Vision-Language Models (VLMs) typically rely on static initial frames for video reasoning, restricting their ability to incorporate essential dynamic information as the reasoning process evolves. Existing methods that augment Chain-of-Thought (CoT) with additional frame information often exhibit suboptimal CoT quality and lack the crucial ability to synthesize visual information for hypothetical or counterfactual scenarios. We introduce Act-to-See (Act2See), a novel framework that enables active visual perception by empowering VLMs to actively interleave video frames within text CoTs. Act2See is developed via Supervised Fine-Tuning (SFT) on a high-quality dataset of reasoning traces generated by a frontier VLM. These traces integrate active calls to either retrieve existing frames or generate new ones, and are rigorously verified against human-annotated CoTs to ensure quality. This approach cultivates an emergent capability: at inference time, the model actively determines when to search for or synthesize the necessary visual evidence. Act2See establishes new state-of-the-art results on challenging benchmarks, including VideoEspresso and ViTIB, and outperforms comparable or larger models on Video-MME, EgoNormia, and VCR-Bench, demonstrating an advancement in enabling VLMs with active visual perception for video reasoning.

Human communication involves a complex interplay of verbal and nonverbal signals, essential for conveying meaning and achieving interpersonal goals. To develop socially intelligent AI technologies, it is crucial to develop models that can both comprehend and generate dyadic behavioral dynamics. To this end, we introduce the Seamless Interaction Dataset, a large-scale collection of over 4,000 hours of face-to-face interaction footage from over 4,000 participants in diverse contexts. This dataset enables the development of AI technologies that understand dyadic embodied dynamics, unlocking breakthroughs in virtual agents, telepresence experiences, and multimodal content analysis tools. We also develop a suite of models that utilize the dataset to generate dyadic motion gestures and facial expressions aligned with human speech. These models can take as input both the speech and visual behavior of their interlocutors. We present a variant with speech from an LLM model and integrations with 2D and 3D rendering methods, bringing us closer to interactive virtual agents. Additionally, we describe controllable variants of our motion models that can adapt emotional responses and expressivity levels, as well as generating more semantically-relevant gestures. Finally, we discuss methods for assessing the quality of these dyadic motion models, which are demonstrating the potential for more intuitive and responsive human-AI interactions.

Recent Large Vision-Language Models (LVLMs) have introduced a new paradigm for understanding and reasoning about image input through textual responses. Although they have achieved remarkable performance across a range of multi-modal tasks, they face the persistent challenge of hallucination, which introduces practical weaknesses and raises concerns about their reliable deployment in real-world applications. Existing work has explored contrastive decoding approaches to mitigate this issue, where the output of the original LVLM is compared and contrasted with that of a perturbed version. However, these methods require two or more queries that slow down LVLM response generation, making them less suitable for real-time applications. To overcome this limitation, we propose ONLY, a training-free decoding approach that requires only a single query and a one-layer intervention during decoding, enabling efficient real-time deployment. Specifically, we enhance textual outputs by selectively amplifying crucial textual information using a text-to-visual entropy ratio for each token. Extensive experimental results demonstrate that our proposed ONLY consistently outperforms state-of-the-art methods across various benchmarks while requiring minimal implementation effort and computational cost. Code is available at https://github.com/zifuwan/ONLY.

While recent Large Vision-Language Models (LVLMs) have shown remarkable performance in multi-modal tasks, they are prone to generating hallucinatory text responses that do not align with the given visual input, which restricts their practical applicability in real-world scenarios. In this work, inspired by the observation that the text-to-image generation process is the inverse of image-conditioned response generation in LVLMs, we explore the potential of leveraging text-to-image generative models to assist in mitigating hallucinations in LVLMs. We discover that generative models can offer valuable self-feedback for mitigating hallucinations at both the response and token levels. Building on this insight, we introduce self-correcting Decoding with Generative Feedback (DeGF), a novel training-free algorithm that incorporates feedback from text-to-image generative models into the decoding process to effectively mitigate hallucinations in LVLMs. Specifically, DeGF generates an image from the initial response produced by LVLMs, which acts as an auxiliary visual reference and provides self-feedback to verify and correct the initial response through complementary or contrastive decoding. Extensive experimental results validate the effectiveness of our approach in mitigating diverse types of hallucinations, consistently surpassing state-of-the-art methods across six benchmarks. Code is available at https://github.com/zhangce01/DeGF.

