.bib52)and in high\-stakes decision domains such as legal judgment prediction[Kang et al\. \(2026a\)](https://arxiv.org/html/2609.02231#bib.bib70), though benchmarks such as Neptune[Nagrani et al\. \(2024\)](https://arxiv.org/html/2609.02231#bib.bib19)show that long\-horizon temporal reasoning remains a bottleneck\. In contrast to these general\-purpose systems, PhoenixNest\-Video targets evidence\-grounded assessment and traceable reasoning over structured interview evaluations\.

## 3Task Formulation

Video interview assessment is widely adopted across educational, professional, and organizational contexts to evaluate candidates through structured multimodal interactions\. In a typical protocol, a candidate delivers a presentation before a panel, sometimes followed by a question\-and\-answer session\. Each panelist then independently rates the candidate across a predefined set of criteria and submits a separate score vector, and a final outcome is derived by aggregating the individual scores\.

Our objective is to develop an evaluation system that emulates this multi\-criterion assessment process\. We define the evaluation functionℱ\\mathcal\{F\}that maps a candidate’s videoVV, optional supplementary materialsMM, a set ofNNcriteriaC=\{c1,…,cN\}C=\\\{c\_\{1\},\\dots,c\_\{N\}\\\}, and their rubricsR⁡\(c,s\)R\(c,s\)providing a textual descriptor for every score levelssin an ordinal scale, to per\-criterion scores𝒮=\{s1,…,sN\}\\mathcal\{S\}=\\\{s\_\{1\},\\dots,s\_\{N\}\\\}, evidence referencesE=\{\(mi,rationalei\)\}E=\\\{\(m\_\{i\},\\text\{rationale\}\_\{i\}\)\\\}that link each judgment to specific elements of the candidate’s materials, and textual feedbackFF

𝒮,E,F←ℱ⁡\(V,M,C,R\)\.\\mathcal\{S\},E,F\\leftarrow\\mathcal\{F\}\(V,M,C,R\)\.\(1\)
This formulation is general\. The number of criteriaNN, the score scale, and the rubric contentRRare parameters supplied by the application setting\. We evaluate on two benchmarks with different configurations, as described in Section[5](https://arxiv.org/html/2609.02231#S5)\.

## 4Methodology

![Refer to caption](https://arxiv.org/html/2609.02231v1/fig3_v3.png)Figure 3:Overview of the PhoenixNest\-Video framework\. Multimodal Preprocessing converts the interview into aligned visual, audio/transcript, and slide streams, while Rubric Decomposition expands each score descriptor into behavioral indicatorsℬ⁡\(c,s\)\\mathcal\{B\}\(c,s\)\. The Grapher constructs a clip\-level semantic index; the Retriever selects criterion\-relevant candidate clips; the Verifier checks the candidates across modalities; and the trained Scorerπθ∗\\pi\_\{\\theta\}^\{\*\}produces criterion\-level scores, evidence references, and feedback\. The Scorer is optimized by supervised fine\-tuning followed by rubrics\-based reinforcement learning with alignment \(RalignR\_\{\\text\{align\}\}\) and differentiation \(RdiffR\_\{\\text\{diff\}\}\) rewards\.Figure[3](https://arxiv.org/html/2609.02231#S4.F3)shows the overall architecture\. Given a raw interview video and a rubric, PhoenixNest\-Video first runs multimodal preprocessing \(Section[4\.1](https://arxiv.org/html/2609.02231#S4.SS1)\) to obtain structured visual and audio streams and to expand rubric descriptors into fine\-grained behavioral indicators\. Four modular components \(Section[4\.2](https://arxiv.org/html/2609.02231#S4.SS2)\) then produce the assessment via the Scorer policy

\(s^,r\)∼πθ\(⋅∣V,c,R\(c,⋅\),ℬ\(c,⋅\),A\),\(\\hat\{s\},r\)\\sim\\pi\_\{\\theta\}\(\\cdot\\mid V,c,R\(c,\\cdot\),\\mathcal\{B\}\(c,\\cdot\),A\),\(2\)wheres^\\hat\{s\}is the predicted score,rrits rationale, andAAthe verified evidence chain assembled by the pipeline \(Section[4\.2](https://arxiv.org/html/2609.02231#S4.SS2)\)\. Onlyπθ∗\\pi\_\{\\theta\}^\{\*\}is trained offline by supervised fine\-tuning followed by rubrics\-based reinforcement learning \(Section[4\.3](https://arxiv.org/html/2609.02231#S4.SS3)\)\.

### 4\.1Multimodal Preprocessing

Before entering the agent framework, PhoenixNest\-Video transforms raw video interviews into structured multimodal inputs through three parallel processing streams\.

Visual Stream\.We uniformly sample 32 frames per video to capture the candidate’s expressions, body language, and overall presentation demeanor, with 32 chosen empirically as the accuracy peak \(Section[5\.5](https://arxiv.org/html/2609.02231#S5.SS5)\)\. We additionally extract slide images to capture presentation content\. This global sample is criterion\-agnostic and runs in parallel with the clip partitioning of the Grapher \(Section[4\.2](https://arxiv.org/html/2609.02231#S4.SS2)\)\. Key moments arise downstream, per criterion, as the clips that survive rubric\-conditioned retrieval and multimodal verification\.

Audio Stream\.We extract the audio track from each video, apply ZipEnhancer for noise reduction and speech clarity enhancement, and transcribe the enhanced audio using Whisper\-Large\-v3[Radford et al\. \(2023\)](https://arxiv.org/html/2609.02231#bib.bib50)\.

Rubric Decomposition\.For each criterionc∈Cc\\in Cwith rubric descriptorsR⁡\(c,s\)R\(c,s\)fors∈\{0,1,2\}s\\in\\\{0,1,2\\\}, we use an MLLM to expand the concise rubric text into fine\-grained behavioral indicators:

ℬ⁡\(c,s\)=\{b1,b2,…,bm\}←MLLM​\(R⁡\(c,s\)\),\\mathcal\{B\}\(c,s\)=\\\{b\_\{1\},b\_\{2\},\\dots,b\_\{m\}\\\}\\leftarrow\\text\{MLLM\}\(R\(c,s\)\),\(3\)wheremmis the number of behavioral indicators generated for the pair\(c,s\)\(c,s\), ranging from three to six in our implementation, and eachbib\_\{i\}describes a concrete behavioral manifestation corresponding to score levelss\. For instance, for Critical Thinking at score levels=2s=2, the decomposition produces indicators such as “analyzes the problem from multiple complementary angles” and “supports claims with specific evidence”; ats=0s=0, it produces “provides circular or unsubstantiated reasoning” and “fails to engage with the substance of the question\.” These behavioral indicators ground the reward signals during training and direct the retrieval queries during inference\.

### 4\.2Pipeline Components

To ensure that every score decision is anchored in verifiable video content, PhoenixNest\-Video passes each video through four sequential stages at inference time, transforming a long\-form video interview into a structured, auditable assessment through the Grapher, Retriever, Verifier, and Scorer\.

Grapher\.We partition the videoVVat 1\.0 FPS into clips\{V1,…,Vn\}\\\{V\_\{1\},\\dots,V\_\{n\}\\\}ofK=64K=64frames each and prompt an MLLM to extract open\-vocabulary semantic mentionsEiE\_\{i\}from each clip together with its aligned transcriptCiC\_\{i\}\. A mention is an entity \(a person, object, or presentation material, stored as a short name with a description\), an action, or a scene label marking the interview phase\. The vocabulary is open and drawn from the clip content itself\. Two mentions are compared by cosine similarity on L2\-normalized\[CLS\]embeddings fromBAAI/bge\-large\-en\-v1\.5,

sim⁡\(ta,tb\)=𝐞a⋅𝐞b‖𝐞a‖​‖𝐞b‖,\\mathrm\{sim\}\(t\_\{a\},t\_\{b\}\)=\\frac\{\\mathbf\{e\}\_\{a\}\\cdot\\mathbf\{e\}\_\{b\}\}\{\\\|\\mathbf\{e\}\_\{a\}\\\|\\,\\\|\\mathbf\{e\}\_\{b\}\\\|\},\(4\)and are merged into one prototype entity whensim\>τ\\mathrm\{sim\}\>\\tau\. The semantic graph𝒢=\(𝒱,ℰ\)\\mathcal\{G\}=\(\\mathcal\{V\},\\mathcal\{E\}\)has clip nodes𝒱=\{vi\}i=1n\\mathcal\{V\}=\\\{v\_\{i\}\\\}\_\{i=1\}^\{n\}, and an edge joins two clips whenever they share a prototype entity\. Edges therefore link temporally separated moments that carry semantically equivalent content, which is what a rubric criterion typically requires: a candidate may demonstrate critical thinking in the self\-introduction, again in the project presentation, and again under questioning\.𝒢\\mathcal\{G\}is built once and reused across all criteria\.

