Source-linked AI summary
Beneath the Scores: Rethinking Hallucination Evaluation for Video Understanding Models
Shuzhi Gong, Fengze Sun, Yuansan Liu
TL;DR
Fragmented benchmarks make it difficult to determine which stages of multi-stage video agents contribute to downstream errors. The paper organizes benchmarks by stage and uses controlled interventions to measure stage effects on a shared task. Grounding has the largest measured downstream impact, while conventional benchmark scores do not reliably predict causal cascade sensitivity.
Problem
Existing video benchmarks evaluate stages on different datasets and protocols, so their scores cannot reliably diagnose which stage contributes to downstream hallucination.
Source: §1 · p. 2Method
The authors overwrite individual stages on a shared item set while holding the downstream task fixed, measuring how interventions change downstream performance.
Source: §1 · p. 2Results
Across three architectures, replacing a length-matched random window with oracle evidence improves accuracy by 17–18 points, roughly four times substantial Watching corruption's effect.
Source: §1 · p. 2Takeaways & Limitations
Wrong evidence is more harmful than missing evidence, and existing benchmark scores do not reliably predict causal cascade sensitivity.
Source: §1 · p. 2, §5 · p. 8Only one checkpoint per architecture is evaluated, and the nine-configuration system-level audit supports only a finding of no detectable relationship.
Source: §5 · p. 8
Abstract
from arXiv · showhide
Video understanding is increasingly performed by multi-stage LLM agents that separate temporal grounding, visual observation, and reasoning. Yet these stages are typically evaluated on different benchmarks and distributions, making it difficult to determine where hallucinations originate. We first organize existing benchmarks around these stages and show that their scores provide inconsistent diagnostic signals: stronger stage-level performance does not reliably imply lower downstream hallucination, and even benchmarks targeting the same capability can disagree. We therefore introduce a causal stage-intervention protocol that overwrites individual stages while holding the downstream task fixed. Across 60,008 runs on three video-agent architectures, we find that grounding is the dominant source of downstream error, with roughly four times the causal impact of corrupting visual observations. Successful grounding depends primarily on locating the correct region rather than precise temporal overlap, explaining why standard mIoU metrics poorly predict downstream reliability. We further find that incorrect evidence is substantially more harmful than missing evidence. Finally, auditing existing benchmarks against these interventions reveals that their scores do not reliably predict causal cascade sensitivity and can fail under distribution shift. These results motivate intervention-based, stage-aware evaluation for trustworthy video agents.
1 Introduction
Multi-stage video agents create several possible sources of downstream error, but fragmented benchmark scores cannot reliably diagnose which stage is responsible. The paper replaces cross-benchmark comparisons with controlled stage interventions and finds grounding has the largest downstream effect.
1 Introduction
- Video agents divide work among Grounding, Watching, and Reasoning, creating multiple points where errors can arise and propagate.A wrong moment, a misread moment, or faulty reasoning over valid evidence can each affect the final answer.Source: §1 · p. 1, Figure 1 · p. 2
- Existing benchmark profiles are inconsistent: stronger upstream scores do not reliably predict downstream hallucination, and benchmarks for the same stage can disagree.Because these measures use different datasets and protocols, their correlations cannot identify which stage changes final outcomes.Source: §1 · pp. 1–2
- The authors overwrite one stage while holding the downstream task fixed, testing oracle, random, off-target, and dilated Grounding windows and omitted or replaced Watching frames.The study uses 2,522 questions, 227 videos, and 60,008 intervened runs across three architectures.Source: §1 · p. 2
- Replacing a length-matched random window with oracle evidence improves accuracy by 17–18 points across three architectures, roughly four times the effect of substantial Watching corruption.On-target evidence matters more than precise temporal overlap, and incorrect evidence is more harmful than missing evidence.Source: §1 · p. 2
- The study also finds that existing answer- and hallucination-level scores can fail to reflect causal cascade sensitivity.Its evidence-level diagnosis finds that 42–65% of correct answers can be unsupported by the video.Source: §1 · pp. 2–3
2 Preliminary Study
The preliminary study formalizes video agents as Grounding, Watching, and Reasoning stages, then organizes benchmarks around the capabilities they probe. Their results show capability scores do not consistently track hallucination robustness or reveal where downstream failures originate.
