Leonid Sigal

Raghav Goyal, Effrosyni Mavroudi, Xitong Yang et al. · meta-ai

Video understanding tasks take many forms, from action detection to visual query localization and spatio-temporal grounding of sentences. These tasks differ in the type of inputs (only video, or video-query pair where query is an image region or sentence) and outputs (temporal segments or spatio-temporal tubes). However, at their core they require the same fundamental understanding of the video, i.e., the actors and objects in it, their actions and interactions. So far these tasks have been tackled in isolation with individual, highly specialized architectures, which do not exploit the interplay between tasks. In contrast, in this paper, we present a single, unified model for tackling query-based video understanding in long-form videos. In particular, our model can address all three tasks of the Ego4D Episodic Memory benchmark which entail queries of three different forms: given an egocentric video and a visual, textual or activity query, the goal is to determine when and where the answer can be seen within the video. Our model design is inspired by recent query-based approaches to spatio-temporal grounding, and contains modality-specific query encoders and task-specific sliding window inference that allow multi-task training with diverse input modalities and different structured outputs. We exhaustively analyze relationships among the tasks and illustrate that cross-task learning leads to improved performance on each individual task, as well as the ability to generalize to unseen tasks, such as zero-shot spatial localization of language queries.

Jiayun Luo, Siddhesh Khandelwal, Leonid Sigal et al.

From image-text pairs, large-scale vision-language models (VLMs) learn to implicitly associate image regions with words, which prove effective for tasks like visual question answering. However, leveraging the learned association for open-vocabulary semantic segmentation remains a challenge. In this paper, we propose a simple, yet extremely effective, training-free technique, Plug-and-Play Open-Vocabulary Semantic Segmentation (PnP-OVSS) for this task. PnP-OVSS leverages a VLM with direct text-to-image cross-attention and an image-text matching loss. To balance between over-segmentation and under-segmentation, we introduce Salience Dropout; by iteratively dropping patches that the model is most attentive to, we are able to better resolve the entire extent of the segmentation mask. PnP-OVSS does not require any neural network training and performs hyperparameter tuning without the need for any segmentation annotations, even for a validation set. PnP-OVSS demonstrates substantial improvements over comparable baselines (+26.2% mIoU on Pascal VOC, +20.5% mIoU on MS COCO, +3.1% mIoU on COCO Stuff and +3.0% mIoU on ADE20K). Our codebase is at https://github.com/letitiabanana/PnP-OVSS.

Mohammed Suhail, Carlos Esteves, Leonid Sigal et al.

Neural rendering has received tremendous attention since the advent of Neural Radiance Fields (NeRF), and has pushed the state-of-the-art on novel-view synthesis considerably. The recent focus has been on models that overfit to a single scene, and the few attempts to learn models that can synthesize novel views of unseen scenes mostly consist of combining deep convolutional features with a NeRF-like model. We propose a different paradigm, where no deep features and no NeRF-like volume rendering are needed. Our method is capable of predicting the color of a target ray in a novel scene directly, just from a collection of patches sampled from the scene. We first leverage epipolar geometry to extract patches along the epipolar lines of each reference view. Each patch is linearly projected into a 1D feature vector and a sequence of transformers process the collection. For positional encoding, we parameterize rays as in a light field representation, with the crucial difference that the coordinates are canonicalized with respect to the target ray, which makes our method independent of the reference frame and improves generalization. We show that our approach outperforms the state-of-the-art on novel view synthesis of unseen scenes even when being trained with considerably less data than prior work.

Tanzila Rahman, Hsin-Ying Lee, Jian Ren et al.

There has been a recent explosion of impressive generative models that can produce high quality images (or videos) conditioned on text descriptions. However, all such approaches rely on conditional sentences that contain unambiguous descriptions of scenes and main actors in them. Therefore employing such models for more complex task of story visualization, where naturally references and co-references exist, and one requires to reason about when to maintain consistency of actors and backgrounds across frames/scenes, and when not to, based on story progression, remains a challenge. In this work, we address the aforementioned challenges and propose a novel autoregressive diffusion-based framework with a visual memory module that implicitly captures the actor and background context across the generated frames. Sentence-conditioned soft attention over the memories enables effective reference resolution and learns to maintain scene and actor consistency when needed. To validate the effectiveness of our approach, we extend the MUGEN dataset and introduce additional characters, backgrounds and referencing in multi-sentence storylines. Our experiments for story generation on the MUGEN, the PororoSV and the FlintstonesSV dataset show that our method not only outperforms prior state-of-the-art in generating frames with high visual quality, which are consistent with the story, but also models appropriate correspondences between the characters and the background.