Retriever\.For each criterioncic\_\{i\}, the Retriever \(i\) takes the behavioral indicatorsℬ⁡\(ci,s\)\\mathcal\{B\}\(c\_\{i\},s\)from Eq\.[3](https://arxiv.org/html/2609.02231#S4.E3)and extracts semantic keywords𝒦\\mathcal\{K\}, \(ii\) prompts an MLLM to refine each indicatorbjb\_\{j\}into targeted queriesQ⁡\(bj\)Q\(b\_\{j\}\), and \(iii\) matches them against𝒢\\mathcal\{G\}under Eq\.[4](https://arxiv.org/html/2609.02231#S4.E4)\. A query keyword that matches a prototype entity aboveθ\\thetapulls in*every*clip node sharing that prototype, so the merged\-entity edges act as an inverted index that surfaces temporally distant evidence a clip\-local match would miss\. A second pass adds nodes whose own attributes exceedθ\\thetadirectly, and the union is re\-ranked by average embedding similarity, keeping the top\-NrN\_\{r\}candidate clipsℛ\\mathcal\{R\}\.ℛ\\mathcal\{R\}is topically relevant but may include false positives\.

Verifier\.For each indicatorbjb\_\{j\}and clipvi∈ℛv\_\{i\}\\in\\mathcal\{R\}, an MLLM is queried with the binary question “Does this clip demonstratebjb\_\{j\}?” across visual, audio, and textual modalities with equal weight;viv\_\{i\}is retained inℛ′\\mathcal\{R\}^\{\\prime\}if at least one modality answers positively above confidenceδ\\delta\. The retained clips form the reasoning chainAA, a criterion\-specific evidence package in which each entry records the clip identifier, its temporal range in the interview, the matched indicatorbjb\_\{j\}, and per\-modality flags marking which channels supported the match\. Clips on which no modality passesδ\\deltaare dropped, and the surviving entries are ordered chronologically, soAAreads as a time\-ordered evidence trail that a reviewer can replay against the recording\.

Scorer\.The trained Scorerπθ∗\\pi\_\{\\theta\}^\{\*\}is invoked on the reasoning chainAA, the rubricR⁡\(ci,⋅\)R\(c\_\{i\},\\cdot\), and the behavioral indicatorsℬ⁡\(ci,⋅\)\\mathcal\{B\}\(c\_\{i\},\\cdot\)for all score levels, producing the predicted scores^\\hat\{s\}, its rationalerr, textual feedbackFF, and references to the candidate’s materials drawn fromℛ′\\mathcal\{R\}^\{\\prime\}\. Becauseπθ∗\\pi\_\{\\theta\}^\{\*\}has been optimized \(Section[4\.3](https://arxiv.org/html/2609.02231#S4.SS3)\) for rubric\-faithful rationales and institutionally calibrated scores, its output is directly consumable as the final report\.

### 4\.3Scorer Training

The Scorer componentπθ∗\\pi\_\{\\theta\}^\{\*\}invoked at the assessment stage is trained offline in two steps: supervised fine\-tuning followed by rubrics\-based reinforcement learning\.

Supervised Fine\-Tuning\.We fine\-tune the base model on expert annotation pairs𝒟SFT=\{\(Vj,cj,sj∗,rj∗\)\}\\mathcal\{D\}\_\{\\text\{SFT\}\}=\\\{\(V\_\{j\},c\_\{j\},s\_\{j\}^\{\*\},r\_\{j\}^\{\*\}\)\\\}, establishing the basic evaluation format, output structure, and initial score distribution\.

Rubrics\-based Reinforcement Learning\.Recent clinical MLLM work shows that continuous rubric\-based rewards can provide informative supervision when binary feedback is sparse[Fan et al\. \(2026a\)](https://arxiv.org/html/2609.02231#bib.bib64), and structured reward shaping has been used more broadly to densify the learning signal in RL post\-training[Shi et al\. \(2026\)](https://arxiv.org/html/2609.02231#bib.bib71)\. A binary exact\-match GRPO reward is likewise too sparse here: with 18 criteria on a three\-level scale and a total spanning\[0,36\]\[0,36\], exact matches are rare in early training and most rollouts receive zero gradient\. We therefore design two complementary reward signals that decompose assessment quality along orthogonal axes\.

Alignment Reward\.The alignment rewardRalignR\_\{\\text\{align\}\}measures whether the generated rationale correctly references and applies the rubric criteria for the predicted score level\. We employ an independent LLM as a judge that receives the rubric descriptorR⁡\(c,s\)R\(c,s\), the behavioral indicatorsℬ⁡\(c,s\)\\mathcal\{B\}\(c,s\), the model’s predicted scores^\\hat\{s\}, and the generated rationalerr, and evaluates:

Ralign=LLMjudge​\(R⁡\(c,s\),ℬ⁡\(c,s\),s^,r\),R\_\{\\text\{align\}\}=\\text\{LLM\}\_\{\\text\{judge\}\}\(R\(c,s\),\\;\\mathcal\{B\}\(c,s\),\\;\\hat\{s\},\\;r\),\(5\)whereLLMjudge\\text\{LLM\}\_\{\\text\{judge\}\}assesses whether the rationale faithfully grounds its reasoning in the rubric descriptors and whether the cited behavioral evidence supports the assigned score level\.

Differentiation Reward\.The differentiation rewardRdiffR\_\{\\text\{diff\}\}measures the agreement between the model’s predicted total score and the expert\-assigned total score at the institutional grading level\. LetL⁡\(⋅\)L\(\\cdot\)denote the level mapping function that assigns a total score to one of four institutional levels:\[0,18\)\[0,18\),\[18,24\)\[18,24\),\[24,30\)\[24,30\),\[30,36\]\[30,36\]\. The reward is:

Rdiff=\{1\.0if​L​\(S^t​o​t​a​l\)=L⁡\(St​o​t​a​l∗\),0\.5if\|L⁡\(S^t​o​t​a​l\)−L⁡\(St​o​t​a​l∗\)\|=1,0\.0if\|L⁡\(S^t​o​t​a​l\)−L⁡\(St​o​t​a​l∗\)\|≥2,R\_\{\\text\{diff\}\}=\\begin\{cases\}1\.0&\\text\{if \}L\(\\hat\{S\}\_\{total\}\)=L\(S^\{\*\}\_\{total\}\),\\\\ 0\.5&\\text\{if \}\|L\(\\hat\{S\}\_\{total\}\)\-L\(S^\{\*\}\_\{total\}\)\|=1,\\\\ 0\.0&\\text\{if \}\|L\(\\hat\{S\}\_\{total\}\)\-L\(S^\{\*\}\_\{total\}\)\|\\geq 2,\\end\{cases\}\(6\)whereiiindexes the rubric criteria,S^t​o​t​a​l=∑i=1Ns^i\\hat\{S\}\_\{total\}=\\sum\_\{i=1\}^\{N\}\\hat\{s\}\_\{i\}is the model’s predicted total score summed over theNNcriterion\-level predictionss^i\\hat\{s\}\_\{i\}, andSt​o​t​a​l∗S^\{\*\}\_\{total\}is the expert total score\.

The total reward isR=λ1​Ralign\+λ2​RdiffR=\\lambda\_\{1\}R\_\{\\text\{align\}\}\+\\lambda\_\{2\}R\_\{\\text\{diff\}\}, and we optimizeπθ\\pi\_\{\\theta\}with standard GRPO[Shao et al\. \(2024\)](https://arxiv.org/html/2609.02231#bib.bib1)using group\-normalized advantages and a KL penaltyβKL\(πθ∥πref\)\\beta\\,\\mathrm\{KL\}\(\\pi\_\{\\theta\}\\\|\\pi\_\{\\text\{ref\}\}\)against the reference policy\.

