PaperlayerView exact revision on arXiv

Source-linked AI summary

Can We Trust Video Hallucination Detectors? VidHalLoc for Evaluating the Evaluators

Xinyu Chen, Adnan Mahmood, Mark Dras

arXiv:2609.09895v1cs.CV

TL;DR

Video-language models and agents can produce hallucinations that conflict with spatiotemporal evidence, while existing benchmarks and heterogeneous detector mechanisms make reliability difficult to compare. VidHalLoc introduces a unified diagnostic benchmark and VideoHALO construction workflow; across fifteen systems, four dedicated detectors peak at 34.63% Overall accuracy. The results identify limited reliability across video hallucination types and emphasize dynamic hallucinations as a detector-development target.

  • Problem

    Existing benchmarks mainly assess model hallucinations, while heterogeneous detector categories and evaluation targets hinder meaningful reliability comparisons.

  • Method

    VidHalLoc uses a unified diagnostic protocol with 2,000 adversarial Video Question Answering and Video Captioning samples spanning ontology and dynamic hallucinations, constructed with VideoHALO.

  • Results

    Mean Overall accuracy is lower on dynamic hallucinations than ontology types, and four dedicated detectors peak at 34.63% Overall accuracy versus 83.63% for the top LVLM.

  • Takeaways & Limitations

    Dynamic hallucinations are a primary target for detector development, and evaluation should integrate factual acceptance with counterfactual rejection.

  • Takeaways & Limitations

    The benchmark uses single-detail adversarial attacks and balanced categories, so it does not test compound errors or deployment-frequency distributions.

Abstract

from arXiv · show

Video-language models and video agents can produce hallucinations that conflict with spatiotemporal evidence. Existing benchmarks mainly evaluate model hallucinations, and heterogeneous mechanisms make detector reliability difficult to compare. We introduce VidHalLoc, a benchmark that evaluates hallucination detection methods under a unified diagnostic evaluation protocol using 2,000 adversarial hallucination samples across Video Question Answering and Video Captioning tasks, spanning Ontology and Dynamic hallucination categories. To construct VidHalLoc efficiently, we introduce VideoHALO, a Harness Engineering-informed multi-agent workflow that decomposes data construction into four executable stages supported by a memory system and a communication protocol. Evaluation of fifteen methods reveals that the four dedicated detectors peak at an Overall accuracy of only 34.63%, indicating limited reliability across video hallucination types [Dataset Repository: https://huggingface.co/datasets/wesfggfd/VidHalLoc].

1 Introduction

Video-language systems can hallucinate against spatiotemporal evidence, while existing evaluations make detector reliability difficult to compare. VidHalLoc addresses this gap with a unified benchmark, taxonomy, and construction workflow, finding limited dedicated-detector accuracy.

  • Motivation: Video hallucination is generated content that misaligns with observable evidence across entities, actions, temporal relations, and camera transitions.Video requires tracing evidence across spatiotemporal dimensions rather than only image-level content.
  • Motivation: Most existing benchmarks assess whether models hallucinate rather than whether detection methods are reliable, and heterogeneous categories hinder meaningful comparisons.Fine-grained video annotation also requires costly evidence localization and manual verification.
  • Evaluation gap and benchmark result: 34.63% Overall accuracy is the peak for four dedicated detectors, versus 83.63% for Flash, revealing limited reliability across evaluated video tasks.The benchmark evaluates fifteen systems, including LVLMs, video agents, and dedicated detection methods.
  • Diagnostic framework: The diagnostic framework organizes hallucinations into five ontological and three dynamic categories for evaluating detection reliability.The taxonomy covers entity-related errors alongside actions, temporal relations, and camera predicates.
  • VidHalLoc: VidHalLoc evaluates hallucination detectors on 2,000 adversarial samples across Video Question Answering and Video Captioning under a unified protocol.The benchmark spans ontology and dynamic hallucination categories.
  • VideoHALO: VideoHALO constructs the benchmark through a harness-engineering-informed multi-agent workflow supported by hierarchical memory and a unified communication protocol.The workflow decomposes benchmark construction into executable stages and is reported to achieve higher throughput than human experts with high audited category accuracy.

2 Related Work

Prior video hallucination detectors use varied units, mechanisms, and task protocols, limiting reliability comparisons. VidHalLoc provides a unified taxonomy and adversarial protocol while addressing the cost of fine-grained benchmark construction.

  • Video hallucination detection: Video detection must connect textual claims with evidence distributed across frames and events.Methods include embedding matching, learned entailment, structured verification, and internal model signals.
  • Hallucination benchmarks: VidHalLoc organizes hallucinations into ontology and dynamic categories covering entities, attributes, relations, actions, temporal order, and camera transitions.The taxonomy separates entity-related misalignments from motion and temporal inconsistencies.
  • Video hallucination detection: VidHalLoc evaluates representative detection methods under a unified hallucination taxonomy and adversarial protocol to enable reliability comparisons.Table 1 compares detection methods, while existing approaches differ in evaluation units, mechanisms, and protocols.
  • Hallucination benchmarks: Existing video benchmarks examine task-specific model hallucinations but do not ensure detector reliability across diverse hallucination types.Fine-grained type annotations also make manual video dataset construction costly.

