MiniVer-V: Identifying Minimal Sufficient Evidence for Short Video Verification

Leran Chen; Lingnan Kong; Zile Cai

cs.CVOct 8, 2026

MiniVer-V: Identifying Minimal Sufficient Evidence for Short Video Verification

Authors: Leran Chen, Lingnan Kong, Zile Cai

Organizations: The University of Sydney

Abstract

A core challenge in short-video fact-checking is identifying which evidence is sufficient to support a verification conclusion. Existing approaches either give the verifier all available evidence, introducing noise, or select evidence by topical relevance, which conflates relatedness with sufficiency. We identify evidential sufficiency as the selection criterion: whether a subset of evidence is adequate to support a confident verdict without redundancy. We introduce MiniVer-V, a benchmark of 195 short videos with three-way verdict annotations (supported, refuted, insufficient) and 5,510 multimodal evidence units spanning visual keyframes, speech transcripts, and web-retrieved external sources. We propose a two-layer verification framework that separates claim-video consistency, assessed from internal evidence, from factual verdict determination, which additionally requires external corroboration. On top of it, a sufficiency-driven greedy search assembles evidence until a sufficiency threshold is met and outputs insufficient when the candidate pool is exhausted, rather than forcing a verdict. With Claude Sonnet 4, the method reaches a Macro-F1 of 0.510 using 4.5 evidence units on average (16% of the full evidence set), statistically indistinguishable from the full-evidence baseline (0.518 with 27.7 units), while significantly improving recognition of insufficient cases over the same search without abstention. The efficiency result replicates with GPT-5.5 and holds only partially with an open-weight Qwen2.5-72B verifier. Ablations show that external evidence is indispensable for factual determination, while internal video evidence grounds the verdict in claim-video consistency. These findings suggest that evidence-efficient verification is achievable, and that explicit abstention is needed when evidence is genuinely inadequate.

Figures & tables

Verdict	Count	Proportion
Supported	97	49.7%
Refuted	39	20.0%
Insufficient	59	30.3%
Total	195	100%

Table 1: Verdict distribution in MiniVer-V.

Modality	Code	Total Units	Avg / Sample	Proportion
Visual keyframes	V	1,207	6.2	21.9%
ASR transcripts	A	2,847	14.6	51.7%
External search	E	1,456	7.5	26.4%
Total		5,510	28.3	100%

Table 2: Evidence unit composition in MiniVer-V (metadata excluded).

Benchmark	Media	Samples	Lang	Labels	Evidence (granularity)	Selection focus
FEVER	Text	$\sim$ 185K	EN	3-way	Text sentences	No
AVeriTeC	Text	$\sim$ 4.6K	EN	4-way	QA pairs + URLs	No
MOCHEG	Multimodal web	$\sim$ 15.6K	EN	3-way	Paragraphs + images	No
FakeSV	Short video	$\sim$ 3.7K ∗	ZH	Binary	V + A + M (per sample)	No
MiniVer-V	Short video	195	ZH/EN	3-way	V + A + E (per unit)	Yes

Table 3: Comparison of MiniVer-V with representative fact-checking benchmarks. Selection focus: whether evidence selection is the primary object of evaluation. ∗ FakeSV: fake and real videos used for binary detection (5,538 videos including debunking videos).

Figure 1: Overview of the MiniVer-V verification framework. Given a claim and a multimodal evidence pool (visual keyframes, speech transcripts, and web-retrieved external sources), five evidence selection strategies of increasing sophistication are evaluated. B5, our proposed method, performs sufficiency-driven greedy search with explicit abstention, feeding the selected compact evidence subset into a two-layer verifier that disentangles claim–video consistency (Layer 1) from factual verdict determination (Layer 2).