Table 1:Main results on the VInterview\-2025 test set\.Boldindicates the best result per column;underlineindicates the second best\. “Proprietary” refers to closed\-source general\-purpose models accessed via commercial APIs; “Open\-Source” refers to publicly available general\-purpose open models; “Video\-Specific” refers to models specifically designed or optimized for video understanding and reasoning\. Total score metrics measure alignment at the institutional grading level; per\-criterion metrics measure agreement on individual 0–2 rubric scores\. Missing or non\-parseable model outputs are excluded from scoring, so the number of scored interviews is at most 200 and varies across models\. For the three backbone\-amplifier rows in theOursblock, the small green value in parentheses, e\.g\.\(Δ\\Delta\+0\.09\), denotes the absolute change of the wrapped configuration relative to the same backbone’s direct\-prompt counterpart in theOpen\-Sourceblock; the sign already encodes the direction of improvement with respect to the metric\.

## 5Experiments

Table 2:Macro\-averaged performance on the RecruitView dataset[Gupta et al\. \(2025\)](https://arxiv.org/html/2609.02231#bib.bib20)across all 12 targets on the 317\-sample user\-grouped test split\.Boldindicates the best result andunderlinethe second best in each column\. All baseline rows report our own evaluation of the corresponding model under the prompt\-based protocol\. Missing or non\-parseable API outputs are excluded from scoring\.We design experiments around four research questions that together evaluate whether PhoenixNest\-Video meets the practical demands of an admissions\-facing copilot:

- •RQ1How does PhoenixNest\-Video compare with proprietary, open\-source, and video\-specialized baselines on real\-world video interview assessment, and does the framework lift existing MLLM backbones beyond their direct\-prompt performance?
- •RQ2What is the individual contribution of each key component, namely the alignment rewardRalignR\_\{\\text\{align\}\}, the differentiation rewardRdiffR\_\{\\text\{diff\}\}, and the retrieval\-and\-verification pipeline formed by the Grapher, Retriever, and Verifier?
- •RQ3Does the framework remain effective when retrained under a different annotation schema, i\.e\., on a benchmark whose targets are continuous personality and performance dimensions rather than rubric\-grounded ordinal scores?
- •RQ4Does PhoenixNest\-Video exhibit any systematic bias along sensitive demographic or disciplinary axes, and how sensitive are its scores to the number of sampled frames per video?

### 5\.1Experimental Setup

Datasets\.We evaluate on two complementary benchmarks\.VInterview\-2025is a self\-collected dataset of 491 real\-world graduate admissions interviews recorded during a live admissions cycle at the participating institution\. The recordings were obtained passively from the existing admissions workflow and at no point fed back into the decision process, so the panel scores reflect a genuine high\-stakes evaluation rather than an annotation task performed for our work; this yields high\-quality rubric labels while keeping the data collection free of any decision\-altering intervention\. Each interview lasts approximately 15 minutes and is rated by a three\-member faculty panel on 18 rubric criteria \(0–2 scale\)\. The rubric is the institution’s standard admissions instrument, maintained by the admissions committee and in operational use across earlier cycles; it was not written or modified for this study\. All three panelists are faculty with admissions experience who received institutional training on the criteria and scoring standards, and each submits a separate 18\-criterion score vector\. Following the institution’s own aggregation rule, the total\-score ground truth is the arithmetic mean of the three faculty totals,SGT=13​∑r=13S\(r\)S\_\{\\mathrm\{GT\}\}=\\frac\{1\}\{3\}\\sum\_\{r=1\}^\{3\}S^\{\(r\)\}, and the per\-criterion reference is the arithmetic mean of the three faculty criterion scores\. We use 100 interviews for supervised fine\-tuning, 191 for rubrics\-based reinforcement learning, and hold out 200 for testing; the two training stages draw on non\-overlapping candidates, and the 200 test interviews come from a different admissions round than the training data, so the evaluation is cross\-round rather than a random split within one batch\. To protect candidate privacy, the corpus stays on institutional infrastructure and is not publicly released\.RecruitView[Gupta et al\. \(2025\)](https://arxiv.org/html/2609.02231#bib.bib20)is a public benchmark of 2,011 job\-interview videos with expert annotations along 12 regression targets covering overall personality, speaking skills, confidence, the Big Five personality traits, and additional interview\-performance dimensions, with the official user\-stratified split of 1,404 training, 290 validation, and 317 test clips, on which we retrain and evaluate our framework\.Baselines\.We compare against three categories of MLLMs under the same prompt\-based protocol: four proprietary[Singh et al\. \(2025\)](https://arxiv.org/html/2609.02231#bib.bib21);[Anthropic \(2025\)](https://arxiv.org/html/2609.02231#bib.bib22);[Team et al\. \(2023\)](https://arxiv.org/html/2609.02231#bib.bib24);[xAI \(2025\)](https://arxiv.org/html/2609.02231#bib.bib23), four open\-source[Bai et al\. \(2025\)](https://arxiv.org/html/2609.02231#bib.bib25);[Team et al\. \(2026\)](https://arxiv.org/html/2609.02231#bib.bib26);[Hong et al\. \(2025\)](https://arxiv.org/html/2609.02231#bib.bib27), and five video\-specialized[Feng et al\. \(2025\)](https://arxiv.org/html/2609.02231#bib.bib17);[Li et al\. \(2025b\)](https://arxiv.org/html/2609.02231#bib.bib18);[Wang et al\. \(2025a\)](https://arxiv.org/html/2609.02231#bib.bib28);[Cheng et al\. \(2024\)](https://arxiv.org/html/2609.02231#bib.bib10);[Zhang et al\. \(2025\)](https://arxiv.org/html/2609.02231#bib.bib11)\.

Implementation Details\.We use Qwen3\-VL\-8B\-Instruct as the backbone, trained for 3 epochs with a batch size of 16 and a learning rate of2×10−52\\times 10^\{\-5\}\. The LLM judge forRalignR\_\{\\text\{align\}\}is GPT\-5\.4\. Reward weights areλ1=λ2=0\.5\\lambda\_\{1\}=\\lambda\_\{2\}=0\.5withβ=0\.05\\beta=0\.05\. The retrieval\-and\-verification pipeline uses the untrained base model for graph construction and retrieval, while the final assessment is generated by the trainedπθ\\pi\_\{\\theta\}\. We use VLMEvalKit[Duan et al\. \(2024\)](https://arxiv.org/html/2609.02231#bib.bib67)to execute model inference and standardize output parsing during evaluation\. Video processing uniformly samples 32 frames per video, while the Grapher partitions the video at 1\.0 FPS intoK=64K=64\-frame clips; entity matching usesBAAI/bge\-large\-en\-v1\.5embeddings with merging thresholdτ=0\.7\\tau=0\.7and retrieval thresholdθ=0\.5\\theta=0\.5\. All experiments run on4×4\\timesA800 GPUs \(80GB\)\.

Metrics\.Following prior work[Ke et al\. \(2024\)](https://arxiv.org/html/2609.02231#bib.bib43);[Gu et al\. \(2024\)](https://arxiv.org/html/2609.02231#bib.bib44);[Li et al\. \(2024a\)](https://arxiv.org/html/2609.02231#bib.bib45);[Gupta et al\. \(2025\)](https://arxiv.org/html/2609.02231#bib.bib20), we report grade\-level accuracy \(ACC, computed over the institution’s four\-tier grading scheme, which aggregates raw rubric scores into the admissions\-relevant grades faculty reviewers act on; a prediction counts as correct when its tier is within one tier of the expert tier\), mean absolute error \(MAE\), Wasserstein distance \(W Dist\) at the total\-score level, and Quadratic Weighted Kappa \(QWK\) at the per\-criterion level for VInterview\-2025 test set\. For RecruitView[Gupta et al\. \(2025\)](https://arxiv.org/html/2609.02231#bib.bib20), we report Spearmanρ\\rho, Kendallτ\\tau\-b, C\-index, and Pearsonrr, macro\-averaged over all 12 targets following the evaluation protocol of the original paper\.

### 5\.2Main Results \(RQ1\)

Comparison with baselines\.Table[1](https://arxiv.org/html/2609.02231#S4.T1)reports results on the VInterview\-2025 test set\. PhoenixNest\-Video achieves the best total\-score alignment and remains competitive on per\-criterion agreement, despite using only an 8B backbone\. Among the baselines, the two strongest proprietary models, Gemini\-3\.1\-pro\-preview and Grok\-4\.1, lead every other baseline on grade\-level accuracy, while GPT\-5\.4 and Claude\-opus\-4\-6 fall below the open\-source and video\-specialized groups; video\-specialized models tend to be stronger on fine\-grained per\-criterion MAE\. The fact that our 8B configuration matches or surpasses substantially larger proprietary and open\-source models indicates that rubric\-grounded training and evidence\-grounded retrieval are more effective than raw scale for structured assessment\.

