Large Vision–Language Models (LVLMs) have achieved strong multimodal performance, yet ensuring the factual correctness of generated content remains challenging. Existing methods that provide statistical guarantees on factuality typically rely on external verifiers or generation-time confidence signals, which introduce auxiliary dependencies or often fail for confident but incorrect outputs. We argue that reliable factuality control can instead be achieved through introspective signals derived from the model itself. We introduce IntroConformal, a training-free Conformal Risk Control (CRC) framework that provides finite-sample, distribution-free factuality guarantees. We first instantiate it with layer-wise semantic stability, a conformity score derived from hidden-state representations, and then propose verification probability, a stronger score capturing the model’s self-administered judgment on claim factuality. Across multiple LVLM architectures, IntroConformal satisfies the conformal risk guarantee while substantially reducing abstention and achieving competitive or superior claim-level discrimination relative to external verifier–based baselines.

Diffusion-based Large Vision-Language Models (dLVLMs) have recently emerged as a compelling alternative to autoregressive (AR) LVLMs, offering advantages in parallel decoding, bidirectional context, and controllable generation. Despite rapid progress, their reliability properties remain largely uncharacterized. We present the first systematic reliability evaluation of hallucination and bias in dLVLMs, benchmarking six diffusion models against competitive AR baselines across four dimensions. Our key findings are: (1) dLVLMs reverse the yes-bias of AR models in binary visual queries; (2) they achieve competitive hallucination rates yet exhibit degraded linguistic quality; (3) they collapse to near-zero accuracy on underrepresented racial groups with opposite-polarity gender bias; and (4) they exhibit accuracy collapse in multiple-choice settings when the correct option is shorter than its distractors, associated with a length prior that emerges at the first denoising step. Tokens committed at late denoising steps with low confidence further correlate with hallucinated content, pointing to a mechanistic signal unique to diffusion generation. These patterns vary across model families, suggesting reliability is shaped by the generative paradigm together with training data.

Large Vision-Language Models (LVLMs) demonstrate strong performance on multiple-choice question answering (MCQA), yet their reliability under biased conditions remains underexplored. We investigate length bias, the systematic preference for longer answer options, through controlled benchmarks spanning fine-grained visual domains. LVLMs consistently favor longer options even when visually inconsistent, with accuracy dropping 2–38% when correct answers are shorter and increasing 1–70% when longer. This bias persists across model scales. Through layer-wise mechanistic analysis, we reveal that length bias emerges from late-layer suppression of visual evidence in favor of language-side length priors. Blurry-image experiments show that bias intensifies when visual evidence weakens, while abstention tests reveal that LVLMs struggle to refuse incorrect predictions. We propose and evaluate multiple inference-time mitigation strategies; while these provide partial improvements, none achieves deployable bias mitigation without relying on oracle information unavailable at inference time. These results underscore that inference-time interventions are insufficient, pointing to the need for architectural solutions and bias-aware training.

Recent progress in vision-language models has enabled processing of increasingly long video sequences, but handling extended token streams does not translate to understanding complex narrative structure in long videos. Existing long-video benchmarks focus on needle-in-a-haystack retrieval rather than evaluating how low-level actions form events, interact across time, and drive narratives, for example whether a model can connect an early job loss to a later relationship breakup despite intervening scenes or flashbacks. We introduce NEST (Narrative Event Structures in Time for Long Video Understanding), a dataset of 1,005 full-length movies (avg. 98 minutes), each annotated with about 103 multimodal narrative events grounded in visual content, dialogue, and audio. NEST links these events through relations that reflect narrative structure, including explicit temporal ordering, hierarchical composition, and long-range dependencies. We introduce baselines for event trigger detection (ETD), event localization (EL), event argument extraction (EAE), and event relation extraction (ERE). The benchmark is highly challenging for grounded event discovery, with the best performance across evaluated zero-shot models on main evaluation subsets remaining below 8% for ETD, 6% for EL, and 11% for EAE. In contrast, ERE is more tractable once events are given, reaching 34.23% F1 zero-shot and 49.20% F1 after fine-tuning.