Sahithya Ravi, Aditya Chinchure, Leonid Sigal et al.

There has been a growing interest in solving Visual Question Answering (VQA) tasks that require the model to reason beyond the content present in the image. In this work, we focus on questions that require commonsense reasoning. In contrast to previous methods which inject knowledge from static knowledge bases, we investigate the incorporation of contextualized knowledge using Commonsense Transformer (COMET), an existing knowledge model trained on human-curated knowledge bases. We propose a method to generate, select, and encode external commonsense knowledge alongside visual and textual cues in a new pre-trained Vision-Language-Commonsense transformer model, VLC-BERT. Through our evaluation on the knowledge-intensive OK-VQA and A-OKVQA datasets, we show that VLC-BERT is capable of outperforming existing models that utilize static knowledge bases. Furthermore, through a detailed analysis, we explain which questions benefit, and which don't, from contextualized commonsense knowledge from COMET.

Junpeng Jing, Jiankun Li, Pengfei Xiong et al.

Correlation based stereo matching has achieved outstanding performance, which pursues cost volume between two feature maps. Unfortunately, current methods with a fixed model do not work uniformly well across various datasets, greatly limiting their real-world applicability. To tackle this issue, this paper proposes a new perspective to dynamically calculate correlation for robust stereo matching. A novel Uncertainty Guided Adaptive Correlation (UGAC) module is introduced to robustly adapt the same model for different scenarios. Specifically, a variance-based uncertainty estimation is employed to adaptively adjust the sampling area during warping operation. Additionally, we improve the traditional non-parametric warping with learnable parameters, such that the position-specific weights can be learned. We show that by empowering the recurrent network with the UGAC module, stereo matching can be exploited more robustly and effectively. Extensive experiments demonstrate that our method achieves state-of-the-art performance over the ETH3D, KITTI, and Middlebury datasets when employing the same fixed model over these datasets without any retraining procedure. To target real-time applications, we further design a lightweight model based on UGAC, which also outperforms other methods over KITTI benchmarks with only 0.6 M parameters.

Siddhesh Khandelwal, Leonid Sigal

The task of scene graph generation entails identifying object entities and their corresponding interaction predicates in a given image (or video). Due to the combinatorially large solution space, existing approaches to scene graph generation assume certain factorization of the joint distribution to make the estimation feasible (e.g., assuming that objects are conditionally independent of predicate predictions). However, this fixed factorization is not ideal under all scenarios (e.g., for images where an object entailed in interaction is small and not discernible on its own). In this work, we propose a novel framework for scene graph generation that addresses this limitation, as well as introduces dynamic conditioning on the image, using message passing in a Markov Random Field. This is implemented as an iterative refinement procedure wherein each modification is conditioned on the graph generated in the previous iteration. This conditioning across refinement steps allows joint reasoning over entities and relations. This framework is realized via a novel and end-to-end trainable transformer-based architecture. In addition, the proposed framework can improve existing approach performance. Through extensive experiments on Visual Genome and Action Genome benchmark datasets we show improved performance on the scene graph generation.

Yanwei Fu, Xiaomei Wang, Hanze Dong et al.

Despite significant progress in object categorization, in recent years, a number of important challenges remain; mainly, the ability to learn from limited labeled data and to recognize object classes within large, potentially open, set of labels. Zero-shot learning is one way of addressing these challenges, but it has only been shown to work with limited sized class vocabularies and typically requires separation between supervised and unsupervised classes, allowing former to inform the latter but not vice versa. We propose the notion of vocabulary-informed learning to alleviate the above mentioned challenges and address problems of supervised, zero-shot, generalized zero-shot and open set recognition using a unified framework. Specifically, we propose a weighted maximum margin framework for semantic manifold-based recognition that incorporates distance constraints from (both supervised and unsupervised) vocabulary atoms. Distance constraints ensure that labeled samples are projected closer to their correct prototypes, in the embedding space, than to others. We illustrate that resulting model shows improvements in supervised, zero-shot, generalized zero-shot, and large open set recognition, with up to 310K class vocabulary on Animal with Attributes and ImageNet datasets.