Framework as a backbone amplifier\.The remainingOursrows wrap three large MLLMs \(Qwen3\.5\-397B\-A17B, Kimi\-K2\.5, GLM\-4\.5v\) inside the PhoenixNest\-Video framework\. Each configuration consistently improves over the same backbone’s direct\-prompt counterpart on every metric, as shown by the green deltas next to each value\. The largest absolute gain appears for GLM\-4\.5v, whose grade\-level accuracy increases by more than ten points and whose per\-criterion MAE drops by a comparable margin; Qwen3\.5\-397B\-A17B and Kimi\-K2\.5 show similar trends\. This confirms that the improvements stem from the framework itself rather than from a particular choice of backbone, supporting the claim that PhoenixNest\-Video acts as a backbone\-agnostic capability amplifier for existing MLLMs\.

### 5\.3Ablation Studies \(RQ2\)

Figure 4:Ablation of PhoenixNest\-Video on VInterview\-2025\. Removing any single component or skipping reinforcement learning consistently degrades grade\-level accuracy and inflates MAE, indicating that the rubrics\-based rewards and the retrieval\-and\-verification pipeline are complementary\.Figure[4](https://arxiv.org/html/2609.02231#S5.F4)isolates the contribution of each component by removing one at a time from the full framework\. Removing either reward signal causes a substantial drop in accuracy and a sharp rise in MAE, with the differentiation rewardRdiffR\_\{\\text\{diff\}\}being slightly more impactful than the alignment rewardRalignR\_\{\\text\{align\}\}\. This pattern indicates that the two rewards are complementary along orthogonal axes:RalignR\_\{\\text\{align\}\}enforces rubric\-faithful rationales whileRdiffR\_\{\\text\{diff\}\}shapes the harder institutional\-level score differentiation\. Removing the Verifier, so that every retrieved clip is passed to the Scorer without cross\-modal checking, produces a comparable degradation, confirming that verification contributes on top of the reward shaping rather than being subsumed by it\. Finally, the SFT\-only baseline yields the largest overall regression, validating that supervised fine\-tuning alone is insufficient and that reinforcement training with structured rewards is necessary to internalize the rubric\.

### 5\.4Results on RecruitView \(RQ3\)

To verify that the framework is not tied to the self\-collected rubric schema, we retrain and evaluate PhoenixNest\-Video on the RecruitView dataset[Gupta et al\. \(2025\)](https://arxiv.org/html/2609.02231#bib.bib20)following its native train/test protocol\. RecruitView differs from VInterview\-2025 in two fundamental ways, its targets are continuous regression scores rather than 0–2 ordinal rubric scores, and they cover the Big Five personality traits, an overall\-personality index, and six interview\-performance dimensions rather than presentation\- and Q&A\-oriented criteria\. We compare against the proprietary and open\-source models of Table[1](https://arxiv.org/html/2609.02231#S4.T1), plus Doubao\-Seed\-1\.8[Guo et al\. \(2025\)](https://arxiv.org/html/2609.02231#bib.bib29), under the same prompt\-based protocol\. As Table[2](https://arxiv.org/html/2609.02231#S5.T2)shows, PhoenixNest\-Video attains the best score across all four rank\-correlation and concordance metrics, while every baseline remains at moderate correlation levels, underscoring the intrinsic difficulty of video interview assessment even for strong proprietary models\. The consistent advantage indicates that the rubric\-grounded training procedure and the retrieval\-and\-verification pipeline are not coupled to the specific labeling convention on which they were originally developed, and that they remain effective when the target structure changes\.

### 5\.5Deep Analysis \(RQ4\)

Beyond aggregate accuracy, we further analyze two properties of PhoenixNest\-Video on VInterview\-2025\. First, we audit predictions for systematic score gaps along two sensitive axes, candidate gender and academic discipline\. Second, we examine the sensitivity of the framework to the number of sampled frames per video, which is the most consequential preprocessing hyperparameter in our pipeline\.

Table 3:Mean predicted total score \(max 36\) by candidate gender on VInterview\-2025\. The aggregate gap is small relative to the score range and within\-group variance\.Gender Bias\.Table[3](https://arxiv.org/html/2609.02231#S5.T3)reports the mean predicted total score on VInterview\-2025 by candidate gender\. The gap is small relative to the score range and the within\-group spread, and the scores show no systematic preference for either gender\. The Limitations section defines the scope of this audit\.

![Refer to caption](https://arxiv.org/html/2609.02231v1/subject_bias.png)Figure 5:Mean predicted total score \(max 36\) by candidate discipline on VInterview\-2025, where each candidate is assigned directly to their reported academic discipline\. Error bars denote 95% bootstrap confidence intervals\.Subject Bias\.Figure[5](https://arxiv.org/html/2609.02231#S5.F5)groups candidates by their academic discipline\. The per\-discipline means span a moderate range, with engineering\-heavy disciplines slightly below the cross\-discipline mean and verbal\- or scenario\-heavy disciplines slightly above\. We attribute the residual spread to differences in evidence density: applicants in disciplines with richer verbal content tend to surface more rubric\-aligned behavioral indicators during retrieval and verification\. The overlap of per\-discipline confidence intervals indicates that PhoenixNest\-Video spreads its scores across fields rather than favoring a few\.

Table 4:Effect of the number of uniformly sampled frames per video on grade\-level accuracy on VInterview\-2025\. Accuracy peaks at 32 frames and then declines, indicating that denser sampling is not always beneficial\.Effect of Frame Sampling Density\.Table[4](https://arxiv.org/html/2609.02231#S5.T4)reports grade\-level accuracy as we vary the number of uniformly sampled frames per video from 10 to 64\. Accuracy rises sharply from 10 to 16 frames, plateaus through 32, and degrades at 64\. Too few frames omit criterion\-relevant evidence, whereas too many inflate the visual token budget and dilute attention across redundant near\-duplicate frames\. We therefore adopt 32 frames as the default throughout this paper\.

## 6Conclusion

We presented PhoenixNest\-Video, a multimodal agent framework for automated video interview assessment\. Rubrics\-based Reinforcement Learning supplies dual rewards for rubric alignment and score\-level differentiation, and a four\-stage agent pipeline anchors each score to the elements of the candidate’s materials that survive cross\-modal verification\. On VInterview\-2025 and RecruitView, PhoenixNest\-Video outperforms substantially larger proprietary and open\-source baselines, showing that rubric grounding and evidence traceability matter more than raw scale for structured assessment\.

## Limitations

Four conditions define where these results apply\.

Scope of the fairness audit\.The audit covers candidate gender and academic discipline in a single admissions cohort at one institution\. Within these groups and this sample, predicted scores show no large aggregate difference, and the per\-group sizes make only large disparities detectable\. The admissions workflow the data comes from records no other candidate attributes, so ethnicity, nationality, accent, disability, socioeconomic background, and first\-language background fall outside the audit; collecting them would require consent and data\-protection provisions beyond the present ethical\-review approval\. The behavioral and linguistic signals the framework retrieves may also track culturally specific communication norms, or unequal access to interview coaching and recording conditions, rather than candidate capability\.

Language proficiency and transcript quality\.The interviews are conducted in English and many candidates are non\-native speakers, which constrains the system in two ways\. First, ASR accuracy degrades under strong accents, background noise, and low\-quality recordings, and these errors propagate into the transcript stream the Grapher and the Retriever consume\. Measuring this effect requires human reference transcripts, which VInterview\-2025 does not include\. Second, disfluent or error\-prone speech can depress criteria that target content rather than language, so a candidate with limited fluency may lose points for a reason the rubric does not intend to measure\. The Scorer is trained on faculty ratings and reproduces this confound\. Separating language proficiency from the competencies the rubric targets calls for interview data with wider variation in recording conditions and speaker language background\.