Creating artificial social intelligence - algorithms that can understand the nuances of multi-person interactions - is an exciting and emerging challenge in processing facial expressions and gestures from multimodal videos. Recent multimodal methods have set the state of the art on many tasks, but have difficulty modeling the complex face-to-face conversational dynamics across speaking turns in social interaction, particularly in a self-supervised setup. In this paper, we propose Face-to-Face Contrastive Learning (F2F-CL), a graph neural network designed to model social interactions using factorization nodes to contextualize the multimodal face-to-face interaction along the boundaries of the speaking turn. With the F2F-CL model, we propose to perform contrastive learning between the factorization nodes of different speaking turns within the same video. We experimentally evaluated the challenging Social-IQ dataset and show state-of-the-art results.

In a wide range of multimodal tasks, contrastive learning has become a particularly appealing approach since it can successfully learn representations from abundant unlabeled data with only pairing information (e.g., image-caption or video-audio pairs). Underpinning these approaches is the assumption of multi-view redundancy - that shared information between modalities is necessary and sufficient for downstream tasks. However, in many real-world settings, task-relevant information is also contained in modality-unique regions: information that is only present in one modality but still relevant to the task. How can we learn self-supervised multimodal representations to capture both shared and unique information relevant to downstream tasks? This paper proposes FactorCL, a new multimodal representation learning method to go beyond multi-view redundancy. FactorCL is built from three new contributions: (1) factorizing task-relevant information into shared and unique representations, (2) capturing task-relevant information via maximizing MI lower bounds and removing task-irrelevant information via minimizing MI upper bounds, and (3) multimodal data augmentations to approximate task relevance without labels. On large-scale real-world datasets, FactorCL captures both shared and unique information and achieves state-of-the-art results on six benchmarks

In this work, we present an approach, which we call Embeddings for Language/Image-aligned X-Rays, or ELIXR, that leverages a language-aligned image encoder combined or grafted onto a fixed LLM, PaLM 2, to perform a broad range of chest X-ray tasks. We train this lightweight adapter architecture using images paired with corresponding free-text radiology reports from the MIMIC-CXR dataset. ELIXR achieved state-of-the-art performance on zero-shot chest X-ray (CXR) classification (mean AUC of 0.850 across 13 findings), data-efficient CXR classification (mean AUCs of 0.893 and 0.898 across five findings (atelectasis, cardiomegaly, consolidation, pleural effusion, and pulmonary edema) for 1% (~2,200 images) and 10% (~22,000 images) training data), and semantic search (0.76 normalized discounted cumulative gain (NDCG) across nineteen queries, including perfect retrieval on twelve of them). Compared to existing data-efficient methods including supervised contrastive learning (SupCon), ELIXR required two orders of magnitude less data to reach similar performance. ELIXR also showed promise on CXR vision-language tasks, demonstrating overall accuracies of 58.7% and 62.5% on visual question answering and report quality assurance tasks, respectively. These results suggest that ELIXR is a robust and versatile approach to CXR AI.