2.1 Pipeline Formalization and Parameters
- Grounding localizes question-relevant temporal evidence, Watching extracts visual information, and Reasoning integrates it with the question to produce an answer.The functional stages need not map to explicit modules; the sequential pipeline is G → W → R.Source: §2.1 · p. 3
2.2 Models, Benchmarks and Metrics
- The benchmark map assigns NEXT-GQA and CG-BENCH to Grounding, temporal-grounding subsets of LVBENCH and ARGUS to Watching, and hallucination or reasoning measures to downstream evaluation.ARGUS spans Watching and Reasoning because it evaluates faithfulness in generated visual descriptions.Source: §2.2 · p. 3, Table 1 · p. 3
2.3 Results and Limits of Fragmented Evaluation
- Strong Grounding or Watching scores do not imply lower downstream hallucination: EVA leads on CG-Bench accuracy and LVBench yet performs poorly on ELV-Halluc and VideoHallucer.ARGUS offers an intermediate signal by assessing hallucination and omission in visual descriptions, but uses a separate dataset.Source: §2.3 · p. 3, Table 2 · p. 4
- Benchmark profiles can rank capabilities but cannot identify where downstream hallucinations originate; changing frame counts or tool budgets yields no consistent robustness improvement.The same nominal stage can produce conflicting conclusions across datasets, and anomalous ELV-HALLUC scores leave model failure versus metric failure ambiguous.Source: §2.3 · p. 4
- Because the benchmark comparisons are correlational and use different datasets, they cannot determine which stage caused an error or how much correcting it would help.The next study therefore uses controlled stage interventions on a shared item set.Source: §2.3 · p. 4
3 Causal Stage Intervention
The intervention protocol changes one stage while keeping the downstream task fixed, enabling direct measurement of stage effects. Across agents, grounding has the largest downstream impact, while on-target windows tolerate substantial temporal widening and misleading evidence is more harmful than missing evidence.
3.1 Intervention Protocol
- The study intervenes on 2,522 questions from 227 long videos, replacing grounding windows and, with grounding fixed to oracle, omitting or replacing sampled frames.The protocol covers 60,008 intervened runs across three architectures; grounding windows include full-video, predicted, oracle, random, and off-target conditions.Source: §3 · p. 4, §3.1 · p. 4
3.2 Analysis and Findings of Causal Stage Intervention
- Expanding an oracle window up to 8× causes little degradation, indicating that locating the relevant region matters more than precise temporal overlap.Performance drops only once the window becomes very broad.Source: Figure 2 · p. 5
- 16.6–17.9 pp: replacing a length-matched random window with the oracle span improves accuracy across all three architectures, significantly more than moderate frame corruption.All oracle−random differences are significant at p < 0.001.Source: Table 3 · p. 5, §3.2 · p. 5
- Deployed grounders recover only 5.0 pp for EVA and 4.5 pp for VIDEOMIND, leaving respective gaps of 12.5 and 13.4 pp against oracle grounding.VIDEOMIND’s predicted grounding is only marginally better than exposing the whole video, while EVA obtains a clearer gain.Source: §3.2 · pp. 5–6
- Replacing frames with plausible content from elsewhere in the video causes a clearer accuracy drop than omitting them, even when 75% of frames are removed.Degradation from omission is small and not consistently significant.Source: §3.2 · p. 6
- Holding the downstream items fixed avoids confounding stage sensitivity with distribution differences across benchmarks.On the shorter DR.V-BENCH slice, predicted grounding does not outperform exposing the whole video for EVA or VIDEOMIND.Source: §3.2 · p. 6
4 Auditing the Benchmark Ecosystem Against the Causal Standard
The audit finds that conventional benchmark scores do not reliably indicate which stage causally determines an answer, and correctness alone can conceal unsupported answers. It also shows that hallucination scores may lose discriminative power in a shifted evaluation regime, motivating evaluation of on-target grounding, metric behavior, and evidential support.