Fixed design choices and deployment scope\.Two components are set once and not ablated\. The entries of the reasoning chainAAreach the Scorer in chronological order, and LLMs are sensitive to the ordering of structured inputs[He et al\. \(2026\)](https://arxiv.org/html/2609.02231#bib.bib72), so an alternative ordering such as by verification confidence may shift the resulting scores\. The pipeline also assumes that the full interview is available before assessment begins; extending it to streaming video, so that intermediate scores and evidence anchors appear while the interview unfolds, remains open\. VInterview\-2025 itself stays closed\. The recordings are graduate admissions interviews with identifiable candidates, and no anonymization we can apply to video and voice removes that identifiability, so we keep the corpus on institutional infrastructure and release neither the recordings nor the derived benchmark\.

## References

- Agrawalet al\.\(2020\)A\. Agrawal, R\. A\. George, S\. S\. Ravi, S\. K\. S, and A\. KumarLeveraging multimodal behavioral analytics for automated job interview performance assessment and feedback\.ArXivabs/2006\.07909\.External Links:[Link](https://api.semanticscholar.org/CorpusID:219686775)Cited by:[§1](https://arxiv.org/html/2609.02231#S1.p2.1),[§2](https://arxiv.org/html/2609.02231#S2.p1.1)\.
- Albaroudiet al\.\(2024\)E\. Albaroudi, T\. Mansouri, and A\. AlameerA comprehensive review of ai techniques for addressing algorithmic bias in job hiring\.Ai5\(1\),pp\. 383–404\.Cited by:[§1](https://arxiv.org/html/2609.02231#S1.p1.1)\.
- Altbachet al\.\(2019\)P\. G\. Altbach, L\. Reisberg, and L\. E\. RumbleyTrends in global higher education: tracking an academic revolution\.Vol\.22,Brill\.Cited by:[§1](https://arxiv.org/html/2609.02231#S1.p1.1)\.
- Anthropic \(2025\)AnthropicIntroducing claude sonnet 4\.5\.https://www\.anthropic\.com/news/claude\-sonnet\-4\-5\.Cited by:[Table 1](https://arxiv.org/html/2609.02231#S4.T1.2.1.4.1),[§5\.1](https://arxiv.org/html/2609.02231#S5.SS1.p1.1),[Table 2](https://arxiv.org/html/2609.02231#S5.T2.2.1.1.1.3.1)\.
- Arakawa and Yakura \(2022\)R\. Arakawa and H\. YakuraAI for human assessment: what do professional assessors need?\.Extended Abstracts of the 2023 CHI Conference on Human Factors in Computing Systems\.External Links:[Link](https://api.semanticscholar.org/CorpusID:248239708)Cited by:[§1](https://arxiv.org/html/2609.02231#S1.p3.1)\.
- Baiet al\.\(2025\)S\. Bai, Y\. Cai, R\. Chen, K\. Chen, X\. Chen, Z\. Cheng, L\. Deng, W\. Ding, C\. Gao, C\. Ge,et al\.Qwen3\-vl technical report\.arXiv preprint arXiv:2511\.21631\.Cited by:[Table 1](https://arxiv.org/html/2609.02231#S4.T1.2.1.7.2),[Table 1](https://arxiv.org/html/2609.02231#S4.T1.2.1.8.1),[§5\.1](https://arxiv.org/html/2609.02231#S5.SS1.p1.1),[Table 2](https://arxiv.org/html/2609.02231#S5.T2.2.1.1.1.7.1)\.
- Baiet al\.\(2026\)T\. Bai, Y\. Fan, Q\. Jiantao, F\. Sun, J\. Song, J\. Han, Z\. Liu, C\. He, W\. Zhang, and B\. YuanHallucination at a glance: controlled visual edits and fine\-grained multimodal learning\.Advances in Neural Information Processing Systems38,pp\. 135360–135393\.Cited by:[§1](https://arxiv.org/html/2609.02231#S1.p3.1)\.
- Campionet al\.\(1997\)M\. A\. Campion, D\. K\. Palmer, and J\. E\. CampionA review of structure in the selection interview\.Personnel psychology50\(3\),pp\. 655–702\.Cited by:[§1](https://arxiv.org/html/2609.02231#S1.p1.1)\.
- Chenet al\.\(2026\)H\. Chen, K\. Chen, Y\. Fan, B\. Wang, Y\. Gao, Y\. Chu, and X\. HuSeen, said, or forgotten? a causal audit of visual kv memory across dialog turns\.arXiv preprint arXiv:2607\.25467\.Cited by:[§1](https://arxiv.org/html/2609.02231#S1.p3.1)\.
- Chenet al\.\(2016\)L\. Chen, G\. Feng, M\. P\. Martin\-Raugh, C\. W\. Leong, C\. Kitchen, S\. Yoon, B\. Lehman, H\. Kell, and C\. M\. LeeAutomatic scoring of monologue video interviews using multimodal cues\.\.InINTERSPEECH,pp\. 32–36\.Cited by:[§2](https://arxiv.org/html/2609.02231#S2.p1.1)\.
- Chenet al\.\(2017\)L\. Chen, R\. Zhao, C\. W\. Leong, B\. Lehman, G\. Feng, and M\. E\. HoqueAutomated video interview judgment on a large\-sized corpus collected online\.In2017 Seventh International Conference on Affective Computing and Intelligent Interaction \(ACII\),pp\. 504–509\.Cited by:[§2](https://arxiv.org/html/2609.02231#S2.p1.1)\.
- Chenget al\.\(2024\)Z\. Cheng, S\. Leng, H\. Zhang, Y\. Xin, X\. Li, G\. Chen, Y\. Zhu, W\. Zhang, Z\. Luo, D\. Zhao,et al\.Videollama 2: advancing spatial\-temporal modeling and audio understanding in video\-llms\.arXiv preprint arXiv:2406\.07476\.Cited by:[§1](https://arxiv.org/html/2609.02231#S1.p2.1),[§2](https://arxiv.org/html/2609.02231#S2.p2.1),[Table 1](https://arxiv.org/html/2609.02231#S4.T1.2.1.14.1),[§5\.1](https://arxiv.org/html/2609.02231#S5.SS1.p1.1)\.
- Conwayet al\.\(1995\)J\. M\. Conway, R\. A\. Jako, and D\. F\. GoodmanA meta\-analysis of interrater and internal consistency reliability of selection interviews\.\.Journal of applied psychology80\(5\),pp\. 565\.Cited by:[§1](https://arxiv.org/html/2609.02231#S1.p1.1)\.
- Duanet al\.\(2024\)H\. Duan, J\. Yang, Y\. Qiao, X\. Fang, L\. Chen, Y\. Liu, X\. Dong, Y\. Zang, P\. Zhang, J\. Wang,et al\.VLMEvalKit: an open\-source toolkit for evaluating large multi\-modality models\.InProceedings of the 32nd ACM International Conference on Multimedia,pp\. 11198–11201\.Cited by:[§5\.1](https://arxiv.org/html/2609.02231#S5.SS1.p2.1)\.
- Escalanteet al\.\(2018\)H\. Escalante, H\. Kaya, A\. Salah, S\. Escalera, Y\. Gucluturk, U\. Guclu, X\. Baró, I\. Guyon, J\. Junior, M\. Madadi,et al\.Explaining first impressions: modeling, recognizing, and explaining apparent personality from videos\. arxiv preprint arxiv: 180200745\.Cited by:[§2](https://arxiv.org/html/2609.02231#S2.p1.1)\.
- Fabeyo \(2025\)S\. FabeyoExplainable ai in employment decision\-making: a systematic review of transparency methods in hiring algorithms\.\.Issues in Information Systems26\(3\)\.Cited by:[§1](https://arxiv.org/html/2609.02231#S1.p3.1)\.
