GroundLie360: Misinformation Benchmark

• GroundLie360 is a real-world dataset benchmarking multimodal misinformation by integrating video titles, ASR transcripts, and raw video with detailed veracity and token-level annotations.
• Its hierarchical annotation scheme labels veracity, misinformation types, and specific localization spans (text, temporal, and spatial), enabling explainable verification and precise grounding.
• The benchmark fosters research into end-to-end multimodal models by highlighting challenges like type confusion, propagation of early errors, and the need for robust cross-modal reasoning.

GroundLie360 most prominently denotes a real-world dataset and benchmark for Grounding Multimodal Misinformation (GroundMM), designed to verify multimodal content and localize misleading segments across text, speech, and visuals (Yang et al., 8 Sep 2025). It brings together video titles, ASR transcripts, and raw video, and provides both binary veracity labels and fine-grained annotations at token, temporal, and spatial levels. In the supplied arXiv-linked technical material, the same label also appears in derivative descriptions for a 360° terrestrial LiDAR ground-surface reconstruction pipeline and for an omnidirectional vision dataset with automatically projected object centroids (Rychkov, 2012, Stamatescu et al., 2017). This suggests that the name is not unique; however, its most explicit benchmark instantiation is the misinformation-grounding dataset introduced with the GroundMM task.

1. Dataset composition and modality structure

GroundLie360 is built from 1,466 video-related events fact-checked on Snopes.com, yielding 2,044 target videos, each under five minutes, together with 588 auxiliary clips comprising original footage, debunking videos, and Snopes evidence (Yang et al., 8 Sep 2025). Of the 2,044 target clips, 995 are labeled real and 1,049 fake, with labels drawn from Snopes’ verdicts. The dataset is described as the first large-scale, real-world video dataset designed explicitly for Grounding Multimodal Misinformation.

Its organization is explicitly multimodal. The three input modalities are text, instantiated as video titles; speech, instantiated as ASR transcripts; and visuals, instantiated as raw video. The annotation taxonomy spans six misinformation types grouped under four modality categories. This design reflects the claim that current datasets and detection methods primarily target binary classification or single-modality localization based on post-processed data, whereas GroundLie360 is intended to support interpretable verification and localization across modalities.

Modality category	Fake types
Text (Title)	False Title
Speech	False Speech
Video	Temporal Edit; CGI
Cross-modal (Text–Video)	Contradictory Content; Unsupported Content

The statistical profile reported for the dataset emphasizes co-occurrence rather than isolated manipulation. Videos per fake type range from approximately 400 for CGI to approximately 800 for unsupported content, and multi-labeling yields a total of over 3,000 fake-type assignments (Yang et al., 8 Sep 2025). Over 60% of fake samples contain two types, and approximately 7% have three or more. The grounding distributions are also structured: most videos show one to two false spans in title or transcript; non-edit manipulations usually occupy less than 50% of video length; scene-insert strategies often appear at video ends; and CGI boxes most often cover small regions below 25% of frame, though full-frame generative content also exists.

2. Annotation hierarchy and quality control

GroundLie360 uses a three-level hierarchical annotation scheme (Yang et al., 8 Sep 2025). At Level 1, each video receives a ground-truth label $y\in\{0,1\}$ for real or fake. At Level 2, fake videos receive multi-label type annotations $c_i\in\{0,1\},\,i=1\ldots6$. At Level 3, the exact locus of misinformation is grounded in a modality-specific way.

The grounding protocol is type-dependent. Textual span annotations are used for the four text-based types. Temporal Edit is localized by timestamps or scene indices. CGI is localized by time intervals together with bounding boxes. This scheme makes the benchmark simultaneously a verification problem, a subtype classification problem, and a structured localization problem.

Quality control is explicit and evidence-based. Annotation was performed by two independent annotators per video, and disagreements on labels or types were adjudicated by a senior expert (Yang et al., 8 Sep 2025). To anchor subjectivity, all decisions must cite the relevant Snopes article, and each annotator provides a concise rationale. Inter-annotator agreement on grounding, measured as IoU overlap, exceeded 0.74 for text spans, 0.84 for temporal segments, and 0.98 for spatial boxes. These values are presented as evidence of high-quality annotation despite the subjectivity inherent in misinformation assessment.

A recurrent misconception in misinformation benchmarks is that a binary label is sufficient. GroundLie360 is explicitly structured against that assumption. Its hierarchical annotation scheme encodes not only whether content is false, but also which manipulation types are present and where the misleading evidence resides. A plausible implication is that the benchmark is intended for explainable systems rather than only end-point classifiers.

