Multimodal Large Language Models

Hallucinations in multimodal large language models (MLLMs), where a model generates content inconsistent with the input image, pose significant risks in real-world applications. Existing hallucination benchmarks primarily test recognition, i.e., whether a model can select the correct answer among distractors. They overlook another capability that is essential for trustworthy AI: recognizing when none of the provided options is supported by the image and declining to commit to a false choice. We present HumbleBench, a hallucination benchmark that evaluates this false-option rejection behavior, a directly observable facet of epistemic humility, through multiple-choice questions that always include a "None of the above" (NOTA) option. HumbleBench is built from the Panoptic Scene Graph dataset: its object and relation annotations, together with candidate attribute cues, are used to prompt GPT-4-Turbo to generate questions, which are then manually verified, yielding 22,831 questions on object, relation, and attribute hallucinations. We evaluate 27 MLLMs, including general-purpose, reasoning-oriented, and proprietary models, and complement standard accuracy with NOTA-aware metrics, diagnostic stress tests, and a human evaluation. Even the strongest models select NOTA for only about half of the questions in which it is correct, cautious prompting shifts models toward NOTA without improving their ability to detect unsupported options, and humans reject unsupported options considerably more reliably than all evaluated models. These results show that recognition accuracy alone is insufficient for assessing hallucination robustness. Our code and dataset are publicly available at https://github.com/maifoundations/HumbleBench.

Connecting pretrained vision and language models usually involves projecting image features into the language model's input space. Inverse-LLaVA reverses this mapping within decoder attention: language states are projected to the visual feature dimension, and modality-specific maps produce residual query, key, and value updates. Fusion and low-rank adaptation (LoRA) learn jointly from 665K visual instructions, with frozen backbones and no separate alignment stage. Across nine primary benchmark evaluations, the final 7B model approaches two-stage LLaVA-1.5 on several tasks. It scores 78.45% on VQAv2 versus 79.13% for official LLaVA-LoRA, and 50.96% versus 48.56% on VizWiz; TextVQA is lower at 56.96% versus 58.47%. Controlled studies examine fusion components, insertion depth, visual features, and language-model size. Representation analysis shows that the text maps preserve much of the pairwise similarity ordering while changing its geometric spread. Additional paired supervision improves celebrity recognition, while instruction replay repairs caption-induced answer-format failures. Analytical cost expressions and fixed-work profiles separate the additional fusion computation from the omitted alignment stage. These findings establish text-to-vision attention fusion as a practical alternative for instruction-only multimodal adaptation.

Document fraud poses a significant threat to industries reliant on secure and verifiable documentation, necessitating robust detection mechanisms. This study investigates the efficacy of state-of-the-art multi-modal large language models (LLMs)-including OpenAI O1, OpenAI 4o, Gemini Flash (thinking), Deepseek Janus, Grok, Llama 3.2 and 4, Qwen 2 and 2.5 VL, Mistral Pixtral, and Claude 3.5 and 3.7 Sonnet-in detecting fraudulent documents. We benchmark these models against each other and prior work on document fraud detection techniques using a standard dataset with real transactional documents. Through prompt optimization and detailed analysis of the models' reasoning processes, we evaluate their ability to identify subtle indicators of fraud, such as tampered text, misaligned formatting, and inconsistent transactional sums. Our results reveal that top-performing multi-modal LLMs demonstrate superior zero-shot generalization, outperforming conventional methods on out-of-distribution datasets, while several vision LLMs exhibit inconsistent or subpar performance. Notably, model size and advanced reasoning capabilities show limited correlation with detection accuracy, suggesting task-specific fine-tuning is critical. This study underscores the potential of multi-modal LLMs in enhancing document fraud detection systems and provides a foundation for future research into interpretable and scalable fraud mitigation strategies.

Recent advances in multimodal large language models (MLLMs) have shown impressive reasoning capabilities. However, further enhancing existing MLLMs necessitates high-quality vision-language datasets with carefully curated task complexities, which are both costly and challenging to scale. Although recent self-improving models that iteratively refine themselves offer a feasible solution, they still suffer from two core challenges: (i) most existing methods augment visual or textual data separately, resulting in discrepancies in data complexity (e.g., over-simplified diagrams paired with redundant textual descriptions); and (ii) the evolution of data and models is also separated, leading to scenarios where models are exposed to tasks with mismatched difficulty levels. To address these issues, we propose C2-Evo, an automatic, closed-loop self-improving framework that jointly evolves both training data and model capabilities. Specifically, given a base dataset and a base model, C2-Evo enhances them by a cross-modal data evolution loop and a data-model evolution loop. The former loop expands the base dataset by generating complex multimodal problems that combine structured textual sub-problems with iteratively specified geometric diagrams, while the latter loop adaptively selects the generated problems based on the performance of the base model, to conduct supervised fine-tuning and reinforcement learning alternately. Consequently, our method continuously refines its model and training data, and consistently obtains considerable performance gains across multiple mathematical reasoning benchmarks. Our code, models, and datasets will be released.