Understanding how the same real-world event is represented across text, video, and audio remains a fundamental challenge in information extraction. While prior datasets have explored events within individual modalities or limited modality pairs, a setting that combines all three modalities remains largely unexplored. We introduce M³EC (Multimodal Multidocument Event Coreference), a new benchmark for multimodal, multidocument event extraction and coreference resolution. It is built from tightly related news clusters that combine eventful videos, audio files, and news articles. We also introduce a semi-automated annotation pipeline, AutoEventAnnotator, to generate automated annotations at scale. Additionally, we construct a human-annotated test set for robust evaluation. To our knowledge, this is the first benchmark to support cluster-level multidocument event extraction and event coreference across text, video, and audio. Evaluation with strong multimodal LLMs shows that the majority of them fail to extract valid events in video and particularly in audio. We further show that models finetuned on the annotations generated by our automated pipeline achieve consistent improvements, including more than +30 F1 on video event extraction, +15 F1 on audio event extraction, and +40 CoNLL F1 on end-to-end coreference resolution, highlighting the usefulness of M³EC for training and evaluation in multimodal event understanding.

Multimodal foundation models that integrate audio, vision, and language achieve strong performance on reasoning and generation tasks, yet their robustness to adversarial manipulation remains poorly understood. We study a realistic and underexplored threat model: untargeted, audio-only adversarial attacks on trimodal audio–video–language models. We analyze six complementary attack objectives that target different stages of multimodal processing, including audio encoder representations, cross-modal attention, hidden states, and output likelihoods. Across four state-of-the-art models and multiple benchmarks, we show that audio-only perturbations can induce severe multimodal failures, achieving up to 96% attack success rate. We further show that attacks can be successful at low perceptual distortions (LPIPS ≤ 0.08, SI-SNR ≥ 0 dB) and benefit more from extended optimization than increased data scale. We evaluate the feasibility of these attacks under physically realistic conditions by incorporating room impulse response (RIR) modeling, showing that audio-only perturbations remain effective under environmental transformations and thus highlight the practical risk of single-modality attacks in real-world multimodal systems. Transferability across models and encoders remains limited, while speech recognition systems such as Whisper primarily respond to perturbation magnitude, achieving >97% attack success under severe distortion. These results expose a previously overlooked single-modality attack surface in multimodal systems and motivate defenses that enforce cross-modal consistency. Our project website is available at https://aafiya-h.github.io/soundbreak/.

This work introduces SteerVLM, a lightweight steering module designed to guide Vision-Language Models (VLMs) towards outputs that better adhere to desired instructions. Our approach learns from the latent embeddings of paired prompts encoding target and converse behaviors to dynamically adjust activations connecting the language modality with image context. This allows for fine-grained, inference-time control over complex output semantics without modifying model weights while preserving performance on off-target tasks. Our steering module requires learning parameters equal to 0.14% of the original VLM’s size. Our steering module gains model control through dimension-wise activation modulation and adaptive steering across layers without requiring pre-extracted static vectors or manual tuning of intervention points. Furthermore, we introduce VNIA (Visual Narrative Intent Alignment), a multimodal dataset specifically created to facilitate the development and evaluation of VLM steering techniques. Our method outperforms existing intervention techniques on steering and hallucination mitigation benchmarks for VLMs and proposes a robust solution for multimodal model control through activation engineering.

Large Vision-Language Models (LVLMs) have demonstrated impressive performance on vision-language reasoning tasks. However, their potential for zero-shot fine-grained image classification, a challenging task requiring precise differentiation between visually similar categories, remains underexplored. We present a novel method that transforms zero-shot fine-grained image classification into a visual question-answering framework, leveraging LVLMs’ comprehensive understanding capabilities rather than relying on direct class name generation. We enhance model performance through a novel attention intervention technique. We also address a key limitation in existing datasets by developing more comprehensive and precise class description benchmarks. We validate the effectiveness of our method through extensive experimentation across multiple fine-grained image classification benchmarks. Our proposed method consistently outperforms the current state-of-the-art (SOTA) approach, demonstrating both the effectiveness of our method and the broader potential of LVLMs for zero-shot fine-grained classification tasks. Code and Datasets: https://github.com/Atabuzzaman/Fine-grained-classification

