E-AVI: Evidence-Grounded Multimodal Assessment for Automated Video Interviews
Haoshen Wang, Dongbo Che, Zeyi Xie, Yuanjie Du, Shicheng Hua, Xingyu Wang
cs.AI
Quick verdict
E-AVI claims evidence-grounded multimodal interview scoring; rank-correlation gains over fine-tuned baselines are plausible, but ablations credit source embeddings more than evidence.
Summary
E-AVI is an evidence-grounded framework for automated video interview assessment. A MiniCPM-o extractor, LoRA-tuned under teacher-generated supervision, emits timestamped text, audio, and visual evidence claims; a dimension-conditioned scorer attends over cached evidence embeddings and pooled source embeddings for each rating dimension, and a GPT-4o organizer serves natural-language feedback and follow-up QA over the shared evidence pool. On RecruitView (2,011 videos), E-AVI reaches MAE 0.607 and Spearman rho 0.541 versus 0.641/0.440 for the strongest fine-tuned baseline (MiniCPM-o-9B); on a private hospitality set (101 hotel employees), it reaches MAE 0.503 and rho 0.639 with the extractor transferred unchanged. Ablations reveal that source embeddings, not the explicit evidence pool, carry most predictive power (No-evidence rho drops only to 0.435; No-source to 0.261). A human audit finds 80.8% of sampled evidence items fully supported, and E-AVI ranks first in human-rated QA quality over GPT-4o and other baselines. Key limitations: gains rest on bootstrap CIs from a small private test set only, the private data is unavailable, extraction errors dominate the difficult-case analysis, and no code is released.
Review complete
Assessment at a glance
6
risk flags identified
● high 0
● medium 3
● low 3
5 strengths · 0 weaknesses · 0 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Extracting timestamped multimodal evidence and scoring with dimension-conditioned attention plus source embeddings improves rank correlation over fine-tuned multimodal baselines on both RecruitView (rho 0.541 vs 0.440) and a private hospitality dataset (rho 0.639 vs 0.481).
- The shared evidence pool supports natural-language feedback and follow-up QA, and human raters prefer E-AVI answers over GPT-4o, Qwen3-VL, MiniCPM-o, and VideoLLaMA2-AV (overall 4.61/5).
- Human auditing shows 80.8% of sampled evidence items are fully supported by the source recordings.
- Ablations show explicit evidence contributes meaningfully to ranking, although source embeddings — not the evidence pool — carry most of the predictive power.
Strengths
Above-average statistical hygiene for a venue-length systems paper
Participant-level splits prevent identity leakage (§3.1); bootstrap CIs and a paired-resampling interval are reported for the small private set (§3.3), and the paper explicitly notes the improvement 'does not establish stability across new splits, hotels, or populations' — an unusually candid inferential statement.
Unusually honest ablation and difficult-case analysis
Table 3 shows removing evidence barely changes MAE (0.607→0.612) while removing source embeddings collapses rho from 0.541 to 0.261, and §3.4 concedes E-AVI loses on 19/30 hardest cases with extraction implicated in 25/30 (83.3%) — analyses many papers would omit.
Concrete, auditable grounding mechanism
Evidence pool as a shared substrate for three downstream tasks
Cross-dataset transfer without extractor fine-tuning
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
E-AVI: Structured Preliminary Review
*Review based on arXiv:2609.20001v1 (17 Sep 2026), full extracted text. Figures are only partially recoverable from text; noted under caveats.*
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 0 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 0 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is E-AVI and what does it do?
E-AVI is a framework that extracts timestamped text, audio, and visual evidence from interview recordings, scores rating dimensions by attending over that evidence plus source-level embeddings, and supports natural-language feedback and follow-up questions over the shared evidence pool.
How accurate is E-AVI compared with baselines?
On the public RecruitView dataset it reports MAE 0.607 and Spearman rho 0.541, versus 0.641 and 0.440 for the strongest fine-tuned baseline (MiniCPM-o-9B); on a private hospitality dataset of 101 hotel employees it reports MAE 0.503 and rho 0.639.
What does the evidence pool contain and is it trustworthy?
It holds timestamped natural-language claims from the text, audio, and visual channels (e.g., a pause between 00:08-00:12 or eye contact between 00:25-00:27). A human audit of 870 items judged 80.8% fully supported by the source recordings, though the audit measured support only, not completeness.
Can I reproduce E-AVI or get its code and data?
No code or evidence-pool release is stated in the paper, and one of the two evaluation datasets is private. The public RecruitView dataset plus the detailed hyperparameters make a reimplementation possible in principle, but the exact gains have not been independently replicated.
Is the evidence pathway actually what drives E-AVI's performance?
Ablations show removing the source embeddings drops Spearman from 0.541 to 0.261 while removing the evidence pool only lowers MAE from 0.607 to 0.612 — so the source-embedding pathway, not the explicit evidence, carries most of the predictive power. The evidence's main demonstrated value is inspectability and QA.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-3 are recovered only as flattened OCR text in the extraction; exact Figure 2 deletion curves and Figure 3 bar-to-system mapping could not be independently verified.
- Private test-set size (participant count) is never stated, so the bootstrap interval width cannot be sanity-checked.
- arXiv v1 full text appears otherwise complete (all sections and tables present); no appendix or code/data statement exists in the source.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.