Tianyu Hua, Yonglong Tian, Sucheng Ren et al.

Inspired by the success of self-supervised autoregressive representation learning in natural language (GPT and its variants), and advances in recent visual architecture design with Vision Transformers (ViTs), in this paper, we explore the effect various design choices have on the success of applying such training strategies for visual feature learning. Specifically, we introduce a novel strategy that we call Random Segments with Autoregressive Coding (RandSAC). In RandSAC, we group patch representations (image tokens) into hierarchically arranged segments; within each segment, tokens are predicted in parallel, similar to BERT, while across segment predictions are sequential, similar to GPT. We illustrate that randomized serialization of the segments significantly improves the performance and results in distribution over spatially-long (across-segments) and -short (within-segment) predictions which are effective for feature learning. We illustrate the pertinence of these design choices and explore alternatives on a number of datasets (e.g., CIFAR10, CIFAR100, ImageNet). While our pre-training strategy works with a vanilla Transformer, we also propose a conceptually simple, but highly effective, addition to the decoder that allows learnable skip-connections to encoder$'$s feature layers, which further improves the performance.

Jiahui Huang, Leonid Sigal, Kwang Moo Yi et al.

We present Interactive Neural Video Editing (INVE), a real-time video editing solution, which can assist the video editing process by consistently propagating sparse frame edits to the entire video clip. Our method is inspired by the recent work on Layered Neural Atlas (LNA). LNA, however, suffers from two major drawbacks: (1) the method is too slow for interactive editing, and (2) it offers insufficient support for some editing use cases, including direct frame editing and rigid texture tracking. To address these challenges we leverage and adopt highly efficient network architectures, powered by hash-grids encoding, to substantially improve processing speed. In addition, we learn bi-directional functions between image-atlas and introduce vectorized editing, which collectively enables a much greater variety of edits in both the atlas and the frames directly. Compared to LNA, our INVE reduces the learning and inference time by a factor of 5, and supports various video editing operations that LNA cannot. We showcase the superiority of INVE over LNA in interactive video editing through a comprehensive quantitative and qualitative analysis, highlighting its numerous advantages and improved performance. For video results, please see https://gabriel-huang.github.io/inve/

Gaurav Bhatt, Deepayan Das, Leonid Sigal et al.

Intelligent systems possess a crucial characteristic of breaking complicated problems into smaller reusable components or parts and adjusting to new tasks using these part representations. However, current part-learners encounter difficulties in dealing with incidental correlations resulting from the limited observations of objects that may appear only in specific arrangements or with specific backgrounds. These incidental correlations may have a detrimental impact on the generalization and interpretability of learned part representations. This study asserts that part-based representations could be more interpretable and generalize better with limited data, employing two innovative regularization methods. The first regularization separates foreground and background information's generative process via a unique mixture-of-parts formulation. Structural constraints are imposed on the parts using a weakly-supervised loss, guaranteeing that the mixture-of-parts for foreground and background entails soft, object-agnostic masks. The second regularization assumes the form of a distillation loss, ensuring the invariance of the learned parts to the incidental background correlations. Furthermore, we incorporate sparse and orthogonal constraints to facilitate learning high-quality part representations. By reducing the impact of incidental background correlations on the learned parts, we exhibit state-of-the-art (SoTA) performance on few-shot learning tasks on benchmark datasets, including MiniImagenet, TieredImageNet, and FC100. We also demonstrate that the part-based representations acquired through our approach generalize better than existing techniques, even under domain shifts of the background and common data corruption on the ImageNet-9 dataset. The implementation is available on GitHub: https://github.com/GauravBh1010tt/DPViT.git

Muchen Li, Jeffrey Yunfan Liu, Leonid Sigal et al.