The rapid advancement of large language models (LLMs) has intensified the need for effective mechanisms to transform continuous multimodal data into discrete representations suitable for language-based processing. Discrete tokenization, with vector quantization (VQ) as a central approach, offers both computational efficiency and compatibility with LLM architectures. Despite its growing importance, there is a lack of a comprehensive survey that systematically examines VQ techniques in the context of LLM-based systems. This work fills this gap by presenting the first structured taxonomy and analysis of discrete tokenization methods designed for LLMs. We categorize 8 representative VQ variants that span classical and modern paradigms and analyze their algorithmic principles, training dynamics, and integration challenges with LLM pipelines. Beyond algorithm-level investigation, we discuss existing research in terms of classical applications without LLMs, LLM-based single-modality systems, and LLM-based multimodal systems, highlighting how quantization strategies influence alignment, reasoning, and generation performance. In addition, we identify key challenges including codebook collapse, unstable gradient estimation, and modality-specific encoding constraints. Finally, we discuss emerging research directions such as dynamic and task-adaptive quantization, unified tokenization frameworks, and biologically inspired codebook learning. This survey bridges the gap between traditional vector quantization and modern LLM applications, serving as a foundational reference for the development of efficient and generalizable multimodal systems. A continuously updated version is available at: https://github.com/jindongli-Ai/LLM-Discrete-Tokenization-Survey.

Current CT report generation frameworks predominantly rely on global feature representations, often failing to capture region-specific details and potentially missing certain abnormalities. To overcome this limitation, we propose MedRegion-CT, a region-focused multimodal large language model framework featuring three key innovations. First, we revisit the SlowFast strategy to jointly model global and fine-grained information and adapt it to the medical domain via a Region-based SlowFast Tokenizer that extracts tokens guided by clinically meaningful regions. Second, generated pseudo-masks guide the model to attend to diagnostically important anatomical regions, facilitating a systematic understanding of the overall scan context. Third, quantitative lesion information, including size, diameter, and spatial location, is encoded as structured textual prompts, enabling context-aware and clinically informed report generation. To enable rigorous evaluation, we validate our framework on multi-institutional structured report generation benchmarks. Experimental results demonstrate that MedRegion-CT achieves state-of-the-art performance, outperforming existing approaches in both linguistic quality and clinical accuracy. All code is publicly available at: https://github.com/babbu3682/MedRegion-CT.

Ensuring the trustworthiness of Multimodal Large Language Models (MLLMs) requires rigorous evaluation of gender bias, particularly in socially sensitive domains. Existing benchmarks primarily assess bias in isolated contexts and are less suited to capturing biases that emerge through interpersonal interactions. We introduce Genres, a benchmark for evaluating gender bias in MLLMs through the lens of social relationships in narrative generation. Grounded in Fiske's relational models theory, Genres comprises 1,440 multimodal narrative elicitation pairs spanning four fundamental relationship types, diverse social scenarios, and three age groups. Its dual-character generation tasks elicit socially situated interpersonal narratives and enable fine-grained assessment across multiple dimensions of gender bias. Experiments on six recent MLLMs reveal persistent yet strongly relationship- and model-dependent gender disparities, with some patterns becoming more pronounced in interpersonal settings. We further evaluate two complementary generation-stage mitigation strategies. Our findings demonstrate the value of relationship-aware evaluation for uncovering subtle, interaction-driven gender bias in MLLMs.

In this work, we propose Dimple, the first Discrete Diffusion Multimodal Large Language Model (DMLLM). We observe that training with a purely discrete diffusion approach leads to significant training instability, suboptimal performance, and severe length bias issues. To address these challenges, we design a novel training paradigm that combines an initial autoregressive phase with a subsequent diffusion phase. This approach yields the Dimple-7B model, trained on the same dataset and using a similar training pipeline as LLaVA-NEXT. Dimple-7B ultimately surpasses LLaVA-NEXT in performance by 3.9%, demonstrating that DMLLM can achieve performance comparable to that of autoregressive models. To improve inference efficiency, we propose a decoding strategy termed confident decoding, which dynamically adjusts the number of tokens generated at each step, significantly reducing the number of generation iterations. In autoregressive models, the number of forward iterations during generation equals the response length. With confident decoding, however, the number of iterations needed by Dimple is even only $\frac{\text{response length}}{3}$. We also re-implement the prefilling technique in autoregressive models and demonstrate that it does not significantly impact performance on most benchmark evaluations, while offering a speedup of 1.5x to 7x. Additionally, we explore Dimple's capability to precisely control its response using structure priors. These priors enable structured responses in a manner distinct from instruction-based or chain-of-thought prompting, and allow fine-grained control over response format and length, which is difficult to achieve in autoregressive models. Overall, this work validates the feasibility and advantages of DMLLM and enhances its inference efficiency and controllability. Code and models are available at https://github.com/yu-rp/Dimple.