3. GroundMM task formulation and evaluation protocol

GroundLie360 underpins the GroundMM task, which requires three outputs from a model: veracity prediction $y\in\{0,1\}$, fake-type multi-label classification $\mathbf{c}\in\{0,1\}^6$, and modality-specific grounding $\mathbf{G}=\{G_i\mid c_i=1\}$ (Yang et al., 8 Sep 2025). Each instance is formalized as $x=(\mathcal{V},\mathcal{T},\mathcal{R})$, where $\mathcal{V}$ is the video, $\mathcal{T}$ the title text, and $\mathcal{R}$ the ASR transcript.

The learning objective jointly minimizes classification and grounding errors. The verification and type-classification terms are binary cross-entropy losses, while the grounding term may combine span-matching and IoU penalties (Yang et al., 8 Sep 2025). For temporal or spatial grounding, the overlap metric is

$$\mathrm{IoU}(G,\hat G)=\frac{|G\cap\hat G|}{|G\cup\hat G|}.$$

Evaluation follows the heterogeneity of the outputs. Verification is measured by precision, recall, and $c_i\in\{0,1\},\,i=1\ldots6$0. Multi-label subtype classification uses macro-averaged precision, recall, and $c_i\in\{0,1\},\,i=1\ldots6$1. Textual grounding uses token-level precision, recall, and $c_i\in\{0,1\},\,i=1\ldots6$2. Temporal grounding uses frame-level precision, recall, and $c_i\in\{0,1\},\,i=1\ldots6$3. Spatial-temporal grounding uses mean temporal IoU, mean visual IoU, and vIoU@$c_i\in\{0,1\},\,i=1\ldots6$4, with examples including $c_i\in\{0,1\},\,i=1\ldots6$5 and $c_i\in\{0,1\},\,i=1\ldots6$6 (Yang et al., 8 Sep 2025).

This evaluation protocol makes a methodological distinction between detecting that misinformation exists and grounding the evidence for it. That distinction is significant because a model can achieve moderate verification performance while failing to localize misleading content in the relevant modality. The reported baseline results reinforce exactly that separation.

4. FakeMark baseline and empirical behavior

The benchmark baseline is FakeMark, described as a unified QA-driven pipeline built on a multimodal LLM/VLM backbone, using GPT-4o-mini for text and InternVL-8B for vision (Yang et al., 8 Sep 2025). Its first stage is an Analysis Module: text analysis surfaces candidate false spans for title and speech; video temporal analysis examines the video for abrupt scene changes to flag temporal edits; video spatial analysis highlights potential CGI-generated segments; and cross-modal analysis inspects title and video for contradictory or unsupported claims.

The second stage contains classification heads. A Binary Authenticity Classifier predicts $c_i\in\{0,1\},\,i=1\ldots6$7, and a Multi-label Type Classifier predicts $c_i\in\{0,1\},\,i=1\ldots6$8 (Yang et al., 8 Sep 2025). The third stage contains grounding heads: a textual localizer for token spans, a temporal localizer for misleading scene indices, and a spatial-temporal CGI localizer built from query generation, interval detection, uniform frame sampling, bounding-box prediction, and box tracking. During training, classification losses guide the verification and type objectives, while grounding heads are aligned with Level 3 annotations via IoU and span-overlap objectives.

Empirically, FakeMark exposes the difficulty of the benchmark rather than saturating it. In the supervised setting, binary detection reaches 62.90% accuracy, 67.8% precision, 52.8% recall, and 59.4% $c_i\in\{0,1\},\,i=1\ldots6$9, remaining behind task-specific baselines such as FakingRecipe at 76.1% accuracy and 74.9% $y\in\{0,1\}$0 (Yang et al., 8 Sep 2025). Subtype classification is highly uneven: false-title $y\in\{0,1\}$1 is approximately 53%, CGI approximately 33%, temporal-edit approximately 10%, and cross-modal types below 17%. Text grounding achieves overall $y\in\{0,1\}$2 of approximately 22%, with false title best at 39.6% and contradictory content at 8.9%. Video grounding is weaker still: temporal-edit $y\in\{0,1\}$3 is approximately 2.4%, while CGI spatio-temporal grounding attains approximately 17.2% $y\in\{0,1\}$4, with $y\in\{0,1\}$5 and $y\in\{0,1\}$6.