4.1 No benchmark score predicts what causally matters
- Existing grounding, faithfulness, and hallucination scores show no detectable association with causal cascade sensitivity at the system level, while item-level grounding quality predicts correctness but not causal dependence.Cascade sensitivity measures the accuracy change when grounding is replaced by an oracle span rather than a length-matched random span.Source: Figure 3 · p. 6, §4 · p. 6, §4.1 · p. 7
4.2 Beyond right and wrong: correct answers can still be hallucinated
- 42–65% of correct answers lack sufficient visual evidence, and DR.V separates agents by more than 20 percentage points in hallucination rate despite smaller accuracy differences.Correctness and evidential support capture different aspects of model behavior.Source: §4.2 · p. 7
4.3 Metric failure versus model failure
- DR.V saturates near 100% on CG-BENCH despite an oracle-span intervention improving accuracy by about 17 pp, consistent with a metric mismatch under distribution shift.The proposed explanation is that long free-form answers yield many checkable claims while diagnosis uses sparsely sampled frames; ELV-HALLUC also gives EVA and VIDEO-R1 near-zero or negative scores outside the range observed for validated monolithic models.Source: §4.3 · p. 8
4.4 Implications for trustworthy evaluation
- Grounding evaluation should measure whether the relevant evidence is included within a practically useful window, not only its overlap with the annotation.The authors suggest checking whether annotated evidence is contained within the selected window at a useful dilation.Source: §4.4 · p. 8
- Before interpreting a hallucination score, test whether it responds to a controlled intervention that changes model behavior, such as an oracle-span probe.A nearly unchanged score despite a large, independently verified accuracy change should not be treated as reliable in that regime.Source: §4.4 · p. 8
- Reliability evaluation should record whether available visual evidence supports an answer in addition to whether the answer is correct.Stage-annotated datasets such as CG-BENCH provide clue intervals that make evidence-level evaluation feasible at benchmark scale.Source: §4.4 · p. 8
5 Conclusion
Existing benchmark scores can rank capabilities but cannot reliably diagnose where multi-stage video-agent failures originate. Controlled interventions identify grounding as the dominant downstream bottleneck, while several study boundaries qualify the conclusions.
5 Conclusion
- Grounding is the dominant downstream bottleneck; being on-target matters more than precise overlap, and incorrect evidence is more harmful than missing evidence.The benchmark audit also finds that conventional capability and hallucination scores do not reliably predict causal cascade sensitivity.Source: §5 · p. 8
- The CG-BENCH interventions measure downstream answer error, not hallucination defined as unsupported content.Unsupported-content hallucination is examined separately with DR.V-BENCH, whose diagnosis instrument retains discriminative power.Source: §5 · p. 8
- One checkpoint per architecture leaves cross-agent differences confounded with model identity, and the system-level audit covers only nine configurations.The authors describe the system-level intervals as supporting only “no detectable relationship”; the item-level analysis uses 2,522 items per agent.Source: §5 · p. 8
- EVA’s full condition used a forced-answer prompt, and 11% of multi-interval clues were served as their temporal hull.The forced-answer prompt was used to suppress tool-denied refusals, which were flagged per run.Source: §5 · p. 8
NeurIPS Paper Checklist
The checklist records the authors’ responses on reproducibility, experimental reporting, ethics, impacts, and asset handling. It also notes that the paper’s core research involves LLMs and points to supplementary benchmark analysis.
NeurIPS Paper Checklist
- The authors state that the contributions are reflected in the abstract and introduction and that limitations and corresponding future work are discussed.They also state that model details are disclosed for reproducibility and that experimental-setting details are provided.Source: NeurIPS Paper Checklist · pp. 12–14
- The authors report standard metrics with error bars and state that computer-resource details are provided.They also say code and instructions will be released upon acceptance.Source: NeurIPS Paper Checklist · pp. 14–15
- The checklist states that the research conforms with the NeurIPS Code of Ethics and that societal impacts are discussed in the conclusion.It also says existing assets are properly credited, no new assets are released, and the paper involves no human-subject research.Source: NeurIPS Paper Checklist · pp. 15–17
- The authors identify LLMs as important components of the paper’s core research.Source: NeurIPS Paper Checklist · p. 18
A Additional Analysis of the Preliminary Study
- The appendix expands the preliminary observational results with further interpretation of benchmark profiles and parameter sweeps.Source: A Additional Analysis of the… · p. 19
A.1 Detailed Benchmark Comparisons
Detailed comparisons show that benchmark scores can disagree across capabilities and models, including when the same inference setting changes. These observational patterns motivate, but do not replace, controlled interventions.