Although recent large multimodal models (LMMs) show impressive progress on vision language tasks, their alignment with human centered (HC) principles such as fairness, ethics, inclusivity, empathy, and robustness is often overlooked. Existing LMM benchmarks are largely accuracy-agnostic. We present HumaniBench, a unified framework for characterizing HC alignment across realistic, socially grounded visual contexts. It contains 32,000 expert-verified image-question pairs from real-world news imagery, each mapped to one or more HC principles through explicit metrics. Comparing 15 state of the art LMMs reveals consistent trade -offs: proprietary systems lead on ethics, reasoning, and empathy, while open-source models show superior visual grounding and resilience. All models show persistent gaps in fairness and multilingual inclusivity. Chain-of-thought prompting and test-time scaling yield 8to 12 % gains on several HC dimensions. HumaniBench enables fine-grained analysis of alignment trade-offs not captured by conventional multimodal benchmarks. https://vectorinstitute.github.io/humanibench/

Forensic deepfake analysis demands more than binary classification: investigators need region-grounded natural language explanations they can verify against the image. Multimodal large language models (MLLMs) are a natural fit, but pretrained MLLMs fail systematically, producing globally coherent text that misses the small localized cues defining manipulations. We argue this is an inductive bias problem rather than a capacity issue: the image-text contrastive objective training MLLM visual encoders optimizes for whole-image semantic summaries, not patch-level forensic detail. The same mismatch explains why prior deepfake reasoning methods target either face manipulation or fully AI-generated content, never both. We propose FORGE, which addresses the mismatch by routing a second visual stream into the language model from a Vision-Only Model (VOM) trained on dense patch prediction rather than image-text alignment. The MLLM's native encoder and the VOM operate on a shared patch grid, which lets us interleave their tokens with preserved spatial correspondence; we show this beats naive concatenation. A two-stage adapter training protocol (generic image-caption alignment, then joint task-specific optimization) prevents the localized stream from overfitting to training-domain manipulations. Across face-manipulated and fully synthetic content, FORGE produces region-referential explanations answering fine-grained attribute queries ("Does the eyes/nose/mouth look real or fake?") and substantially outperforms in-domain baselines on cross-domain evaluations; region-specific evaluation and human studies confirm explanation faithfulness.

Unimodal vision models are known to rely on spurious correlations, but it remains unclear to what extent Multimodal Large Language Models (MLLMs) exhibit similar biases despite language supervision. In this paper, we investigate spurious bias in MLLMs and introduce SpurLens, a pipeline that leverages GPT-4 and open-set object detectors to automatically identify spurious visual cues without human supervision. Our findings reveal that spurious correlations cause two major failure modes in MLLMs: (1) over-reliance on spurious cues for object recognition, where removing these cues reduces accuracy, and (2) object hallucination, where spurious cues amplify the hallucination by over 10x. We validate our findings in various MLLMs and datasets. Beyond diagnosing these failures, we explore potential mitigation strategies, such as prompt ensembling and reasoning-based prompting, and conduct ablation studies to examine the root causes of spurious bias in MLLMs. By exposing the persistence of spurious correlations, our study calls for more rigorous evaluation methods and mitigation strategies to enhance the reliability of MLLMs.

To overcome the black-box nature of predictive AI and the hallucination risks of generative models, retrieval-based models offer an interpretable, evidence-based paradigm for pathology clinical workflow. However, real-world clinical queries are inherently interleaved (e.g., pathology images and text). Current dual-encoders suffer from an Architectural Mismatch, lacking the mechanism to fuse such composed queries. To address this, we formalize the task of Pathology Composed Retrieval (PCR). While Multimodal Large Language Models (MLLMs) offer deep-fusion capabilities, directly applying them exposes a Task Mismatch and a Domain Mismatch. To resolve these challenges, we propose HOMIE, a model-agnostic adaptation framework that transforms any generative MLLM into a specialized pathology retrieval expert. Evaluated on our newly introduced PCR Benchmark, a lightweight 2B-parameter HOMIE variant substantially outperforms existing paradigms, surpassing specialized 7B pathology MLLMs and dual-encoders by large margins on composed retrieval, while maintaining strong performance on traditional simple retrieval. The project page is available at https://qfchou.github.io/HOMIE_page/.

