Publications by Lorenzo Baraldi

Video-language models (VLMs) achieve strong multimodal understanding but remain prone to hallucinations, especially when reasoning about actions and temporal order. Existing mitigation strategies, such as textual filtering or random video perturbations, often fail to address the root cause: over-reliance on language priors rather than fine-grained visual dynamics. We propose a scalable framework for counterfactual video generation that synthesizes videos differing only in actions or temporal structure while preserving scene context. Our pipeline combines multimodal LLMs for action proposal and editing guidance with diffusion-based image and video models to generate semantic hard negatives at scale. Using this framework, we build CounterVid, a synthetic dataset of ~26k preference pairs targeting action recognition and temporal reasoning. We further introduce MixDPO, a unified Direct Preference Optimization approach that jointly leverages textual and visual preferences. Fine-tuning diverse VLMs with MixDPO yields consistent improvements, notably in temporal ordering, and transfers effectively to standard video hallucination benchmarks.

Large Language Models (LLMs) have revolutionized language modeling through autoregressive generation, enabling strong performance across a wide range of tasks. Recently, Diffusion Language Models (DLMs) have emerged as an alternative paradigm that generates text through iterative denoising rather than next-token prediction, allowing parallel refinement of entire sequences. While numerous diffusion-based architectures have been proposed, differences in evaluation protocols, datasets, inference budgets, and generation hyperparameters make it difficult to compare their capabilities and understand the trade-offs they offer. In this work, we present a systematic experimental analysis of modern DLMs. Specifically, we evaluate eight state-of-the-art DLMs across eight benchmarks spanning reasoning, coding, translation, knowledge, and structured problem solving, while explicitly considering both generation quality and computational efficiency. Beyond downstream evaluation, we analyze the impact of key inference-time factors, including denoising steps, context length, block size, and parallel unmasking strategies, and complement large-scale experiments with controlled comparisons of smaller models trained under identical conditions. Our analysis highlights the strengths and limitations of diffusion-based language modeling across different tasks, architectures, and inference budgets. We show that the behavior of DLMs is strongly influenced by generation-time design choices, leading to distinct trade-offs between performance and computational efficiency. Overall, our study provides practical insights into the capabilities and deployment characteristics of contemporary DLMs.

Answer grounding in visual question answering over documents remains an open challenge: most existing benchmarks lack grounding annotations or provide annotations of limited quality, and the development of answer grounding datasets is still bounded to expensive manual effort. We introduce DocAttriBench (DAB), an improved large-scale benchmark for fine-grained, element-level source attribution in Document VQA, grounding each answer to specific layout elements such as text blocks, tables, and images. To build DAB, we propose a Mask-based Perplexity-Derived Attribution method (MAPPET) that leverages document layout and language modeling to assign each answer to its most informative document element. MAPPET measures the increase in perplexity when masking candidate elements, attributing the answer to the element that contributes most to model confidence. Applying MAPPET to multiple existing Document VQA datasets yields DAB, which contains 237k documents and 296k question-answer pairs, where each answer is grounded to the corresponding layout elements. We benchmark a range of Multimodal Large Language Models (MLLMs) with grounding capabilities on DAB, evaluating answer accuracy, attribution accuracy, and overall answer quality. Our results show that while larger models tend to achieve higher answer accuracy, even the most capable models often fail to correctly localize the supporting elements. DAB thus provides a scalable benchmark for developing grounded, verifiable, and trustworthy Document VQA models.

Vision-and-Language Navigation (VLN) requires embodied agents to follow natural language instructions while grounding their decisions in complex visual environments. Despite recent progress, most VLN agents remain difficult to interpret, providing limited insight into why a navigation action is selected. In this work, we introduce ExplaiNavility, an interpretable VLN framework that augments navigation with explicit reasoning supervision. Our model combines a visual scene encoder with a multimodal LLM, using structured prompts to jointly condition on the instruction, the navigation history, and the candidate viewpoints available at each step. Beyond predicting the next navigable location, the model is trained to produce human-readable explanations of its decisions and descriptive low-level action labels, such as turning left, turning right, moving forward, or stopping. To supervise these outputs, we augment R2R trajectories with speaker-style action explanations generated from consecutive observations, together with action labels obtained from the simulator. By fine-tuning the embodied multimodal agent with these auxiliary reasoning objectives, ExplaiNavility encourages more transparent decision-making while preserving strong navigation ability. Experiments on the R2R benchmark show that our approach improves over prior explainable VLN methods on unseen environments, while producing interpretable rationales that make the agent's behavior easier to inspect. These results suggest that explicit reasoning supervision can benefit both navigation performance and interpretability in embodied instruction-following agents.