Method	Macro-F1	Accuracy	Avg Units	Abst. Prec.	Abst. Recall	OverConf.
B1 Full	0.518	56.9%	27.7	0.500	0.254	0.746
B2 Top-3	0.466	49.7%	3.0	0.378	0.525	0.475
B2 Top-5	0.459	50.3%	5.0	0.370	0.339	0.661
B2 Top-7	0.460	51.3%	7.0	0.350	0.237	0.763
B3 CoT	0.453	57.4%	9.4	1.000	0.051	0.949
B4 Greedy	0.469	52.3%	4.1	0.333	0.203	0.797

Table 4: Main results (195 samples, Claude Sonnet 4). The lower block applies confidence-threshold abstention to three base strategies: any committed verdict with confidence below

\tau=0.8

(the threshold B4 and B5 use for the sufficiency check) is converted to insufficient . Bold marks the highest Macro-F1 among efficient methods (

\leq

10 units).

Component ablations on B4 ( $\Delta$ F1 relative to B4)
Setting	Macro-F1	Acc.	Units	$\Delta$ F1 [95% CI]
B5 (proposed)	0.510	54.4%	4.5	—
w/o Abstention (= B4)	0.469	52.3%	4.1	$-$ 0.041 [ $-$ 0.10, +0.02]
w/o External	0.273	32.3%	7.3	$-$ 0.196 [ $-$ 0.28, $-$ 0.11]
w/o V+A (E-only)	0.484	53.3%	3.5	+0.016 [ $-$ 0.05, +0.08]
w/o ASR	0.465	51.3%	4.6	$-$ 0.004 [ $-$ 0.07, +0.06]

Table 5: Ablation study results. The abstention gate is ablated by comparing B5 with B4; all other components are ablated on B4 (the selection pipeline without the abstention gate), and their

\Delta

F1 is relative to B4. Brackets give 95% paired-bootstrap confidence intervals.

Figure 2: Macro-F1 vs. average evidence units per sample across all strategies. B5 (star) achieves near-full-evidence performance (98.5% of B1) using only 16% of the evidence. Hollow markers apply confidence-threshold abstention (

\tau=0.8

) to B1, B2 Top-5, and B4. The dashed line marks B1’s full-evidence Macro-F1.

Metric	Value
Avg Jaccard similarity	0.882
Avg Precision (B5 $\cap$ human / B5)	0.899
Avg Recall (B5 $\cap$ human / human)	0.970
Exact match rate	71.8%
Human $\subseteq$ B5 rate	92.8%
Avg human set size	3.74

Table 6: Comparison of B5’s evidence selection against human annotations (195 samples).

Verifier	B1 F1	B5 F1	B5/B1	Units	Ins-F1 B4 $\to$ B5
Claude Sonnet 4	0.518	0.510	98.5%	4.5 (16%)	0.253 $\to$ 0.383
GPT-5.5	0.586	0.559	95.3%	4.9 (18%)	0.407 $\to$ 0.508
Qwen2.5-72B	0.580	0.489	84.3%	2.4 (9%)	0.413 $\to$ 0.423

Table 7: Cross-model results on the identical frozen evidence pool. B5/B1: share of B1’s Macro-F1 recovered by B5. Units: B5’s average evidence units (share of B1’s). Ins-F1: insufficient-class F1.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Set Size	Count	Proportion	Cumulative
1	32	16.4%	16.4%
2	47	24.1%	40.5%
3	31	15.9%	56.4%
4	31	15.9%	72.3%
5	18	9.2%	81.5%
6	9	4.6%	86.2%

Appendix

Table 8: Distribution of human-annotated minimal sufficient evidence set sizes.

	Supported			Refuted			Insufficient
Method	P	R	F1	P	R	F1	P	R	F1
B1 Full	0.632	0.742	0.682	0.471	0.615	0.533	0.500	0.254	0.337
B2 Top-3	0.679	0.546	0.606	0.371	0.333	0.351	0.378	0.525	0.440
B2 Top-5	0.606	0.649	0.627	0.405	0.385	0.395	0.370	0.339	0.354
B2 Top-7	0.615	0.691	0.650	0.413	0.487	0.447	0.350	0.237	0.283
B3 CoT	0.596	0.866	0.706	0.490	0.641	0.556	1.000	0.051	0.097

Appendix

Table 9: Per-class precision, recall, and F1 across all strategies (195 samples). Bold marks the best value in each column.