3 VIDHALLOC Benchmark

VIDHALLOC is a 2,000-sample benchmark for diagnosing video hallucination detectors across Video QA and Video Captioning, organized into Ontology and Dynamic categories. VIDEOHALO constructs the benchmark through a four-stage multi-agent workflow with hierarchical memory, structured communication, and human auditing.

  • VIDHALLOC contains 2,000 adversarial hallucination samples spanning Video QA and Video Captioning, with Ontology and Dynamic hallucination categories.
  • VIDEOHALO: VIDEOHALO decomposes benchmark construction into four executable stages coordinated by role-specific agents, hierarchical memory, and a unified communication protocol.The stages cover category retrieval, fact extraction and reflection, adversarial-pair generation and verification, and reliability validation.
  • Data construction: The dataset is built from 31,771 videos using quality filtering, distribution normalization, representation validation, and human auditing.CLIP supports dataset-wide visual diversity, while LaViLa captures temporal changes within videos.
  • Data quality: Human auditing reports 98.75% sample accuracy, 96.63% category agreement, and multiclass Cohen’s κ of 0.962.
  • Data statistics: Each of the eight hallucination categories contains exactly 250 adversarial samples, drawn from 1,090 unique source videos.

4 Experiments

Experiments evaluate fifteen systems and show that video hallucination evaluation remains difficult, especially for Dynamic hallucinations. Dedicated detectors transfer unevenly across hallucination types and do not reliably resolve adversarial attacks.

  • Evaluation on VIDHALLOC: Most LVLMs and all three video agents score at least 17.60% and 15.60% lower Overall accuracy on Dynamic than Ontology hallucinations, respectively.Dynamic categories require reasoning across actions, event order, and camera transitions.
  • Evaluation on VIDHALLOC: Model scale does not correlate directly with performance, while Gemini-3-Flash reaches 55.77% on APH Video Captioning and 58.62% on CPH Video QA.GPT-5 scores 35.71% Overall on CPH Video Captioning versus 59.52% for Qwen3.6-27B.
  • Detection methods: None of the four dedicated detectors correctly resolves a majority of adversarial attacks, with performance deteriorating further on Dynamic hallucinations.
  • Detection methods: FIFA peaks at 64.44% on EQH Video QA but falls to 7.14% on CPH Video Captioning, while Owl-Con drops from 71.43% on EEH Video Captioning to 9.43% on TRH Video QA.
  • Detection methods: Embedding similarity, learned entailment, and structured verification transfer unevenly across hallucination categories under the unified protocol.PAC-S and EMScore do not explicitly encode event order, while FIFA depends on both fact decomposition and evidence verification.

5 Conclusion

VIDHALLOC evaluates hallucination detectors with adversarial candidates across multiple video hallucination categories, while VIDEOHALO provides a multi-agent construction workflow. Results identify Dynamic hallucinations as a primary detector-development target.

  • VIDHALLOC supports integrated evaluation of factual acceptance and counterfactual rejection through adversarial candidates across hallucination categories.
  • The four dedicated detectors reach a peak Overall accuracy of only 34.63%, versus 83.63% for the top LVLM.
  • Mean Overall accuracy across fifteen evaluated systems is lower on Dynamic hallucinations than on Ontology types.

Limitations

The benchmark’s conclusions are bounded by its adversarial single-detail design, balanced categories, moderate video lengths, and confounded system configurations.

  • Limitations: VidHalLoc targets single-detail adversarial attacks, so it does not evaluate compound errors or deployment-frequency distributions.The authors propose authentic responses and target-distribution evaluation as extensions.
  • Limitations: The benchmark uses 1,090 videos averaging 48.63–61.05 seconds, leaving detector performance on longer recordings untested.Longer videos would test integration of distant evidence and event relations.
  • Limitations: Because systems vary in visual encoders, frame sampling, language models, and inference pipelines, the evaluation cannot isolate individual component contributions.Single-factor evaluations are needed to identify design choices affecting detection reliability.

Appendices

The appendices document VIDEOHALO and provide uncertainty analyses for the human audit and main experiments, alongside qualitative examples.

  • Appendices: The appendices document VIDEOHALO, uncertainty analyses of the human audit and main experiments, and qualitative examples.

A Data Processing

The data-processing pipeline collects videos from public datasets, filters and normalizes them, and validates inter-video diversity and intra-video temporal richness before VIDEOHALO.