Diffusion Transformers (DiTs) and related flow-based architectures are now among the strongest text-to-image generators, yet the internal mechanisms through which prompts shape image semantics remain poorly understood. In this work, we study massive activations: a small subset of hidden-state channels whose responses are consistently much larger than the rest. We show that, despite their sparsity, these few channels effectively draw the whole picture, in three complementary senses. First, they are functionally critical: a controlled disruption probe that zeroes the massive channels causes a sharp collapse in generation quality, while disrupting an equally-sized set of low-statistic channels has marginal effect. Second, they are spatially organized: restricting image-stream tokens to massive channels and clustering them yields coherent partitions that closely align with the main subject and salient regions, exposing a structured spatial code hidden inside an apparently outlier-like subspace. Third, they are transferable: transporting massive activations from one prompt-conditioned trajectory into another, shifts the final image toward the source prompt while preserving substantial content from the target, producing localized semantic interpolation rather than unstructured pixel blending. We exploit this property in two use cases: text-conditioned and image-conditioned semantic transport, where massive activations transport enables prompt interpolation and subject-driven generation without any additional training. Together, these results recast massive activations not as activation anomalies, but as a sparse prompt-conditioned carrier subspace that organizes and controls semantic information in modern DiT models.

Multimodal Large Language Models (MLLMs) have achieved impressive performance across a variety of vision–language tasks. However, their internal working mechanisms remain largely underexplored. In this work, we introduce FG-TRACER, a framework designed to analyze the information flow between visual and textual modalities in MLLMs in free-form generation. Notably, our numerically stabilized computational method enables the first systematic analysis of multimodal information flow in underexplored domains such as image captioning and chain-of-thought (CoT) reasoning. We apply FG-TRACER to three state-of-the-art MLLMs—LLaVA 1.5, LLaMA 3.2-Vision, and Qwen 2.5-VL—across three vision–language benchmarks—TextVQA, COCO 2014, and ChartQA—and we conduct a word-level analysis of multimodal integration. Our findings uncover distinct patterns of multimodal fusion across models and tasks, demonstrating that fusion dynamics are both model- and task-dependent. Overall, FG-TRACER offers a robust methodology for probing the internal mechanisms of MLLMs in free-form settings, providing new insights into their multimodal reasoning strategies. Our source code is publicly available at https://github.com/AImageLab-zip/FG-TRACER

Despite recent advances, single-image super-resolution (SR) remains challenging, especially in real-world scenarios with complex degradations. Diffusion-based SR methods, particularly those built on Stable Diffusion, leverage strong generative priors but commonly rely on text conditioning derived from semantic captioning. Such textual descriptions provide only high-level semantics and lack the spatially aligned visual information required for faithful restoration, leading to a representation gap between abstract semantics and spatially aligned visual details. To address this limitation, we propose GramSR, a one-step diffusion-based SR framework that replaces text conditioning with dense visual features extracted from the low-resolution input using a pre-trained DINOv3 encoder. GramSR adopts a three-stage LoRA architecture, where pixel-level, semantic-level, and texture-level LoRA modules are trained sequentially. The pixel-level module focuses on degradation removal using L2 loss, the semantic-level module enhances perceptual details via LPIPS and CSD losses, and the texture-level module enforces feature correlation consistency through a Gram matrix loss computed from DINOv3 features. At inference, independent guidance scales enable flexible control over degradation removal, semantic enhancement, and texture preservation. Extensive experiments on standard SR benchmarks demonstrate that GramSR consistently outperforms existing one-step diffusion-based methods, achieving superior structural fidelity and texture realism.

Large Language Models (LLMs) are traditionally evaluated on multiple-choice question answering (MCQA) tasks using First-Token Probability (FTP), which selects the answer option whose initial token has the highest likelihood. While efficient, FTP can be fragile: models may assign high probability to unrelated tokens (misalignment) or use a valid token merely as part of a generic preamble rather than as a clear answer choice (misinterpretation), undermining the reliability of symbolic evaluation. We propose a simple solution: output prefilling, a structured natural-language prefix (e.g., 'The correct option is:') prepended to the model output. Originally explored in AI safety as an attack strategy, we repurpose prefilling to steer the model to respond with a clean, valid option, without modifying its parameters. Through extensive evaluation, we find that the FTP with prefilling strategy substantially improves accuracy, calibration, and output consistency across a broad set of LLMs and MCQA benchmarks. It outperforms standard FTP and often matches the performance of open-ended generation approaches that require full decoding and external classifiers, while being significantly more efficient. Our analysis suggests that prefilling is a simple, robust, and zero-cost method to enhance the reliability of FTP-based evaluation in multiple-choice settings.

Representation alignment has emerged as an effective approach to improve Multimodal Large Language Models (MLLMs) by regularizing their internal representations toward those of an external vision encoder. However, existing methods typically align a fixed layer of the language backbone, overlooking the fine-grained structure of Transformer models. In this work, we propose Head-Wise Representation Alignment (HeRA), a method that enforces cross-modal alignment at the level of individual attention heads. Our approach is grounded in the Platonic Representation Hypothesis, focusing on preserving the topological structure of representations (i.e., their local neighborhood relationships) across modalities. Following the Mutual K-Nearest Neighbor (MKNN) alignment metric, we introduce a contrastive objective that acts as a differentiable proxy for matching local structures. HeRA applies this objective during multimodal training to specific attention heads in the LLM, selected by their alignment score according to the MKNN metric. Counterintuitively, we find that aligning the least aligned heads yields the largest gains. Extensive evaluations across multiple MLLMs and 18 benchmarks demonstrate that HeRA consistently improves performance on challenging vision-centric tasks and serves as an effective regularizer against visual hallucinations by naturally curbing the over-reliance on linguistic priors.