Neural architectures can be naturally viewed as computational graphs. Motivated by this perspective, we, in this paper, study neural architecture search (NAS) through the lens of learning random graph models. In contrast to existing NAS methods which largely focus on searching for a single best architecture, i.e, point estimation, we propose GraphPNAS a deep graph generative model that learns a distribution of well-performing architectures. Relying on graph neural networks (GNNs), our GraphPNAS can better capture topologies of good neural architectures and relations between operators therein. Moreover, our graph generator leads to a learnable probabilistic search method that is more flexible and efficient than the commonly used RNN generator and random search methods. Finally, we learn our generator via an efficient reinforcement learning formulation for NAS. To assess the effectiveness of our GraphPNAS, we conduct extensive experiments on three search spaces, including the challenging RandWire on TinyImageNet, ENAS on CIFAR10, and NAS-Bench-101/201. The complexity of RandWire is significantly larger than other search spaces in the literature. We show that our proposed graph generator consistently outperforms RNN-based one and achieves better or comparable performances than state-of-the-art NAS methods.

Siddhesh Khandelwal, Anirudth Nambirajan, Behjat Siddiquie et al.

Methods for object detection and segmentation often require abundant instance-level annotations for training, which are time-consuming and expensive to collect. To address this, the task of zero-shot object detection (or segmentation) aims at learning effective methods for identifying and localizing object instances for the categories that have no supervision available. Constructing architectures for these tasks requires choosing from a myriad of design options, ranging from the form of the class encoding used to transfer information from seen to unseen categories, to the nature of the function being optimized for learning. In this work, we extensively study these design choices, and carefully construct a simple yet extremely effective zero-shot recognition method. Through extensive experiments on the MSCOCO dataset on object detection and segmentation, we highlight that our proposed method outperforms existing, considerably more complex, architectures. Our findings and method, which we propose as a competitive future baseline, point towards the need to revisit some of the recent design trends in zero-shot detection / segmentation.

Peyman Bateni, Leonid Sigal

Stress is considered to be the epidemic of the 21st-century. Yet, mobile apps cannot directly evaluate the impact of their content and services on user stress. We introduce the Beam AI SDK to address this issue. Using our SDK, apps can monitor user stress through the selfie camera in real-time. Our technology extracts the user's pulse wave by analyzing subtle color variations across the skin regions of the user's face. The user's pulse wave is then used to determine stress (according to the Baevsky Stress Index), heart rate, and heart rate variability. We evaluate our technology on the UBFC dataset, the MMSE-HR dataset, and Beam AI's internal data. Our technology achieves 99.2%, 97.8% and 98.5% accuracy for heart rate estimation on each benchmark respectively, a nearly twice lower error rate than competing methods. We further demonstrate an average Pearson correlation of 0.801 in determining stress and heart rate variability, thus producing commercially useful readings to derive content decisions in apps. Our SDK is available for use at www.beamhealth.ai.

Shih-Han Chou, James J. Little, Leonid Sigal

Existing dense or paragraph video captioning approaches rely on holistic representations of videos, possibly coupled with learned object/action representations, to condition hierarchical language decoders. However, they fundamentally lack the commonsense knowledge of the world required to reason about progression of events, causality, and even the function of certain objects within a scene. To address this limitation we propose a novel video captioning Transformer-based model, that takes into account both implicit (visuo-lingual and purely linguistic) and explicit (knowledge-base) commonsense knowledge. We show that these forms of knowledge, in isolation and in combination, enhance the quality of produced captions. Further, inspired by imitation learning, we propose a new task of instruction generation, where the goal is to produce a set of linguistic instructions from a video demonstration of its performance. We formalize the task using the ALFRED dataset [54] generated using an AI2-THOR environment. While instruction generation is conceptually similar to paragraph captioning, it differs in the fact that it exhibits stronger object persistence, as well as spatially-aware and causal sentence structure. We show that our commonsense knowledge enhanced approach produces significant improvements on this task (up to 57% in METEOR and 8.5% in CIDEr), as well as the state-of-the-art result on more traditional video captioning in the ActivityNet Captions dataset [29].

Ailar Mahdizadeh, Puria Azadi, Muchen Li et al.