  • Video Collection: 31,771 source videos are collected from five public datasets consisting of authentic recordings and purpose-recorded scenes.
  • Filter and Normalization: Quality filtering removes duplicates and checks media integrity and visual richness, while distribution normalization quantifies motion using consecutive pixel differences.
  • Representation Validation: Representation validation uses CLIP features for inter-video semantic breadth and LaViLa encodings for intra-video temporal dynamics.

B Multi-Agent Video Data Construction Workflow

VIDEOHALO constructs benchmark samples through four verifiable, agent-specialized stages coordinated by hierarchical memory, structured communication, and state-preserving validation.

  • B. Multi-Agent Video Data Construction Workflow: VIDEOHALO decomposes benchmark construction into four verifiable subtasks executed by specialized agents with strict operational boundaries.
  • B. Multi-Agent Video Data Construction Workflow: The workflow separates discovery, generation, and verification, preserving traceability from each finalized sample to its visual evidence.
  • B. Multi-Agent Video Data Construction Workflow: A dual-layer memory system combines global construction constraints with category-specific definitions, examples, boundaries, and retrieval rules.
  • B. Multi-Agent Video Data Construction Workflow: A schema-driven communication protocol preserves verified task state, prevents overwriting accepted conclusions, and allows at most one focused visual re-inspection.
  • B. Multi-Agent Video Data Construction Workflow: Algorithm 1 formalizes agent specifications, task orchestration, state transitions, validity checks, and final dataset projection.

C External Human Audit

The external audit samples 100 instances from each of eight categories and uses independent worker reviews plus an author reference to assess label quality.

  • Sampling and Reviewers: 800 audited samples cover eight categories, with 100 instances sampled per category.The audit set was drawn from the 2,000-sample benchmark.
  • Task Templates: The audit uses exact source templates for Video QA categories and caption formats spanning ontology and dynamic hallucinations.The listed templates include temporal-order, camera-edit, observable-fact, and action questions.
  • Annotation Protocol: Two workers independently reviewed each sample after calibration, while one author reviewed all samples to establish reference labels.The workers’ judgments and original labels were hidden, and presentation order was independently randomized.
  • Annotation Protocol: Workers verified video support for factual answers, contradiction of counterfactual answers, and difference only in the targeted content.Each sample received a final category label after review.
  • Metrics: Sample accuracy measures agreement with author reference labels, while worker agreement and Cohen’s κ use the workers’ independent annotations.Per-category metrics are computed within equally sized 100-sample strata, making overall values macroaverages.

D Main Experiment Settings

The main experiment evaluates fifteen methods on a stratified 800-item subset using fixed thresholds, category-aware metrics, and comparisons between Ontology and Dynamic hallucinations.

  • Evaluation Subset and Split: 800 evaluation items are selected as 100 instances from each of eight categories, with a 100-item calibration set drawn from distinct videos.The evaluation set spans 436 unique videos and has no overlap with calibration data.
  • VideoHALO Workflow: VideoHALO initializes six agents with dual-layer memory and executes staged feasibility, fact extraction, grounding, and structural checks to produce accepted samples.The taxonomy requires one normalized atomic fact, one hallucination category, and one conflict slot per contradicted fact.
  • Threshold Settings: The fixed threshold is selected on validation scores and remains unchanged across evaluation splits, tasks, and hallucination categories.Overall accuracy jointly evaluates factual acceptance and counterfactual rejection.
  • Primary Evaluation: 19.12 percentage points separates Ontology and Dynamic accuracy within matched videos, close to the 19.53-point full-evaluation gap.The primary analysis contains 500 Ontology and 300 Dynamic samples per method.

E.2 Method-Level Intervals

Method-level analyses show that most detectors perform better on Ontology than Dynamic hallucinations, while isolated factual or counterfactual successes often fail to coincide.

  • Method-Level Differences: 14 of 15 methods have strict Ontology-minus-Dynamic intervals entirely above zero; LLaVA-NeXT-Video is the sole exception at −0.87 points.Its interval spans [−3.15, 1.17].
  • Dynamic-Sample Behavior: EMScore and Owl-Con more often accept factual answers on Dynamic samples but less often reject counterfactual answers than on Ontology samples.This pattern reflects greater acceptance of candidates, including incorrect counterfactual answers, on Dynamic samples.
  • Outcome Breakdown: PAC-S and EMScore can resolve individual factual or counterfactual assertions, yet their correct classifications rarely coincide on the same instance.The outcome breakdown separates isolated decisions from joint correctness.
  • Representative Cases: Only GPT-5, Deep Video Discovery, and FIFA jointly accept the factual answer and reject the counterfactual answer in the representative APH example.The table marks correct and incorrect decisions with green checks and red crosses.
  • Representative Cases: None of the evaluated methods correctly resolve both candidates in the representative CPH example.Most methods accept or reject both options simultaneously, while GPT-5 reverses the required labels.
Loading 2609.09895v1…