- Fanet al\.\(2026a\)Y\. Fan, J\. Hao, H\. Chen, J\. Bao, Y\. Shao, Y\. Liang, K\. F\. Hung, and H\. TangOralGPT\-plus: learning to use visual tools via reinforcement learning for panoramic x\-ray analysis\.arXiv preprint arXiv:2603\.06366\.Cited by:[§2](https://arxiv.org/html/2609.02231#S2.p2.1),[§4\.3](https://arxiv.org/html/2609.02231#S4.SS3.p3.1)\.
- Fanet al\.\(2026b\)Y\. Fan, G\. Seo, J\. Hao, J\. Cho, M\. Bansal, and J\. YoonMuseBench: benchmarking intent\-level audiovisual arts understanding in mllms\.arXiv preprint arXiv:2606\.30026\.Cited by:[§2](https://arxiv.org/html/2609.02231#S2.p2.1)\.
- Fenget al\.\(2025\)K\. Feng, K\. Gong, B\. Li, Z\. Guo, Y\. Wang, T\. Peng, J\. Wu, X\. Zhang, B\. Wang, and X\. YueVideo\-r1: reinforcing video reasoning in mllms\.arXiv preprint arXiv:2503\.21776\.Cited by:[§2](https://arxiv.org/html/2609.02231#S2.p2.1),[Table 1](https://arxiv.org/html/2609.02231#S4.T1.2.1.11.2),[§5\.1](https://arxiv.org/html/2609.02231#S5.SS1.p1.1)\.
- Franklin \(2024\)T\. FranklinWork in progress: development of a taxonomy of undergraduate engineering admissions practices and protocols\.In2024 ASEE Annual Conference & Exposition,Cited by:[§1](https://arxiv.org/html/2609.02231#S1.p1.1)\.
- Guet al\.\(2024\)J\. Gu, X\. Jiang, Z\. Shi, H\. Tan, X\. Zhai, C\. Xu, W\. Li, Y\. Shen, S\. Ma, H\. Liu,et al\.A survey on llm\-as\-a\-judge\.arXiv preprint arXiv:2411\.15594\.Cited by:[§5\.1](https://arxiv.org/html/2609.02231#S5.SS1.p3.1)\.
- Guoet al\.\(2025\)D\. Guo, F\. Wu, F\. Zhu, F\. Leng, G\. Shi, H\. Chen, H\. Fan, J\. Wang, J\. Jiang, J\. Wang,et al\.Seed1\. 5\-vl technical report\.arXiv preprint arXiv:2505\.07062\.Cited by:[§5\.4](https://arxiv.org/html/2609.02231#S5.SS4.p1.1),[Table 2](https://arxiv.org/html/2609.02231#S5.T2.2.1.1.1.6.1)\.
- Guptaet al\.\(2025\)A\. K\. Gupta, F\. Sheth, H\. Shaikh, D\. Kumar, A\. Puniya, D\. Panwar, S\. Chaurasia, and P\. MathurRecruitView: a multimodal dataset for predicting personality and interview performance for human resources applications\.arXiv preprint arXiv:2512\.00450\.Cited by:[§1](https://arxiv.org/html/2609.02231#S1.p4.1),[§2](https://arxiv.org/html/2609.02231#S2.p1.1),[§5\.1](https://arxiv.org/html/2609.02231#S5.SS1.p1.1),[§5\.1](https://arxiv.org/html/2609.02231#S5.SS1.p3.1),[§5\.4](https://arxiv.org/html/2609.02231#S5.SS4.p1.1),[Table 2](https://arxiv.org/html/2609.02231#S5.T2)\.
- Haoet al\.\(2026a\)J\. Hao, S\. Dai, Y\. Zhang, Y\. Liang, J\. Wu, J\. Bao, Y\. Fan, Z\. Ye, Y\. Sun, X\. Zhang,et al\.OralAgent: integrating reasoning, tools, and knowledge for interactive dental image analysis\.arXiv preprint arXiv:2605\.27378\.Cited by:[§2](https://arxiv.org/html/2609.02231#S2.p2.1)\.
- Haoet al\.\(2026b\)J\. Hao, Y\. Liang, L\. Lin, Y\. Fan, W\. Zhou, K\. Guo, Z\. Ye, Y\. Sun, X\. Zhang, Y\. Yang,et al\.OralGPT\-omni: a versatile dental multimodal large language model\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,pp\. 38509–38519\.Cited by:[§2](https://arxiv.org/html/2609.02231#S2.p2.1)\.
- Heet al\.\(2026\)Y\. He, Z\. Kang, K\. Jiang, Q\. Zhang, J\. Qian, C\. Meng, Y\. Feng, Y\. Wang, J\. Dou, A\. Wu,et al\.How order\-sensitive are llms? orderprobe for deterministic structural reconstruction\.arXiv preprint arXiv:2601\.08626\.Cited by:[Limitations](https://arxiv.org/html/2609.02231#Sx1.p4.1)\.
- Hemamouet al\.\(2019\)L\. Hemamou, G\. Felhi, V\. Vandenbussche, J\. Martin, and C\. ClavelHireNet: a hierarchical attention model for the automatic analysis of asynchronous video job interviews\.InAAAI Conference on Artificial Intelligence,External Links:[Link](https://api.semanticscholar.org/CorpusID:69424678)Cited by:[§1](https://arxiv.org/html/2609.02231#S1.p2.1),[§2](https://arxiv.org/html/2609.02231#S2.p1.1)\.
- Honget al\.\(2025\)G\. T\. W\. Hong, W\. Yu, X\. Gu, G\. Wang, G\. Gan, H\. Tang, J\. Cheng, J\. Qi, J\. Ji, L\. Pan, S\. Duan, W\. Wang, Y\. Wang, Y\. Cheng, Z\. He, Z\. Su, Z\. Yang, Z\. Pan, A\. Zeng, B\. Wang, B\. Shi, C\. Pang, C\. Zhang, D\. Yin, F\. Yang, G\. Chen, J\. Xu, J\. Chen, J\. Chen, J\. Chen, J\. Lin, J\. Wang, J\. Chen, L\. Lei, L\. Gong, L\. Pan, M\. Zhang, Q\. Zheng, S\. Yang, S\. Zhong, S\. Huang, S\. Zhao, S\. Xue, S\. Tu, S\. Meng, T\. Zhang, T\. Luo, T\. Hao, W\. Li, W\. Jia, X\. Lyu, X\. Huang, Y\. Wang, Y\. Xue, Y\. Wang, Y\. An, Y\. Du, Y\. Shi, Y\. Huang, Y\. Niu, Y\. Wang, Y\. Yue, Y\. Li, Y\. Zhang, Y\. Zhang, Z\. Du, Z\. Hou, Z\. Xue, Z\. Du, Z\. Wang, P\. Zhang, D\. Liu, B\. Xu, J\. Li, M\. Huang, Y\. Dong, and J\. TangGLM\-4\.5v and glm\-4\.1v\-thinking: towards versatile multimodal reasoning with scalable reinforcement learning\.External Links:[Link](https://api.semanticscholar.org/CorpusID:280049141)Cited by:[Table 1](https://arxiv.org/html/2609.02231#S4.T1.2.1.10.1),[§5\.1](https://arxiv.org/html/2609.02231#S5.SS1.p1.1),[Table 2](https://arxiv.org/html/2609.02231#S5.T2.2.1.1.1.8.1)\.
- Inamet al\.\(2026\)S\. A\. Inam, A\. Kabeer, M\. A\. Abbasi, and A\. A\. KhanIVAS: a multimodal ai system for objective video interview assessment with facial emotion, gaze, and audio analysis\.Edelweiss Applied Science and Technology10\(1\),pp\. 525–543\.Cited by:[§2](https://arxiv.org/html/2609.02231#S2.p1.1)\.
- Kanget al\.\(2026a\)Z\. Kang, J\. Gong, Q\. Chen, H\. Zhang, J\. Liu, R\. Fu, Z\. Feng, Y\. Wang, S\. Fong, and K\. ZhouMultimodal multi\-agent empowered legal judgment prediction\.InICASSP 2026\-2026 IEEE International Conference on Acoustics, Speech and Signal Processing \(ICASSP\),pp\. 12202–12206\.Cited by:[§2](https://arxiv.org/html/2609.02231#S2.p2.1)\.
- Kanget al\.\(2026b\)Z\. Kang, J\. Gong, J\. Yan, W\. Xia, Y\. Wang, Z\. Cheng, W\. Cao, Z\. Wang, Z\. Feng, H\. Ding,et al\.HSSBench: benchmarking humanities and social sciences ability for multimodal large language models\.InInternational Conference on Learning Representations,Vol\.2026,pp\. 74664–74719\.Cited by:[§2](https://arxiv.org/html/2609.02231#S2.p2.1)\.
- Keet al\.\(2024\)P\. Ke, B\. Wen, A\. Feng, X\. Liu, X\. Lei, J\. Cheng, S\. Wang, A\. Zeng, Y\. Dong, H\. Wang,et al\.Critiquellm: towards an informative critique generation model for evaluation of large language model generation\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 13034–13054\.Cited by:[§5\.1](https://arxiv.org/html/2609.02231#S5.SS1.p3.1)\.