Streaming video understanding with large vision-language models (VLMs) requires a compact memory that can support future reasoning over an ever-growing visual history. A common solution is to compress the key-value (KV) cache, but existing streaming methods typically rely on local token-wise heuristics, such as recency, temporal redundancy, or saliency, which do not explicitly optimize whether the retained cache is representative of the accumulated history. We propose to view KV-cache compression as a coreset selection problem: rather than scoring tokens independently for retention, we select a small subset that covers the geometry of the accumulated visual cache. Our method operates in a joint KV representation and introduces a bicriteria objective that balances coverage in key and value spaces, preserving both retrieval structure and output-relevant information. To encourage a more diverse retained subset, we further introduce an orthogonality-driven diversity criterion that favors candidates contributing new directions beyond the current selection, and connect this criterion to log-determinant subset selection. Across four open-source VLMs and five long-video and streaming-video benchmarks, our method improves over heuristic streaming compression baselines under a fixed cache budget. These results highlight that representative coreset selection offers a more effective principle, than token-wise pruning, for memory-constrained streaming video understanding.

Wan-Cyuan Fan, Yen-Chun Chen, Mengchen Liu et al.

Recent studies customizing Multimodal Large Language Models (MLLMs) for domain-specific tasks have yielded promising results, especially in the field of scientific chart comprehension. These studies generally utilize visual instruction tuning with specialized datasets to enhance question and answer (QA) accuracy within the chart domain. However, they often neglect the fundamental discrepancy between natural image-caption pre-training data and digital chart image-QA data, particularly in the models' capacity to extract underlying numeric values from charts. This paper tackles this oversight by exploring the training processes necessary to improve MLLMs' comprehension of charts. We present three key findings: (1) Incorporating raw data values in alignment pre-training markedly improves comprehension of chart data. (2) Replacing images with their textual representation randomly during end-to-end fine-tuning transfer the language reasoning capability to chart interpretation skills. (3) Requiring the model to first extract the underlying chart data and then answer the question in the fine-tuning can further improve the accuracy. Consequently, we introduce CHOPINLLM, an MLLM tailored for in-depth chart comprehension. CHOPINLLM effectively interprets various types of charts, including unannotated ones, while maintaining robust reasoning abilities. Furthermore, we establish a new benchmark to evaluate MLLMs' understanding of different chart types across various comprehension levels. Experimental results show that CHOPINLLM exhibits strong performance in understanding both annotated and unannotated charts across a wide range of types.

Bicheng Xu, Renjie Liao, Leonid Sigal

The goal of scene graph generation is to predict a graph from an input image, where nodes correspond to identified and localized objects and edges to their corresponding interaction predicates. Existing methods are trained in a fully supervised manner and focus on message passing mechanisms, loss functions, and/or bias mitigation. In this work we introduce a simple-yet-effective self-supervised relational alignment regularization designed to improve the scene graph generation performance. The proposed alignment is general and can be combined with any existing scene graph generation framework, where it is trained alongside the original model's objective. The alignment is achieved through distillation, where an auxiliary relation prediction branch, that mirrors and shares parameters with the supervised counterpart, is designed. In the auxiliary branch, relational input features are partially masked prior to message passing and predicate prediction. The predictions for masked relations are then aligned with the supervised counterparts after the message passing. We illustrate the effectiveness of this self-supervised relational alignment in conjunction with two scene graph generation architectures, SGTR and Neural Motifs, and show that in both cases we achieve significantly improved performance.

Ailar Mahdizadeh, Puria Azadi Moghadam, Xiangteng He et al.

Vision-language models (VLMs) have demonstrated strong cross-modal capabilities, yet most work remains limited to 2D data and assumes binary supervision (i.e., positive vs. negative pairs), overlooking the continuous and structured dependencies present in volumetric data such as CT. Existing approaches often treat volumetric scans as independent 2D slices, compromising spatial coherence and underutilizing rich clinical semantics. We propose SCALE-VLP, a soft-weighted contrastive vision-language pre-training framework that integrates (i) volumetric spatial semantics to preserve anatomical structure and (ii) domain-aware, knowledge-infused semantics (e.g., radiological ontologies) to guide alignment. This yields structurally consistent and semantically grounded representations under limited supervision, demonstrating strong cross-task transferability (retrieval, report generation, and classification), and cross-domain generalizability with consistent gains without further fine-tuning. In particular, compared to the previous state of the art, SCALE-VLP achieves up to 4.3x higher top-1 CT-report retrieval, improves abnormality classification by 10 points, and reaches ROUGE-L 0.44 and BERT-F1 0.89 for report generation. Further, in zero-shot evaluation on an out-of-domain external dataset, we observe consistent gains, indicating the cross-task and cross-domain generalization ability of SCALE-VLP.