Backbone	Method	Macro-F1	Accuracy	Units	Ins-F1
Claude Sonnet 4	B1 Full	0.518	0.569	27.7	0.337
	B2 Top-5	0.459	0.503	5.0	0.354
	B4 Greedy	0.469	0.523	4.1	0.253
	B5 Abstention	0.510	0.544	4.5	0.383
GPT-5.5 (closed)	B1 Full	0.586	0.651	27.7	0.464
	B2 Top-5	0.551	0.605	5.0	0.404

Appendix

Table 10: Cross-model generalization: full results for three verifier backbones on the identical frozen evidence pool. Claude Sonnet 4 (

n{=}195

) and GPT-5.5 (

n{=}195

) cover all samples; for Qwen2.5-72B, a few samples whose claims concern disaster and violent-incident reporting were refused by the DashScope provider’s content-safety filter and excluded (

n{=}193

for B1/B2;

n{=}192

for B4/B5). These are provider-side content refusals unrelated to the selection method. “Units” = average selected evidence units; “Ins-F1” = insufficient-class F1. Bold marks the best value per column within each backbone. The efficiency effect (B5 recovers most of B1’s Macro-F1 with a fraction of the evidence) holds on GPT-5.5 and only partially on Qwen2.5-72B; the abstention gain (B5

>

B4 on Ins-F1) is clear on Claude Sonnet 4 and GPT-5.5 but small on Qwen2.5-72B.

Claude Sonnet 4 ( $n{=}195$ )
Method	Macro-F1	Units	$\Delta$ Macro-F1 [95% CI]	$\Delta$ Ins-F1 [95% CI]	CV- $\tau$ Macro-F1
B1 Full	0.518	27.7	$-$ 0.008 [ $-$ 0.08, +0.06]	+0.046 [ $-$ 0.08, +0.18]	—
B2 Top-3	0.466	3.0	+0.044 [ $-$ 0.03, +0.12]	$-$ 0.056 [ $-$ 0.18, +0.07]	—
B2 Top-5	0.459	5.0	+0.051 [ $-$ 0.02, +0.13]	+0.029 [ $-$ 0.11, +0.17]	—
B2 Top-7	0.460	7.0	+0.050 [ $-$ 0.02, +0.12]	+0.101 [ $-$ 0.02, +0.22]	—
B3 CoT	0.453	9.4	+0.057 [ $-$ 0.02, +0.13]	+0.287 [+0.16, +0.41] ∗	—

Appendix

Table 11: Paired-bootstrap comparisons of B5 against every other method, on all three verifier backbones.

\Delta

= B5 minus method (positive favors B5), with 95% confidence intervals; ∗ marks

p<0.05

. “+

\tau

” rows apply confidence-threshold abstention with

\tau=0.8

; the last column gives the Macro-F1 of the same base method with a cross-validated

\tau

(mean

\pm

s.d. over 20 random 5-fold splits). For Qwen2.5-72B, the comparison uses the 192 samples on which all four strategies returned a verdict, so B1 and Top-5 differ slightly from Table 10 (

n{=}193

True $\backslash$ Pred	Supported	Refuted	Insufficient
Supported (97)	66	8	23
Refuted (39)	7	17	15
Insufficient (59)	30	6	23

Appendix

Table 12: Confusion matrix for B5 (Greedy + Abstention).

True $\backslash$ Pred	Supported	Refuted	Insufficient
Supported (97)	72	16	9
Refuted (39)	9	24	6
Insufficient (59)	33	11	15

Appendix

Table 13: Confusion matrix for B1 (Full Evidence).