Large Vision-Language Models (LVLMs) have achieved strong performance on vision-language tasks, particularly Visual Question Answering (VQA). While prior work has explored unimodal biases in VQA, the problem of selection bias in Multiple-Choice Question Answering (MCQA), where models may favor specific option tokens (e.g., “A”) or positions, remains underexplored. In this paper, we investigate both the presence and nature of selection bias in LVLMs through fine-grained MCQA benchmarks spanning easy, medium, and hard difficulty levels, defined by the semantic similarity of the options. We further propose an inference-time logit-level debiasing method that estimates an ensemble bias vector from general and contextual prompts and applies confidence-adaptive corrections to the model’s output. Our method mitigates bias without retraining and is compatible with frozen LVLMs. Extensive experiments across several state-of-the-art models reveal consistent selection biases that intensify with task difficulty, and show that our mitigation approach significantly reduces bias while improving accuracy in challenging settings. This work offers new insights into the limitations of LVLMs in MCQA and presents a practical approach to improve their robustness in fine-grained visual reasoning. Datasets and code are available at: https://github.com/Atabuzzaman/Selection-Bias-of-LVLMs

Discrete diffusion models are a new class of text generators that offer advantages such as bidirectional context use, parallelizable generation, and flexible prompting compared to autoregressive models. However, a critical limitation of discrete diffusion models is their inability to perform flexible-length or flexible-position text infilling without access to ground-truth positional data. We introduce DDOT (Discrete Diffusion with Optimal Transport Position Coupling), the first discrete diffusion model to overcome this challenge. DDOT jointly denoises token values and token positions, employing a novel sample-level Optimal Transport (OT) coupling. This coupling preserves relative token ordering while dynamically adjusting the positions and length of infilled segments, a capability previously missing in text diffusion. Our method is orthogonal to existing discrete text diffusion methods and is compatible with various pretrained text denoisers. Extensive experiments on text infilling benchmarks such as One-Billion-Word and Yelp demonstrate that DDOT outperforms naive diffusion baselines. Furthermore, DDOT achieves performance on par with state-of-the-art non-autoregressive models and enables significant improvements in training efficiency and flexibility.

Cross-modal image-text retrieval is challenging because of the diverse possible associations between content from different modalities. Traditional methods learn a single-vector embedding to represent semantics of each sample, but struggle to capture nuanced and diverse relationships that can exist across modalities. Set-based approaches, which represent each sample with multiple embeddings, offer a promising alternative, as they can capture richer and more diverse relationships. In this paper, we show that, despite their promise, these set-based representations continue to face issues including sparse supervision and set collapse, which limits their effectiveness. To address these challenges, we propose Maximal Pair Assignment Similarity to optimize one-to-one matching between embedding sets which preserve semantic diversity within the set. We also introduce two loss functions to further enhance the representations: Global Discriminative Loss to enhance distinction among embeddings, and Intra-Set Divergence Loss to prevent collapse within each set. Our method achieves state-of-the-art performance on MS-COCO and Flickr30k without relying on external data.

Fact-checking claims is a highly laborious task that involves understanding how each factual assertion within the claim relates to a set of trusted source materials. Existing approaches make sample-level predictions but fail to identify the specific aspects of the claim that are troublesome and the specific evidence relied upon. In this paper, we introduce a method and new benchmark for this challenging task. Our method predicts the fine-grained logical relationship of each aspect of the claim from a set of multimodal documents, which include text, image(s), video(s), and audio(s). We also introduce a new benchmark (M3DC) of claims requiring multimodal multidocument reasoning, which we construct using a novel claim synthesis technique. Experiments show that our approach outperforms other models on this challenging task on two benchmarks while providing finer-grained predictions, explanations, and evidence.