- Kimet al\.\(2023\)C\. Kim, J\. Choi, J\. Yoon, D\. Yoo, and W\. LeeFairness\-aware multimodal learning in automatic video interview assessment\.IEEE Access11,pp\. 122677–122693\.External Links:[Link](https://api.semanticscholar.org/CorpusID:264371524)Cited by:[§1](https://arxiv.org/html/2609.02231#S1.p3.1)\.
- Kugoet al\.\(2025\)N\. Kugo, X\. Li, Z\. Li, A\. Gupta, A\. Khatua, N\. Jain, C\. Patel, Y\. Kyuragi, Y\. Ishii, M\. Tanabiki,et al\.Videomultiagents: a multi\-agent framework for video question answering\.arXiv preprint arXiv:2504\.20091\.Cited by:[§2](https://arxiv.org/html/2609.02231#S2.p2.1)\.
- Landy and Farr \(1980\)F\. J\. Landy and J\. L\. FarrPerformance rating\.\.Psychological bulletin87\(1\),pp\. 72\.Cited by:[§1](https://arxiv.org/html/2609.02231#S1.p1.1)\.
- Leonget al\.\(2019\)C\. W\. Leong, K\. Roohr, V\. Ramanarayanan, M\. P\. Martin\-Raugh, H\. Kell, R\. Ubale, Y\. Qian, Z\. Mladineo, and L\. McCullaTo trust, or not to trust? a study of human bias in automated video interview assessments\.arXiv preprint arXiv:1911\.13248\.Cited by:[§2](https://arxiv.org/html/2609.02231#S2.p1.1)\.
- Liet al\.\(2024a\)D\. Li, B\. Jiang, L\. Huang, A\. Beigi, C\. Zhao, Z\. Tan, A\. Bhattacharjee, Y\. Jiang, C\. Chen, T\. Wu,et al\.From generation to judgment: opportunities and challenges of llm\-as\-a\-judge\.arXiv preprint arXiv:2411\.16594\.Cited by:[§5\.1](https://arxiv.org/html/2609.02231#S5.SS1.p3.1)\.
- Liet al\.\(2024b\)D\. Li, X\. Liu, B\. Xing, B\. Xia, Y\. Zong, B\. Wen, and H\. KälviäinenEald\-mllm: emotion analysis in long\-sequential and de\-identity videos with multi\-modal large language model\.arXiv preprint arXiv:2405\.00574\.Cited by:[§1](https://arxiv.org/html/2609.02231#S1.p2.1)\.
- Liet al\.\(2025a\)J\. Li, Y\. Wang, W\. Qian, J\. Hu, Z\. Hu, R\. Hong, and M\. WangListening to the unspoken: exploring’365’aspects of multimodal interview performance assessment\.InProceedings of the 33rd ACM International Conference on Multimedia,pp\. 13909–13916\.Cited by:[§2](https://arxiv.org/html/2609.02231#S2.p1.1)\.
- Liet al\.\(2025b\)X\. Li, Z\. Yan, D\. Meng, L\. Dong, X\. Zeng, Y\. He, Y\. Wang, Y\. Qiao, Y\. Wang, and L\. WangVideochat\-r1: enhancing spatio\-temporal perception via reinforcement fine\-tuning\.arXiv preprint arXiv:2504\.06958\.Cited by:[§2](https://arxiv.org/html/2609.02231#S2.p2.1),[Table 1](https://arxiv.org/html/2609.02231#S4.T1.2.1.12.1),[§5\.1](https://arxiv.org/html/2609.02231#S5.SS1.p1.1)\.
- Liaoet al\.\(2024\)R\. Liao, S\. Song, and H\. GunesAn open\-source benchmark of deep learning models for audio\-visual apparent and self\-reported personality recognition\.IEEE Transactions on Affective Computing15\(3\),pp\. 1590–1607\.Cited by:[§2](https://arxiv.org/html/2609.02231#S2.p1.1)\.
- Liffet al\.\(2024\)J\. Liff, N\. Mondragon, C\. Gardner, C\. J\. Hartwell, and A\. BradshawPsychometric properties of automated video interview competency assessments\.\.Journal of Applied Psychology109\(6\),pp\. 921\.Cited by:[§2](https://arxiv.org/html/2609.02231#S2.p1.1)\.
- Maazet al\.\(2024\)M\. Maaz, H\. Rasheed, S\. Khan, and F\. KhanVideo\-chatgpt: towards detailed video understanding via large vision and language models\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 12585–12602\.Cited by:[§1](https://arxiv.org/html/2609.02231#S1.p2.1),[§2](https://arxiv.org/html/2609.02231#S2.p2.1)\.
- Maude and Kirby \(2022\)J\. M\. Maude and D\. KirbyHolistic admissions in higher education: a systematic literature review\.Journal of Higher Education Theory and Practice22\(8\)\.Cited by:[§1](https://arxiv.org/html/2609.02231#S1.p1.1)\.
- Mujtaba and Mahapatra \(2025\)D\. F\. Mujtaba and N\. R\. MahapatraBehind the screens: uncovering bias in ai\-driven video interview assessments using counterfactuals\.arXiv preprint arXiv:2505\.12114\.Cited by:[§2](https://arxiv.org/html/2609.02231#S2.p1.1)\.
- Nagraniet al\.\(2024\)A\. Nagrani, M\. Zhang, R\. Mehran, R\. Hornung, N\. B\. Gundavarapu, N\. Jha, A\. Myers, X\. Zhou, B\. Gong, C\. Schmid,et al\.Neptune: the long orbit to benchmarking long video understanding\.arXiv preprint arXiv:2412\.09582\.Cited by:[§2](https://arxiv.org/html/2609.02231#S2.p2.1)\.
- Naimet al\.\(2015\)I\. Naim, Md\. I\. Tanveer, D\. Gildea, and E\. HoqueAutomated analysis and prediction of job interview performance\.IEEE Transactions on Affective Computing9,pp\. 191–204\.External Links:[Link](https://api.semanticscholar.org/CorpusID:6579869)Cited by:[§1](https://arxiv.org/html/2609.02231#S1.p2.1),[§2](https://arxiv.org/html/2609.02231#S2.p1.1)\.
- Putraet al\.\(2024\)B\. Putra, K\. Azizah, C\. O\. Mawalim, I\. A\. Hanif, S\. Sakti, C\. W\. Leong, and S\. OkadaMAG\-bert\-arl for fair automated video interview assessment\.IEEE Access12,pp\. 145188–145205\.Cited by:[§2](https://arxiv.org/html/2609.02231#S2.p1.1)\.
- Radfordet al\.\(2023\)A\. Radford, J\. W\. Kim, T\. Xu, G\. Brockman, C\. McLeavey, and I\. SutskeverRobust speech recognition via large\-scale weak supervision\.InInternational conference on machine learning,pp\. 28492–28518\.Cited by:[§4\.1](https://arxiv.org/html/2609.02231#S4.SS1.p3.1)\.
- Renet al\.\(2024\)S\. Ren, L\. Yao, S\. Li, X\. Sun, and L\. HouTimechat: a time\-sensitive multimodal large language model for long video understanding\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,pp\. 14313–14323\.Cited by:[§2](https://arxiv.org/html/2609.02231#S2.p2.1)\.
- Shaoet al\.\(2025\)Y\. Shao, H\. He, S\. Li, S\. Chen, X\. Long, F\. Zeng, Y\. Fan, M\. Zhang, Z\. Yan, A\. Ma,et al\.EventVAD: training\-free event\-aware video anomaly detection\.InProceedings of the 33rd ACM International Conference on Multimedia,pp\. 2586–2595\.Cited by:[§2](https://arxiv.org/html/2609.02231#S2.p2.1)\.
- Shaoet al\.\(2024\)Z\. Shao, P\. Wang, Q\. Zhu, R\. Xu, J\. Song, X\. Bi, H\. Zhang, M\. Zhang, Y\. Li, Y\. Wu,et al\.Deepseekmath: pushing the limits of mathematical reasoning in open language models\.arXiv preprint arXiv:2402\.03300\.Cited by:[§4\.3](https://arxiv.org/html/2609.02231#S4.SS3.p6.1)\.
- Shiet al\.\(2026\)Q\. Shi, Z\. Kang, Y\. Zhou, D\. Weng, and Y\. WuSPADER: step\-wise peer advantage with diversity\-aware exploration rewards for multi\-answer question answering\.arXiv preprint arXiv:2606\.00593\.Cited by:[§4\.3](https://arxiv.org/html/2609.02231#S4.SS3.p3.1)\.