Given an unsupervised outlier detection task, how should one select i) a detection algorithm, and ii) associated hyperparameter values (jointly called a model)? E ective outlier model selection is essential as di erent algorithms may work well for varying detection tasks, and moreover their performance can be quite sensitive to the values of the hyperparameters (HPs). On the other hand, unsupervised model selection is notoriously difficult, in the absence of hold-out validation data with ground-truth labels. Therefore, the problem is vastly understudied in the outlier mining literature. There exists a body of work that propose internal model evaluation strate- gies for selecting a model. These so-called internal strategies solely rely on the input data (without labels) and the output (outlier scores) of the candidate models. In this paper, we rst survey internal model evaluation strategies including both those proposed speci cally for outlier detection, as well as those that can be adapted from the unsupervised deep representation learning literature. Then, we investigate their e ectiveness empirically in comparison to simple baselines such as random selection and the popular state-of-the-art detector Isolation Forest (iForest) with default HPs. To this end, we set up (and open-source) a large testbed with 39 detection tasks and 297 candidate models comprised of 8 different detectors and various HP con gurations. We evaluate internal strategies from 7 di erent families on their ability to discriminate between models w.r.t. detection performance, without using any labels. Our study reports a striking nding, that none of the existing and adapted strategies would be practically useful: stand-alone ones are not signi cantly di erent from random, and consensus-based ones do not outperform iForest (w/ default HPs) while being more expensive (as all candidate models need to be trained for evaluation). Our survey stresses the importance of and the standing need for e ective unsupervised outlier model selection, and acts as a call for future work on the problem.

Masked autoencoder (MAE), a simple and effective self-supervised learning framework based on the reconstruction of masked image regions, has recently achieved prominent success in a variety of vision tasks. Despite the emergence of intriguing empirical observations on MAE, a theoretically principled understanding is still lacking. In this work, we formally characterize and justify existing empirical in-sights and provide theoretical guarantees of MAE. We formulate the underlying data-generating process as a hierarchical latent variable model, and show that under reasonable assumptions, MAE provably identifies a set of latent variables in the hierarchical model, explaining why MAE can extract high-level information from pixels. Further, we show how key hyperparameters in MAE (the masking ratio and the patch size) determine which true latent variables to be recovered, therefore influencing the level of semantic information in the representation. Specifically, extremely large or small masking ratios inevitably lead to low-level representations. Our theory offers coherent explanations of existing empirical observations and provides insights for potential empirical improvements and fundamental limitations of the masked-reconstruction paradigm. We conduct extensive experiments to validate our theoretical insights.

Conditional contrastive learning frameworks consider the conditional sampling procedure that constructs positive or negative data pairs conditioned on specific variables. Fair contrastive learning constructs negative pairs, for example, from the same gender (conditioning on sensitive information), which in turn reduces undesirable information from the learned representations; weakly supervised contrastive learning constructs positive pairs with similar annotative attributes (conditioning on auxiliary information), which in turn are incorporated into the representations. Although conditional contrastive learning enables many applications, the conditional sampling procedure can be challenging if we cannot obtain sufficient data pairs for some values of the conditioning variable. This paper presents Conditional Contrastive Learning with Kernel (CCL-K) that converts existing conditional contrastive objectives into alternative forms that mitigate the insufficient data problem. Instead of sampling data according to the value of the conditioning variable, CCL-K uses the Kernel Conditional Embedding Operator that samples data from all available data and assigns weights to each sampled data given the kernel similarity between the values of the conditioning variable. We conduct experiments using weakly supervised, fair, and hard negatives contrastive learning, showing CCL-K outperforms state-of-the-art baselines.

Given an unsupervised outlier detection task, how should one select a detection algorithm as well as its hyperparameters (jointly called a model)? Unsupervised model selection is notoriously difficult, in the absence of hold-out validation data with ground-truth labels. Therefore, the problem is vastly understudied. In this work, we study the feasibility of employing internal model evaluation strategies for selecting a model for outlier detection. These so-called internal strategies solely rely on the input data (without labels) and the output (outlier scores) of the candidate models. We setup (and open-source) a large testbed with 39 detection tasks and 297 candidate models comprised of 8 detectors and various hyperparameter configurations. We evaluate 7 different strategies on their ability to discriminate between models w.r.t. detection performance, without using any labels. Our study reveals room for progress—we find that none would be practically useful, as they select models only comparable to a state-of-the-art detector (with random configuration).