Shivam Chandhok, Wan-Cyuan Fan, Leonid Sigal

Vision-Language Models (VLMs) have emerged as general purpose tools for addressing a variety of complex computer vision problems. Such models have been shown to be highly capable, but, at the same time, also lacking some basic visual understanding skills. In this paper, we set out to understand the limitations of SoTA VLMs on fundamental visual tasks: object classification, understanding spatial arrangement, and ability to delineate individual object instances (through counting), by constructing a series of tests that probe which components of design, specifically, maybe lacking. Importantly, we go significantly beyond the current benchmarks, that simply measure final performance of VLM, by also comparing and contrasting it to performance of probes trained directly on features obtained from visual encoder (image embeddings), as well as intermediate vision-language projection used to bridge image-encoder and LLM-decoder ouput in many SoTA models (e.g., LLaVA, BLIP, InstructBLIP). In doing so, we uncover nascent shortcomings in VLMs response and make a number of important observations which could help train and develop more effective VLM models in future.

Sadegh Mahdavi, Muchen Li, Kaiwen Liu et al.

Advances in Large Language Models (LLMs) have sparked interest in their ability to solve Olympiad-level math problems. However, the training and evaluation of these models are constrained by the limited size and quality of available datasets, as creating large-scale data for such advanced problems requires extensive effort from human experts. In addition, current benchmarks are prone to contamination, leading to unreliable evaluations. In this paper, we present an automated pipeline that leverages the rich resources of the Art of Problem Solving (AoPS) forum, which predominantly features Olympiad-level problems and community-driven solutions. Using open-source LLMs, we develop a method to extract question-answer pairs from the forum, resulting in AoPS-Instruct, a dataset of more than 600,000 high-quality QA pairs. Our experiments demonstrate that fine-tuning LLMs on AoPS-Instruct improves their reasoning abilities across various benchmarks. Moreover, we build an automatic pipeline that introduces LiveAoPSBench, an evolving evaluation set with timestamps, derived from the latest forum data, providing a contamination-resistant benchmark for assessing LLM performance. Notably, we observe a significant decline in LLM performance over time, suggesting their success on older examples may stem from pre-training exposure rather than true reasoning ability. Our work presents a scalable approach to creating and maintaining large-scale, high-quality datasets for advanced math reasoning, offering valuable insights into the capabilities and limitations of LLMs in this domain. Our benchmark and code is available at https://github.com/DSL-Lab/aops

Tanzila Rahman, Renjie Liao, Leonid Sigal

Training multimodal large language models (MLLMs) for video understanding requires large-scale annotated data spanning diverse tasks such as object counting, question answering, and segmentation. However, collecting and annotating multimodal video data in real-world is costly, slow, and inherently limited in diversity and coverage. To address this challenge, we propose a unified synthetic data generation pipeline capable of automatically producing unlimited multimodal video data with rich and diverse supervision. Our framework supports multiple task formats within a single pipeline, enabling scalable and consistent data creation across tasks. To further enhance reasoning ability, we introduce a VQA-based fine-tuning strategy that trains models to answer structured questions about visual content rather than relying solely on captions or simple instructions. This formulation encourages deeper visual grounding and reasoning. We evaluate our approach in three challenging tasks: video object counting, video-based visual question answering, and video object segmentation. Experimental results demonstrate that models trained predominantly on synthetic data generalize effectively to real-world datasets, often outperforming traditionally trained counterparts. Our findings highlight the potential of unified synthetic data pipelines as a scalable alternative to expensive real-world annotation for multimodal video understanding.

Aditya Chinchure, Sahithya Ravi, Raymond Ng et al.