- Shuet al\.\(2025\)Y\. Shu, Z\. Liu, P\. Zhang, M\. Qin, J\. Zhou, Z\. Liang, T\. Huang, and B\. ZhaoVideo\-xl: extra\-long vision language model for hour\-scale video understanding\.InProceedings of the Computer Vision and Pattern Recognition Conference,pp\. 26160–26169\.Cited by:[§2](https://arxiv.org/html/2609.02231#S2.p2.1)\.
- Singhet al\.\(2025\)A\. Singh, A\. Fry, A\. Perelman, A\. Tart, A\. Ganesh, A\. El\-Kishky, A\. McLaughlin, A\. Low, A\. Ostrow, A\. Ananthram,et al\.Openai gpt\-5 system card\.arXiv preprint arXiv:2601\.03267\.Cited by:[Table 1](https://arxiv.org/html/2609.02231#S4.T1.2.1.3.2),[§5\.1](https://arxiv.org/html/2609.02231#S5.SS1.p1.1),[Table 2](https://arxiv.org/html/2609.02231#S5.T2.2.1.1.1.2.1)\.
- Singhaniaet al\.\(2020\)A\. Singhania, A\. Unnam, and V\. AggarwalGrading video interviews with fairness considerations\.arXiv preprint arXiv:2007\.05461\.Cited by:[§2](https://arxiv.org/html/2609.02231#S2.p1.1)\.
- Songet al\.\(2026\)P\. Song, N\. Sarah, K\. Brian Heseung, K\. Trent, F\. Mark, H\. Rodney, and W\. HerringDeadline update, 2025–2026: first\-year application trends through january 1\.Common App reports and insights\.Cited by:[§1](https://arxiv.org/html/2609.02231#S1.p1.1)\.
- Subramaniamet al\.\(2016\)A\. Subramaniam, V\. Patel, A\. Mishra, P\. Balasubramanian, and A\. MittalBi\-modal first impressions recognition using temporally ordered deep audio and stochastic visual features\.ArXivabs/1610\.10048\.External Links:[Link](https://api.semanticscholar.org/CorpusID:16767231)Cited by:[§1](https://arxiv.org/html/2609.02231#S1.p2.1),[§2](https://arxiv.org/html/2609.02231#S2.p1.1)\.
- Sunet al\.\(2026\)G\. Sun, R\. Yu, L\. Yin, Y\. Yang, B\. Zhang, and Z\. XuCoMAI: a collaborative multi\-agent framework for robust and equitable interview evaluation\.arXiv preprint arXiv:2603\.16215\.Cited by:[§2](https://arxiv.org/html/2609.02231#S2.p1.1)\.
- Takeuchi and Koda \(2021\)N\. Takeuchi and T\. KodaInitial assessment of job interview training system using multimodal behavior analysis\.Proceedings of the 9th International Conference on Human\-Agent Interaction\.External Links:[Link](https://api.semanticscholar.org/CorpusID:243864753)Cited by:[§1](https://arxiv.org/html/2609.02231#S1.p3.1)\.
- Tanget al\.\(2025\)B\. Tang, K\. Pan, M\. Zheng, N\. Zhou, J\. Sui, D\. Zhu, C\. Deng, and S\. KuaiPose as a modality: a psychology\-inspired network for personality recognition with a new multimodal dataset\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.39,pp\. 1538–1546\.Cited by:[§2](https://arxiv.org/html/2609.02231#S2.p1.1)\.
- Taoet al\.\(2025\)S\. Tao, J\. Li, Y\. Yan, J\. Zhang, Y\. Gao, H\. Li, S\. Xun, Y\. Fan, H\. Chen, J\. He,et al\.MOSS\-chatv: reinforcement learning with process reasoning reward for video temporal reasoning\.arXiv preprint arXiv:2509\.21113\.Cited by:[§2](https://arxiv.org/html/2609.02231#S2.p2.1)\.
- Teamet al\.\(2023\)G\. Team, R\. Anil, S\. Borgeaud, J\. Alayrac, J\. Yu, R\. Soricut, J\. Schalkwyk, A\. M\. Dai, A\. Hauth, K\. Millican,et al\.Gemini: a family of highly capable multimodal models\.arXiv preprint arXiv:2312\.11805\.Cited by:[Table 1](https://arxiv.org/html/2609.02231#S4.T1.2.1.5.1),[§5\.1](https://arxiv.org/html/2609.02231#S5.SS1.p1.1),[Table 2](https://arxiv.org/html/2609.02231#S5.T2.2.1.1.1.4.1)\.
- Teamet al\.\(2026\)K\. Team, T\. Bai, Y\. Bai, Y\. Bao, S\. Cai, Y\. Cao, Y\. Charles, H\. Che, C\. Chen, G\. Chen,et al\.Kimi k2\. 5: visual agentic intelligence\.arXiv preprint arXiv:2602\.02276\.Cited by:[Table 1](https://arxiv.org/html/2609.02231#S4.T1.2.1.9.1),[§5\.1](https://arxiv.org/html/2609.02231#S5.SS1.p1.1),[Table 2](https://arxiv.org/html/2609.02231#S5.T2.2.1.1.1.9.1)\.
- Wanget al\.\(2025a\)Q\. Wang, Y\. Yu, Y\. Yuan, R\. Mao, and T\. ZhouVideorft: incentivizing video reasoning capability in mllms via reinforced fine\-tuning\.arXiv preprint arXiv:2505\.12434\.Cited by:[Table 1](https://arxiv.org/html/2609.02231#S4.T1.2.1.13.1),[§5\.1](https://arxiv.org/html/2609.02231#S5.SS1.p1.1)\.
- Wanget al\.\(2025b\)X\. Wang, Z\. Kang, W\. Zhai, X\. Lou, Y\. Lai, Z\. Wang, Y\. Wang, K\. Huang, Y\. Wang, P\. Li,et al\.MUCAR: benchmarking multilingual cross\-modal ambiguity resolution for multimodal large language models\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,pp\. 15037–15059\.Cited by:[§2](https://arxiv.org/html/2609.02231#S2.p2.1)\.
- Wiens \(1976\)A\. N\. WiensThe assessment interview\.\.Cited by:[§1](https://arxiv.org/html/2609.02231#S1.p1.1)\.
- Wingate and Bourdage \(2024\)T\. G\. Wingate and J\. S\. BourdageWhat are interviews for? a qualitative study of employment interview goals and design\.Human Resource Management\.External Links:[Link](https://api.semanticscholar.org/CorpusID:268472924)Cited by:[§1](https://arxiv.org/html/2609.02231#S1.p3.1)\.
- xAI \(2025\)xAIGrok 4\.https://x\.ai/news/grok\-4\.Cited by:[Table 1](https://arxiv.org/html/2609.02231#S4.T1.2.1.6.1),[§5\.1](https://arxiv.org/html/2609.02231#S5.SS1.p1.1),[Table 2](https://arxiv.org/html/2609.02231#S5.T2.2.1.1.1.5.1)\.
- Xuet al\.\(2024\)M\. Xu, M\. Gao, Z\. Gan, H\. Chen, Z\. Lai, H\. Gang, K\. Kang, and A\. DehghanSlowfast\-llava: a strong training\-free baseline for video large language models\.arXiv preprint arXiv:2407\.15841\.Cited by:[§2](https://arxiv.org/html/2609.02231#S2.p2.1)\.
- Zhanget al\.\(2025\)B\. Zhang, K\. Li, Z\. Cheng, Z\. Hu, Y\. Yuan, G\. Chen, S\. Leng, Y\. Jiang, H\. Zhang, X\. Li,et al\.Videollama 3: frontier multimodal foundation models for image and video understanding\.arXiv preprint arXiv:2501\.13106\.Cited by:[§1](https://arxiv.org/html/2609.02231#S1.p2.1),[§2](https://arxiv.org/html/2609.02231#S2.p2.1),[Table 1](https://arxiv.org/html/2609.02231#S4.T1.2.1.15.1),[§5\.1](https://arxiv.org/html/2609.02231#S5.SS1.p1.1)\.
- Zhanget al\.\(2023\)H\. Zhang, X\. Li, and L\. BingVideo\-llama: an instruction\-tuned audio\-visual language model for video understanding\.InProceedings of the 2023 conference on empirical methods in natural language processing: system demonstrations,pp\. 543–553\.Cited by:[§2](https://arxiv.org/html/2609.02231#S2.p2.1)\.