This paper introduces Relative Predictive Coding (RPC), a new contrastive representation learning objective that maintains a good balance among training stability, minibatch size sensitivity, and downstream task performance. The key to the success of RPC is two-fold. First, RPC introduces the relative parameters to regularize the objective for boundedness and low variance. Second, RPC contains no logarithm and exponential score functions, which are the main cause of training instability in prior contrastive objectives. We empirically verify the effectiveness of RPC on benchmark vision and speech self-supervised learning tasks. Lastly, we relate RPC with mutual information (MI) estimation, showing RPC can be used to estimate MI with low variance.

Contrastive self-supervised learning (SSL) learns an embedding space that maps similar data pairs closer and dissimilar data pairs farther apart. Despite its success, one issue has been overlooked: the fairness aspect of representations learned using contrastive SSL. Without mitigation, contrastive SSL techniques can incorporate sensitive information such as gender or race and cause potentially unfair predictions on downstream tasks. In this paper, we propose a Conditional Contrastive Learning (CCL) approach to improve the fairness of contrastive SSL methods. Our approach samples positive and negative pairs from distributions conditioning on the sensitive attribute, or empirically speaking, sampling positive and negative pairs from the same gender or the same race. We show that our approach provably maximizes the conditional mutual information between the learned representations of the positive pairs, and reduces the effect of the sensitive attribute by taking it as the conditional variable. On seven fairness and vision datasets, we empirically demonstrate that the proposed approach achieves state-of-the-art downstream performances compared to unsupervised baselines and significantly improves the fairness of contrastive SSL models on multiple fairness metrics.

Self-supervised learning is a form of unsupervised learning that leverages rich information in data to learn representations. However, data sometimes contains certain information that may be undesirable for downstream tasks. For instance, gender information may lead to biased decisions on many gender-irrelevant tasks. In this paper, we develop conditional contrastive learning to remove undesirable information in self-supervised representations. To remove the effect of the undesirable variable, our proposed approach conditions on the undesirable variable (i.e., by fixing the variations of it) during the contrastive learning process. In particular, inspired by the contrastive objective InfoNCE, we introduce Conditional InfoNCE (C-InfoNCE), and its computationally efficient variant, Weak-Conditional InfoNCE (WeaC-InfoNCE), for conditional contrastive learning. We demonstrate empirically that our methods can successfully learn self-supervised representations for downstream tasks while removing a great level of information related to the undesirable variables. We study three scenarios, each with a different type of undesirable variables: task-irrelevant meta-information for self-supervised speech representation learning, sensitive attributes for fair representation learning, and domain specification for multi-domain visual representation learning.

The human language can be expressed through multiple sources of information known as modalities, including tones of voice, facial gestures, and spoken language. Recent multimodal learning with strong performances on human-centric tasks such as sentiment analysis and emotion recognition are often black-box, with very limited interpretability. In this paper we propose Multimodal Routing, which dynamically adjusts weights between input modalities and output representations differently for each input sample. Multimodal routing can identify relative importance of both individual modalities and cross-modality features. Moreover, the weight assignment by routing allows us to interpret modality-prediction relationships not only globally (i.e. general trends over the whole dataset), but also locally for each single input sample, meanwhile keeping competitive performance compared to state-of-the-art methods.

While deep learning has received a surge of interest in a variety of fields in recent years, major deep learning models barely use complex numbers. However, speech, signal and audio data are naturally complex-valued after Fourier Transform, and studies have shown a potentially richer representation of complex nets. In this paper, we propose a Complex Transformer, which incorporates the transformer model as a backbone for sequence modeling; we also develop attention and encoder-decoder network operating for complex input. The model achieves state-of-the-art performance on the MusicNet dataset and an In-phase Quadrature (IQ) signal dataset.