The commonsense reasoning capabilities of vision-language models (VLMs), especially in abductive reasoning and defeasible reasoning, remain poorly understood. Most benchmarks focus on typical visual scenarios, making it difficult to discern whether model performance stems from keen perception and reasoning skills, or reliance on pure statistical recall. We argue that by focusing on atypical events in videos, clearer insights can be gained on the core capabilities of VLMs. Explaining and understanding such out-of-distribution events requires models to extend beyond basic pattern recognition and regurgitation of their prior knowledge. To this end, we introduce BlackSwanSuite, a benchmark for evaluating VLMs' ability to reason about unexpected events through abductive and defeasible tasks. Our tasks artificially limit the amount of visual information provided to models while questioning them about hidden unexpected events, or provide new visual information that could change an existing hypothesis about the event. We curate a comprehensive benchmark suite comprising over 3,800 MCQ, 4,900 generative and 6,700 yes/no questions, spanning 1,655 videos. After extensively evaluating various state-of-the-art VLMs, including GPT-4o and Gemini 1.5 Pro, as well as open-source VLMs such as LLaVA-Video, we find significant performance gaps of up to 32% from humans on these tasks. Our findings reveal key limitations in current VLMs, emphasizing the need for enhanced model architectures and training strategies. Our data and leaderboard is available at blackswan.cs.ubc.ca.

Yasamin Medghalchi, Moein Heidari, Clayton Allard et al.

Deep neural networks (DNNs) offer significant promise for improving breast cancer diagnosis in medical imaging. However, these models are highly susceptible to adversarial attacks--small, imperceptible changes that can mislead classifiers--raising critical concerns about their reliability and security. Traditional attacks rely on fixed-norm perturbations, misaligning with human perception. In contrast, diffusion-based attacks require pre-trained models, demanding substantial data when these models are unavailable, limiting practical use in data-scarce scenarios. In medical imaging, however, this is often unfeasible due to the limited availability of datasets. Building on recent advancements in learnable prompts, we propose Prompt2Perturb (P2P), a novel language-guided attack method capable of generating meaningful attack examples driven by text instructions. During the prompt learning phase, our approach leverages learnable prompts within the text encoder to create subtle, yet impactful, perturbations that remain imperceptible while guiding the model towards targeted outcomes. In contrast to current prompt learning-based approaches, our P2P stands out by directly updating text embeddings, avoiding the need for retraining diffusion models. Further, we leverage the finding that optimizing only the early reverse diffusion steps boosts efficiency while ensuring that the generated adversarial examples incorporate subtle noise, thus preserving ultrasound image quality without introducing noticeable artifacts. We show that our method outperforms state-of-the-art attack techniques across three breast ultrasound datasets in FID and LPIPS. Moreover, the generated images are both more natural in appearance and more effective compared to existing adversarial attacks. Our code will be publicly available https://github.com/yasamin-med/P2P.

Mir Rayat Imtiaz Hossain, Leo Feng, Leonid Sigal et al.

Large Language Models (LLMs) have emerged as promising recommendation systems, offering novel ways to model user preferences through generative approaches. However, many existing methods often rely solely on text semantics or incorporate collaborative signals in a limited manner, typically using only user or item embeddings. These methods struggle to handle multiple item embeddings representing user history, reverting to textual semantics and neglecting richer collaborative information. In this work, we propose a simple yet effective solution that projects user and item embeddings, learned from collaborative filtering, into the LLM token space via separate lightweight projector modules. A finetuned LLM then conditions on these projected embeddings alongside textual tokens to generate recommendations. Preliminary results show that this design effectively leverages structured user-item interaction data, improves recommendation performance over text-only LLM baselines, and offers a practical path for bridging traditional recommendation systems with modern LLMs.

Wan-Cyuan Fan, Jiayun Luo, Declan Kutscher et al.

