Jump to: overview, details, publications
Vision-Language Models / AI Safety
VLMs, Alignment, Safety & Multimodal Knowledge Graphs
Prior work at AIISC on VLM hallucination benchmarking, multimodal safety, AI-generated content detection, and knowledge-graph-grounded vision-language reasoning. Foundation work for recent developments in vision-language model safety and grounding.
Overview
Prior work at AIISC on VLM hallucination benchmarking, multimodal safety, AI-generated content detection, and knowledge-graph-grounded vision-language reasoning. Foundation work for recent developments in vision-language model safety and grounding.
Details
This research cluster investigates the reliability, safety, and structural grounding of Vision-Language Models (VLMs). Work spans hallucination taxonomy and benchmarking across text, image, and video modalities; safety alignment for multimodal systems; detection of AI-generated visual content; and knowledge-graph-conditioned reasoning to reduce model drift and factual error. This foundational work was conducted at the AI Institute of South Carolina (AIISC) and forms the basis for our recent developments in vision-language model safety and alignment.
Project Motive
VLMs inherit the hallucination and alignment vulnerabilities of LLMs while introducing new failure modes through the visual modality — unsafe outputs triggered by images, identity drift in video generation, and AI-generated content that evades detection. This project builds the diagnostic tools, benchmarks, and grounding methods needed to make VLMs safer and more factually reliable. The insights and methods developed here continue to inform our work on advanced safety mechanisms and grounding strategies for modern multimodal systems.
Publications
-
The Visual Counter Turing Test (VCT²): A Benchmark for Evaluating AI-Generated Image Detection and the Visual AI Index (V_AI)
Introduces a 166K-image benchmark across six state-of-the-art text-to-image systems to stress-test AI-generated image detectors. Reveals that 17 leading detection models achieve only ~58% accuracy in zero-shot settings, exposing a critical generalization gap and motivating the Visual AI Index (V_AI) as a standardized detectability metric.
-
ViBe: A Text-to-Video Benchmark for Evaluating Hallucination in Large Multimodal Models
The first large-scale benchmark specifically for hallucination in text-to-video generation, covering 3,782 human-annotated videos from ten T2V models. Defines five hallucination types — Vanishing Subject, Omission Error, Numeric Variability, Subject Dysmorphia, and Visual Incongruity and establishes classifier baselines for automated detection.
-
The Troubling Emergence of Hallucination in Large Language Models — An Extensive Definition, Quantification, and Prescriptive Remediations
Offers the first fine-grained taxonomy of LLM hallucination organized by degree (mild, moderate, alarming), orientation (factual mirage vs. silver lining), and intrinsic vs. extrinsic type. Goes beyond detection to prescribe targeted remediation strategies for each hallucination category.
-
A Survey of Hallucination in Large Foundation Models
A broad survey covering hallucination phenomena across language, vision, and multimodal foundation models. Synthesizes causes, detection methods, and mitigation approaches into a unified framework applicable across modalities.
-
A Comprehensive Survey of Hallucination Mitigation Techniques in Large Language Models
Systematically reviews over 30 mitigation techniques spanning decoding strategies, retrieval augmentation, self-consistency methods, and fine-tuning approaches. Provides a practical taxonomy for practitioners selecting mitigation strategies based on deployment constraints.
-
Exploring the Relationship between LLM Hallucinations and Prompt Linguistic Nuances: Readability, Formality, and Concreteness
Empirically investigates how surface-level prompt properties — readability level, formality register, and noun concreteness — affect hallucination rates in LLMs. Finds that low-readability and highly abstract prompts consistently increase hallucination, with implications for prompt engineering and safety evaluation.
-
“Sorry, Come Again?” Prompting — Enhancing Comprehension and Diminishing Hallucination with [PAUSE]-Injected Optimal Paraphrasing
Proposes injecting deliberate [PAUSE] tokens into prompts alongside optimal paraphrasing to reduce comprehension failures and hallucination without any model retraining. Demonstrates consistent gains across multiple LLMs on factuality benchmarks, offering a lightweight inference-time intervention.
-
FACTOID: Factual Entailment for Hallucination Detection
Frames hallucination detection as a factual entailment problem, using long-context embeddings and multi-task learning to identify unsupported claims in model outputs. Achieves a 40% average accuracy improvement over leading textual entailment baselines on the FACTOID benchmark.
-
IMAGINATOR: Pre-Trained Image+Text Joint Embeddings using Word-Level Grounding of Images
Extends the distributional hypothesis to the visual modality by training joint image+text embeddings on 1M image-text pairs at the object level. IMAGINATOR captures object-object co-location, word-object co-location, and word-object correlation — enabling visual analogy reasoning that convolution-based embeddings like VGGNet cannot perform.
-
Counter Turing Test (CT²): AI-Generated Text Detection is Not as Easy as You May Think — Introducing AI Detectability Index (ADI)
Challenges the assumption that current AI-generated text detectors are reliable, systematically showing that state-of-the-art methods fail under distributional shift. Introduces the AI Detectability Index (ADI) as a standardized metric for measuring how detectable a given model’s outputs are — the textual precursor to VCT².
-
Visual Causal Question Answering with Knowledge Graph Link Prediction
Grounds visual question answering in causal knowledge graphs by treating answer inference as a link prediction problem over a structured scene representation. This approach improves explainability and reduces spurious correlations that purely neural VQA models rely on.
-
Causal Knowledge Graph for Scene Understanding in Autonomous Driving
Builds causal knowledge graphs from driving scenes to provide structured, causally-grounded representations for autonomous perception systems. Addresses reliability gaps in neural scene understanding by encoding object relationships and causal dependencies explicitly rather than learning them implicitly from data.
-
AAAI 2025 Tutorial: Hallucination in Large Multimodal Models
A full-day tutorial covering hallucination theory, detection, and mitigation across text, image, and video modalities, with practical black-box and gray-box techniques. Bridges the gap between emerging research and deployment, drawing directly from the group’s ViBe, VCT², and survey work.
What We Did New
- Introduced the first fine-grained taxonomy of hallucination across degree, orientation (factual mirage vs. silver lining), and category for both LLMs and multimodal models.
- Built ViBe, the first large-scale benchmark specifically for hallucination in text-to-video generation, annotating five distinct failure types across 3,782 videos.
- Introduced VCT², a 166K-image benchmark revealing that state-of-the-art AI-generated image detectors achieve only ~58% accuracy on outputs from contemporary models — exposing a critical generalization gap.
- Proposed IMAGINATOR, a joint image+text embedding that extends the distributional hypothesis to the visual modality for the first time at object-level grounding.
- Applied causal knowledge graphs to ground visual scene understanding in autonomous driving and VQA, bridging neuro-symbolic reasoning with VLM outputs.
Status
Active and ongoing. Current open threads include hallucination mitigation in video generation, prompt-level interventions for multimodal factuality, and knowledge-graph-conditioned grounding for VLMs.