Fact-checking real-world claims often requires reviewing multiple multimodal documents in order to assess the claim’s truthfulness, a highly laborious and time-consuming task. In this paper, we present a summarization model crafted to generate claim-specific summaries useful for fact-checking from multimodal multi-document datasets. The model takes inputs in the form of documents, images, and a claim, with the objective of assisting in fact-checking tasks. We introduce a dynamic perceiver-based model that is able to handle inputs from multiple modalities of arbitrary lengths. To train our model, we leverage a novel reinforcement learning-based entailment objective in order to generate summaries that provide evidence distinguishing between different truthfulness labels. To assess the efficacy of our approach, we conduct experiments on both an existing benchmark as well as a new dataset of multi-document claims which we contribute. Our approach outperforms the SOTA approach by 4.6% in the claim verification task on the MOCHEG dataset and demonstrates strong performance on our new Multi-News-Fact-Checking dataset.

Building cross-model intelligence that can understand charts and communicate the salient information hidden behind them is an appealing challenge in the vision and language (V+L) community. The capability to uncover the underlined table data of chart figures is a critical key to automatic chart understanding. We introduce ChartT5, a V+L model that learns how to interpret table information from chart images via cross-modal pre-training on plot table pairs. Specifically, we propose two novel pre-training objectives: Masked Header Prediction (MHP) and Masked Value Prediction (MVP) to facilitate the model with different skills to interpret the table information. We have conducted extensive experiments on chart question answering and chart summarization to verify the effectiveness of the proposed pre-training strategies. In particular, on the ChartQA benchmark, our ChartT5 outperforms the state-of-the-art non-pretraining methods by over 8% performance gains.

Given a long untrimmed video and natural language queries, video grounding (VG) aims to temporally localize the semantically-aligned video segments. Almost all existing VG work holds two simple but unrealistic assumptions: 1) All query sentences can be grounded in the corresponding video. 2) All query sentences for the same video are always at the same semantic scale. Unfortunately, both assumptions make today’s VG models fail to work in practice. For example, in real-world multimodal assets (eg, news articles), most of the sentences in the article can not be grounded in their affiliated videos, and they typically have rich hierarchical relations (ie, at different semantic scales). To this end, we propose a new challenging grounding task: Weakly-Supervised temporal Article Grounding (WSAG). Specifically, given an article and a relevant video, WSAG aims to localize all “groundable” sentences to the video, and these sentences are possibly at different semantic scales. Accordingly, we collect the first WSAG dataset to facilitate this task: YouwikiHow, which borrows the inherent multi-scale descriptions in wikiHow articles and plentiful YouTube videos. In addition, we propose a simple but effective method DualMIL for WSAG, which consists of a two-level MIL loss and a single-/cross- sentence constraint loss. These training objectives are carefully designed for these relaxed assumptions. Extensive ablations have verified the effectiveness of DualMIL.

Visual and textual modalities contribute complementary information about events described in multimedia documents. Videos contain rich dynamics and detailed unfoldings of events, while text describes more high-level and abstract concepts. However, existing event extraction methods either do not handle video or solely target video while ignoring other modalities. In contrast, we propose the first approach to jointly extract events from both video and text articles. We introduce the new task of Video MultiMedia Event Extraction and propose two novel components to build the first system towards this task. First, we propose the first self-supervised cross-modal event coreference model that can determine coreference between video events and text events without any manually annotated pairs. Second, we introduce the first cross-modal transformer architecture, which extracts structured event information from both videos and text documents. We also construct and will publicly release a new benchmark of video-article pairs, consisting of 860 video-article pairs with extensive annotations for evaluating methods on this task. Our experimental results demonstrate the effectiveness of our proposed method on our new benchmark dataset. We achieve 6.0% and 5.8% absolute F-score gain on multimodal event coreference resolution and multimedia event extraction.

To defend against machine-generated fake news, an effective mechanism is urgently needed. We contribute a novel benchmark for fake news detection at the knowledge element level, as well as a solution for this task which incorporates cross-media consistency checking to detect the fine-grained knowledge elements making news articles misinformative. Due to training data scarcity, we also formulate a novel data synthesis method by manipulating knowledge elements within the knowledge graph to generate noisy training data with specific, hard to detect, known inconsistencies. Our detection approach outperforms the state-of-the-art (up to 16.8% accuracy gain), and more critically, yields fine-grained explanations.