A.1 Detailed Benchmark Comparisons
- EVA leads on CG-Bench accuracy and LVBench yet performs poorly on hallucination-oriented metrics; ELV-Halluc also gives EVA and VIDEO-R1 near-zero or negative values outside its usual operating regime.VIDEOMIND has the strongest grounding mIoU, while EVA has the best CG-Bench accuracy; Qwen2.5-VL-7B has weaker LVBench accuracy but better ELV-Halluc and VideoHallucer results than EVA.Source: §A.1 · p. 19
A.2 Detailed Parameter-Sweep Analysis
- For EVA, sampled-frame increases raise CG-Bench long-video accuracy from 36.93% to 41.00% but decrease NeXT-GQA mIoU.ARGUS changes mildly for EVA and VIDEOMIND, while VideoHallucer is largely flat for EVA and degrades more substantially for VIDEO-R1.Source: §A.2 · p. 19
A.3 Detailed Parameter-Sweep Analysis
- For EVA, sampled-frame increases raise CG-Bench long-video accuracy from 36.93% to 41.00% but decrease NeXT-GQA mIoU.ARGUS changes mildly for EVA and VIDEOMIND, while VideoHallucer is largely flat for EVA and degrades more substantially for VIDEO-R1.Source: §A.3 · p. 19
A.4 Inference-Parameter Sweeps
- More sampled frames or tool calls do not uniformly improve hallucination robustness, and the same inference change can move benchmarks in opposing directions.For EVA, frame increases improve CG-Bench while decreasing NeXT-GQA; tool-budget increases do not monotonically improve Grounding or hallucination metrics.Source: §A.4 · p. 19, Figure 5 · p. 20, Figure 6 · p. 20
A.5 Interpretation
- These observational differences cannot attribute failure to model capability alone because the benchmarks use different datasets and protocols.They motivate, but cannot replace, the controlled intervention study.Source: §A.5 · p. 20
B Full Intervention Ladders
The intervention ladders show differing effects from omitted versus fabricated evidence and from expanding oracle windows. The system-level audit compares retained benchmark scores with causally measured cascade sensitivity across nine configurations.
B Full Intervention Ladders
- At corruption rate ρ=0.75, clean−corrupted contrasts are omission +5.7 pp [−0.3, +11.6] for EVA and +1.7 pp [−3.0, +6.2] for VIDEO-R1; fabrication +6.7 pp [+1.0, +12.6] and +4.3 pp [0.0, +8.9], respectively.Even discarding three quarters of evidence frames within a correct window is not reliably detectable, whereas replacing them with plausible frames from elsewhere in the video is.Source: B Full Intervention Ladders · p. 20, Table 4 · p. 20
- Oracle-window dilation has a plateau through 8× across all three agents; only 16× expansion degrades accuracy.The table evaluates accuracy when the oracle window is expanded about its centre.Source: Table 5 · p. 21
C Audit Details
- The system-level audit uses nine agent×frame-budget configurations and measures cascade sensitivity as oracle−random on set S with 400 items.Retained benchmark scores were recomputed from per-item outputs of the preliminary study.Source: C Audit Details · p. 21
- Every reported benchmark-score correlation with cascade sensitivity has a 95% BCa interval spanning zero.Reported Spearman correlations range from −0.17 to +0.25 across LVBench-TG, ARGUS ROUGE-L, and VideoHallucer scores.Source: Table 6 · p. 21
D Protocol Details
Video agents divide understanding into stages, but existing hallucination evaluations mostly judge final answers and provide little evidence about where errors originate. The protocol details include controls for grounding interventions and describe the limits of stage-specific diagnosis.