The ablations identify three main tendencies. First, textual cues are the most reliable. Second, temporal and cross-modal reasoning remain major bottlenecks. Third, errors in early classification stages propagate and degrade later grounding accuracy (Yang et al., 8 Sep 2025). These findings are consistent with the benchmark’s emphasis on fine-grained, modality-aware explanation rather than only label prediction.

5. Challenges, misconceptions, and research directions

GroundMM surfaces three core challenges: Type Confusion, Multimodal Complexity, and Annotation Subjectivity & Error Propagation (Yang et al., 8 Sep 2025). Type confusion arises because real-world posts often combine multiple manipulations, such as a miscaption atop CGI. Multimodal complexity arises because distinctions such as contradictory versus unsupported content require deep semantic alignment rather than simple feature fusion. Annotation subjectivity remains unavoidable because subtle falsehoods often require external evidence from fact-checking sources, and multi-stage pipelines amplify upstream mistakes.

A common misconception is that multimodal misinformation can be reduced to a standard multimodal classification problem. GroundLie360 directly contradicts that framing. The benchmark requires localization of false text spans, misleading temporal segments, and CGI spatial regions, and it grounds these decisions in cited Snopes evidence and annotator rationales (Yang et al., 8 Sep 2025). Another misconception is that stronger vision-language backbones alone resolve the task. The reported baseline results instead suggest that reasoning over chronology, semantic contradiction, and evidence grounding remains difficult even with multimodal LLM/VLM components.

The future directions stated for the benchmark are correspondingly structural. They include end-to-end models jointly optimizing detection and grounding to reduce error cascades; stronger cross-modal representations such as contrastive grounding and fusion transformers tailored for video, text, and speech; integration of external knowledge retrieval from fact-check sites directly into model reasoning; more robust temporal models for subtle chronology manipulations; and semi-supervised or active-learning strategies to expand domain coverage with limited manual annotation (Yang et al., 8 Sep 2025). A plausible implication is that progress on GroundLie360 depends not only on larger models but also on architectures that couple evidence retrieval, multimodal alignment, and localized explanation.

6. Other technical uses of the name in 360° research

In the supplied literature, “GroundLie360” is also attached to two unrelated technical artifacts, which makes the label intrinsically ambiguous. One is a three-stage pipeline for reconstructing a globally smooth, curvature-continuous ground surface $y\in\{0,1\}$7 from a massive, 360° terrestrial LiDAR point cloud (Rychkov, 2012). Its stages are hierarchical local LSQR plane fitting with hole detection and multi-level filling, optional Hermite RBF interpolation for undersampled regions, and Partition-of-Unity blending of local models into a $y\in\{0,1\}$8 global surface. The method projects the registered 360° point cloud onto $y\in\{0,1\}$9, fits total least-squares planes in nonempty cells, fills holes top-down from coarser levels, smooths slopes with a $\mathbf{c}\in\{0,1\}^6$0 kernel, and blends local planes using tensor-product cubic B-splines or compactly supported exponential functions. The complexity claims are explicitly local and scalable: preprocessing is $\mathbf{c}\in\{0,1\}^6$1, hierarchy construction is $\mathbf{c}\in\{0,1\}^6$2, PU blending is $\mathbf{c}\in\{0,1\}^6$3 with $\mathbf{c}\in\{0,1\}^6$4 per sample, and only the optional HRBF stage is a bottleneck.

The other is an omnidirectional vision dataset description tied to “Automatic Ground Truths: Projected Image Annotations for Omnidirectional Vision” (Stamatescu et al., 2017). In that description, the data consist of 43 spherical video sequences captured at 15 fps using a RICOH THETA m15 dual-fisheye spherical camera, synchronized with VICON motion capture in the UniSA Mechatronics Laboratory. The annotations comprise raw 3D ground truth from VICON and mapped 2D XML annotations containing centroids, placeholder bounding boxes, and visibility flags. The projection framework transforms VICON world coordinates into camera coordinates through extrinsics, selects the hemisphere by spherical angle, and projects onto the image plane with a pinhole model plus radial and tangential distortion. Calibration uses a 9×6 chessboard with 35.3 mm square side for intrinsics and a VICON calibration wand for extrinsics. The reported systematic reprojection error ranges, by session and lens, from means of 5.17 px to 8.22 px, with larger errors near the fisheye boundary.

These additional uses do not alter the GroundMM benchmark itself, but they matter terminologically. The same surface form, “GroundLie360,” is associated in the provided materials with multimodal misinformation grounding, terrestrial LiDAR ground-surface reconstruction, and omnidirectional projection-based annotation. This suggests that the name should be interpreted contextually rather than assumed to denote a single established resource across arXiv-adjacent technical writing.