Method	Consistent	Inconsistent	Unclear
B1 Full	168 (86.2%)	18 (9.2%)	9 (4.6%)
B2 Top-5	117 (60.0%)	9 (4.6%)	69 (35.4%)
B3 CoT	141 (72.3%)	54 (27.7%)	0 (0.0%)
B4 Greedy	83 (42.6%)	13 (6.7%)	99 (50.8%) †
B5 Abstention	79 (40.5%)	14 (7.2%)	102 (52.3%) †

Appendix

Table 14: Distribution of Layer 1 (consistency) assessments across strategies.

Setting	F1-Sup	F1-Ref	F1-Ins	Macro-F1
B5 (proposed)	0.660	0.486	0.383	0.510
B4 (= B5 w/o Abstention)	0.654	0.500	0.253	0.469
B4 w/o External	0.096	0.270	0.453	0.273
B4 w/o V+A (E-only)	0.660	0.453	0.340	0.484
B4 w/o ASR	0.638	0.494	0.263	0.465
B4 w/o Pruning	0.660	0.506	0.265	0.477

Appendix

Table 15: Per-class F1 for ablation variants. Rows below the rule are ablations of B4 (the selection pipeline without the abstention gate) and should be compared with the B4 row.

Topic	Count	Proportion
International Politics & Military	45	23.1%
Domestic News & Society	30	15.4%
Health & Medicine	29	14.9%
Science & Technology	24	12.3%
Natural Disaster & Accident	22	11.3%
Food Safety & Consumer	14	7.2%

Appendix

Table 16: Topic distribution of MiniVer-V samples.

Platform	Count	Proportion
Douyin	112	57.4%
Xiaohongshu	64	32.8%
X (Twitter)	8	4.1%
Bilibili	3	1.5%
Weibo	3	1.5%
News outlets (CNN, BBC, Xinhua, Southern Daily)	5	2.6%

Appendix

Table 17: Source platform distribution of MiniVer-V samples.

Modality	Total	Mean/Sample	Proportion	Source
Visual (V)	1,207	6.2	21.9%	Claude Sonnet 4
ASR (A)	2,847	14.6	51.7%	Whisper (base)
External (E)	1,456	7.5	26.4%	Web Search
Total	5,510	28.3	100%

Appendix

Table 18: Evidence unit statistics by modality (metadata excluded).

Modality	Units Selected	Proportion	Contrast with Pool
Visual (V)	116	15.9%	21.9% in pool
ASR (A)	237	32.5%	51.7% in pool
External (E)	376	51.6%	26.4% in pool
Total	729	100%

Appendix

Table 19: Modality composition in human-annotated minimal sufficient sets.

Method	Avg (All)	Avg (Correct)	Avg (Incorrect)
B1 Full	27.7	27.7	27.7
B2 Top-3	3.0	3.0	3.0
B2 Top-5	5.0	5.0	5.0
B2 Top-7	7.0	7.0	7.0
B3 CoT	9.4	9.2	9.6
B4 Greedy	4.1	3.5	4.8

Appendix

Table 20: Evidence efficiency: average units used for correct vs. incorrect predictions.

Explore similar work

Aug 9, 2026cs.CV

REVEAL: A Rubric-Guided Agent for Explicit Evidence Sufficiency Verificationin Long-Video Question Answering