D Protocol Details
- A forced-answer follow-up reduced EVA’s refusal rate from 20% to 4% when its frame-selection tool was denied; nudged runs were flagged in the logs.Without this control, oracle–full contrasts would be inflated by a format artefact rather than a grounding effect.Source: D Protocol Details · p. 21
- Items with multiple ground-truth clue intervals (11%) were represented by their temporal hull because each agent consumes one contiguous window.Source: D Protocol Details · p. 21
- Dr.V diagnosis failure rates were 3.3–16.7% on DR.V-BENCH and 1.1–3.4% on CG-BENCH; graded CG-BENCH verdicts remained flat across cells.The graded verdict values ranged from 0.746 to 0.768.Source: D Protocol Details · p. 21
- The 60,008-run campaign used seeded manifests and span generation to give all three agents byte-identical spans and reproducible results.Run records include condition, spans, frame timestamps, evidence rate, IoU, parse path, and latency.Source: D Protocol Details · p. 21
E.1 Video Understanding Pipelines
- Video agents decompose understanding into stages such as temporal localization, visual observation, and answer synthesis, while end-to-end evaluation offers limited insight into error propagation.Source: §E.1 · p. 22
E.2 Video Hallucination Evaluation
- Hallucination benchmarks assess final outputs but provide little evidence about whether errors originate in localization, observation, or reasoning over evidence.Source: §E.2 · p. 22
E.3 Benchmark Diagnosis
The paper examines whether existing benchmark scores diagnose hallucinations in multi-stage video agents and uses controlled stage interventions as a reference for auditing their predictive value.
E.3 Benchmark Diagnosis
- Prior benchmark-diagnosis work examines saturation, dataset bias, ranking inconsistency, shortcut learning, and sensitivity to inference configurations.Source: §E.3 · p. 22
- Controlled intervention overwrites one pipeline stage while holding the downstream task fixed, measuring stage attribution on a shared stage-annotated item set.This contrasts with observational benchmark comparisons that inherit distribution confounds.Source: §E.3 · p. 23
- Existing benchmarks can yield inconsistent, incomplete, or misleading signals when used collectively to evaluate modern agentic architectures.Source: §E.3 · p. 23
F EVA’s Sweep on CG-Bench
- Table 7 reports EVA variants on CG-Bench and identifies maximum tool-call turns and frames per tool call as configuration dimensions.Source: Table 7 · p. 23
I Per-Sample Cascade Analysis: VideoMind Grounding Quality on CGBench
The per-sample analysis relates VIDEOMIND-7B’s grounding IoU to final-answer correctness on 3,000 CGBench questions, using IoU ≥0.1 to mark partial localization.
I Per-Sample Cascade Analysis: VideoMind Grounding Quality on CGBench
- Mean grounding IoU was 0.065 across 3,000 CGBench questions, indicating that accurate temporal localization was rare and the primary bottleneck.IoU ≥0.1 separates samples with at least partial localization from those where the Grounder missed entirely.Source: I Per-Sample Cascade Analysis:… · p. 23
- Table 8 covers EVA sweep configurations on NExT-GQA, with maximum tool-call turns and frames per tool call as configuration dimensions.Source: Table 8 · p. 24
- Table 9 reports accuracy by LVBench category, but the supplied caption gives no category-level values or comparison.Source: Table 9 · p. 24
J Performance on ELV-Halluc Dataset
The supplied tables describe cost, grounding-cascade, ELV-Halluc, and VideoHallucer evaluations, but provide little performance data beyond the cascade caption’s reported pattern.
J Performance on ELV-Halluc Dataset
- VideoHallucer tables report Basic, Hallucination, and Both Accuracy across categories and configurations for Video-R1, VideoMind-7B, and EVA.The Both Accuracy table counts cases where both basic and hallucination questions are correct.Source: K Performance on VideoHallucer · p. 24, Table 13 · p. 27, Table 14 · p. 27
- Table 10 compares cost across frame counts and maximum tool budgets, with lower values preferred.Source: Table 10 · p. 25
- Scene, Hallucination, and Text categories show cascade lift above 20 percentage points, while Time Cognition reverses; overall r = 0.072, p < 0.001.Table 11 compares final-answer accuracy for IoU <0.1 versus IoU ≥0.1 on CGBench.Source: Table 11 · p. 25
- Table 12 defines ELV-Halluc’s in-video and out-video accuracies and their difference; lower SAH Ratio indicates better robustness to semantic aggregation hallucinations.Source: Table 12 · p. 26