Large language models (LLMs) are widely applied in various fields of society due to their powerful reasoning, understanding, and generation capabilities. However, the security issues associated with these models are becoming increasingly severe. Jailbreaking attacks, as an important method for detecting vulnerabilities in LLMs, have been explored by researchers who attempt to induce these models to generate harmful content through various attack methods. Nevertheless, existing jailbreaking methods face numerous limitations, such as excessive query counts, limited coverage of jailbreak modalities, low attack success rates, and simplistic evaluation methods. To overcome these constraints, this paper proposes a multimodal jailbreaking method: JMLLM. This method integrates multiple strategies to perform comprehensive jailbreak attacks across text, visual, and auditory modalities. Additionally, we contribute a new and comprehensive dataset for multimodal jailbreaking research: TriJail, which includes jailbreak prompts for all three modalities. Experiments on the TriJail dataset and the benchmark dataset AdvBench, conducted on 13 popular LLMs, demonstrate advanced attack success rates and significant reduction in time overhead.

Video platforms, used by billions globally, are criticized for exposing users to harmful content. Yet, detecting online harm remains challenging due to inconsistent approaches to what constitutes harm and limitations of human annotation. We address both challenges. Theoretically, we integrate and expand existing taxonomies and platform guidelines to multimodal contexts and develop a comprehensive taxonomy for online harm. We categorize online harm on video platforms into six content-based categories: Information, Hate and harassment, Addictive, Clickbait, Sexual, and Physical harms. Methodologically, we propose and systematically test multimodal large language models (MLLMs) as alternative annotators of harmful videos. We curate a dataset of 19,422 potentially harmful YouTube videos and annotate them using GPT-4-Turbo, non-trained crowdworkers, and trained annotators. We discuss the potential applications of MLLMs in platform content moderation and social science research, particularly in contexts where they perform comparably to human annotators in binary and multi-label classification tasks.

Recent advances in Multimodal Large Language Models (MLLMs) have shown promise for remote sensing tasks such as visual question answering and scene understanding. However, existing models remain limited to basic instruction-following and struggle with real-world scenarios that require multi-source data integration, fine-grained spatial reasoning, and domain expertise. To address this gap, we propose RS-Agent, a domain-adapted intelligent agent that connects user intent with professional remote sensing workflows through structured task planning and tool orchestration. RS-Agent consists of four components aligned with typical remote sensing workflows: a Central Controller for intent understanding and process planning, a dynamic toolkit for tool execution, a Solution Space for task-specific expert guidance, and a Knowledge Space for domain knowledge support. We further introduce Task-Aware Retrieval, which improves planning by identifying task types and retrieving expert-defined solutions, and DualRAG, a weighted dual-path retrieval-augmented generation method that enhances the relevance and completeness of retrieved knowledge. RS-Agent natively supports multiple imaging modalities, including optical and SAR imagery, and can automatically organize dedicated SAR processing tools into executable workflows. Experiments on 9 datasets and 18 remote sensing tasks show that RS-Agent significantly outperforms state-of-the-art MLLMs, achieving over 95% task planning accuracy and strong results in scene classification, object counting, and remote sensing visual question answering. These results demonstrate the value of combining LLM reasoning with remote sensing expertise for intelligent geospatial analysis.

Radiology report generation (RRG) has attracted significant attention due to its potential to reduce the workload of radiologists. The performance of current RRG approaches remains unsatisfactory against clinical standards. This paper introduces a novel RRG method, MLLM-RRG, that integrates multimodal large language models (MLLMs) with various types of clinical knowledge to generate accurate and comprehensive chest X-ray reports. Our method first designs a referring anatomical feature extractor that leverages anatomical knowledge to analyze different regions of the chest X-ray image and extract visual features without explicitly detecting regions. Next, based on the MLLM's decoder, we develop a multimodal report generator that leverages multimodal prompts constructed from dedicated visual features and textual instructions to produce the radiology report in an auto-regressive way. Finally, we introduce a disease-oriented clinical classification and alignment scheme in a multi-task learning manner to leverage disease knowledge to better preserve the clinical relevance among the generated reports. Once the model is trained, we also introduce a novel clinical quality reinforcement learning strategy to enhance the MLLM with report knowledge, further refining the tones of the generated reports towards radiologists. Extensive experiments on the MIMIC-CXR and IU X-Ray datasets demonstrate the superiority of our method over the state of the art. Our codes will be available at https://github.com/viscom-tongji/MLLM-RRG.