Recently, retrieval-augmented and memory-augmented methods have emerged as two promising paradigms for long-video question answering. However, existing methods typically rely on rigid, fixed-length temporal chunking (e.g., 10s) and static offline memory banks, which not only fragment coherent continuous events but also fail to adapt during real-time reasoning. Moreover, whether using multi-scale summaries or multimodal knowledge graphs, current approaches prioritize retrieval relevance while overlooking evidence sufficiency, often stopping to answer once only semantically relevant clues are retrieved, even when key temporal, causal, or fine-grained action evidence is still missing. To tackle these challenges, we propose REVEAL, a rubric-guided agent framework. As a foundation, we introduce an adaptive visual-similarity-based preprocessing pipeline that groups visually coherent adjacent frames into natural event units to construct an offline-online video memory---capturing global video context offline while dynamically maintaining question-conditioned memory online. Built upon this structured memory, REVEAL uses an automatically constructed rubric library to explicitly verify whether retrieved evidence satisfies sufficiency criteria, pinpoints missing clues upon verification failure, and directs targeted re-retrieval for complementary information. Without any extra training, REVEAL consistently outperforms both closed-source and open-source state-of-the-art methods across extensive experiments. These results show that explicitly verifying evidence sufficiency, rather than stopping at semantic relevance, retrieves the decisive clues that prior methods miss and yields more reliable long-video reasoning.

Caijun Yan, Yang Zhou, Meixing Shi +4

1Zhejiang University · 2Shanghai AI Laboratory

Video QA Long-Video QA

Ask PDF

Apr 2, 2026cs.CV

VideoZeroBench: Probing the Limits of Video MLLMs with Spatio-Temporal Evidence Verification

Video multimodal large language models achieve strong results on existing benchmarks, but answer accuracy alone does not establish whether they can locate the evidence needed to answer a question. We introduce VideoZeroBench, a challenging long-video benchmark with manually annotated question-answer pairs spanning 13 video domains. Questions target fine-grained cues, fleeting events, and evidence distributed across multiple segments. Temporal intervals and key-frame boxes are annotated where applicable. All questions undergo two rounds of cross-verification for answer validity and evidence quality. Our five-level diagnostic protocol compares answering with and without evidence hints, then combines answer correctness with independently evaluated temporal and spatial grounding. Across 19 evaluated models, the best standard QA accuracy is 24.8% (Level-3), achieved by Gemini-3.7-Flash. No model exceeds 1.8% when correct answers and accurate spatio-temporal localization are jointly required (Level-5). Analyses of atomic abilities, evidence spans, input modalities, and thinking-with-videos inference further characterize where the evaluated systems struggle. These findings motivate more precise evidence search and localization for long-video question answering. Our code and data are publicly released.

Jiahao Meng, Yue Tan, Qi Xu +9

PKU · WHU · CASIA +2

Video-Language Model Evaluation Long-Video QA Temporal Video Grounding

Ask PDF

Jul 20, 2026cs.CV

Sparse Evidence Can Suffice: Agentic Evidence Seeking for Multimodal Video Misinformation Detection

Multimodal video misinformation detection is commonly formulated as a holistic video-understanding task, where the entire video and its associated content are processed and judged in a single pass. However, real-world misinformation often exhibits a sparse and compositional evidence structure: a reliable decision may depend on only a few coupled clues, while most video content contributes limited additional information. Exhaustive multimodal reasoning may therefore introduce substantial redundancy and obscure decisive evidence. This motivates decoupling evidence acquisition from verification: first identifying sparse, decision-relevant clues and then judging veracity based on the acquired evidence. Accordingly, we propose SIEVE, a framework for Sparse Interactive Evidence Verification via Extraction in multimodal video misinformation detection. An evidence-seeking agent actively explores the available multimodal evidence and constructs a compact evidence package, which is then used by a verifier to determine veracity. The agent is trained with supervised evidence-seeking trajectories and an evidence-aware reinforcement learning objective that promotes informative evidence acquisition while discouraging unnecessary or invalid interactions. Experiments on multiple video misinformation benchmarks show that SIEVE consistently outperforms the evaluated baselines and supports reliable verification using compact evidence packages. Moreover, the resulting acquisition process provides an explicit and inspectable evidence trail, improving the transparency and groundedness of multimodal misinformation detection.

Haochen Zhao, Yongxiu Xu, Xinkui Lin +6

Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China · School of Cyber Security, University of Chinese Academy of Sciences, Beijing, China · Baidu, Beijing, China

Automated Fact-Checking Multimodal Agents

Ask PDF