Vision-Language Models (VLMs) have been shown to be blind, often underutilizing their visual inputs even on tasks that require visual reasoning. In this work, we demonstrate that VLMs are selectively blind. They modulate the amount of attention applied to visual inputs based on linguistic framing even when alternative framings demand identical visual reasoning. Using visual attention as a probe, we quantify how framing alters both the amount and distribution of attention over the image. Constrained framings, such as multiple choice and yes/no, induce substantially lower attention to image context compared to open-ended, reduce focus on task-relevant regions, and shift attention towards uninformative tokens. We further demonstrate that this attention misallocation is the principal cause of degraded accuracy and cross-framing inconsistency. Building on this mechanistic insight, we introduce a lightweight prompt-tuning method using learnable tokens that encourages the robust, visually grounded attention patterns observed in open-ended settings, improving visual grounding and improving performance across framings.

Xiangteng He, Shunsuke Sakai, Kun Yuan et al.

Image-based Joint-Embedding Predictive Architecture (I-JEPA) learns visual representations by predicting latent embeddings of masked regions from visible context. However, it treats all regions uniformly and independently, lacking an explicit notion of where or in what order predictions should be made. Inspired by human visual perception, which deploys attention selectively and sequentially from the most informative to secondary regions, we propose DSeq-JEPA, a Discriminative Sequential Joint-Embedding Predictive Architecture that bridges predictive and autoregressive self-supervised learning, integrating JEPA-style latent prediction with GPT-style sequential reasoning. Specifically, DSeq-JEPA (i) first identifies primary discriminative regions based on a transformer-derived saliency map, emphasizing the distribution of visual importance, and then (ii) predicts subsequent regions in this discriminative order, progressively forming a curriculum-like semantic progression from primary to secondary cues -- a form of GPT-style pre-training. Extensive experiments across diverse tasks, including image classification (ImageNet), fine-grained visual categorization (iNaturalist21, CUB-200-2011, Stanford-Cars), detection and segmentation (MS-COCO, ADE20K), and low-level reasoning tasks (Clevr/Count, Clevr/Dist), demonstrate that DSeq-JEPA consistently focuses on more discriminative and generalizable representations than I-JEPA variants. Project page: https://github.com/SkyShunsuke/DSeq-JEPA.

Jia Jun Cheng Xian, Muchen Li, Haotian Yang et al.

Recent advances in diffusion-based text-to-image (T2I) models have led to remarkable success in generating high-quality images from textual prompts. However, ensuring accurate alignment between the text and the generated image remains a significant challenge for state-of-the-art diffusion models. To address this, existing studies employ reinforcement learning with human feedback (RLHF) to align T2I outputs with human preferences. These methods, however, either rely directly on paired image preference data or require a learned reward function, both of which depend heavily on costly, high-quality human annotations and thus face scalability limitations. In this work, we introduce Text Preference Optimization (TPO), a framework that enables "free-lunch" alignment of T2I models, achieving alignment without the need for paired image preference data. TPO works by training the model to prefer matched prompts over mismatched prompts, which are constructed by perturbing original captions using a large language model. Our framework is general and compatible with existing preference-based algorithms. We extend both DPO and KTO to our setting, resulting in TDPO and TKTO. Quantitative and qualitative evaluations across multiple benchmarks show that our methods consistently outperform their original counterparts, delivering better human preference scores and improved text-to-image alignment. Our Open-source code is available at https://github.com/DSL-Lab/T2I-Free-Lunch-Alignment.

Gaurav Bhatt, Kiran Koshy Thekumparampil, Tanmay Gangwani et al.

Traditional ranking systems optimize offline proxy objectives that rely on oversimplified assumptions about user behavior, often neglecting factors such as position bias and item diversity. Consequently, these models fail to improve true counterfactual utilities such as such as click-through rate or purchase probability, when evaluated in online A/B tests. We introduce RewardRank, a data-driven learning-to-rank (LTR) framework for counterfactual utility maximization. RewardRank first learns a reward model that predicts the utility of any ranking directly from logged user interactions, and then trains a ranker to maximize this reward using a differentiable soft permutation operator. To enable rigorous and reproducible evaluation, we further propose two benchmark suites: (i) Parametric Oracle Evaluation (PO-Eval), which employs an open-source click model as a counterfactual oracle on the Baidu-ULTR dataset, and (ii) LLM-as-User Evaluation (LAU-Eval), which simulates realistic user behavior via large language models on the Amazon-KDD-Cup dataset. RewardRank achieves the highest counterfactual utility across both benchmarks and demonstrates that optimizing classical metrics such as NDCG is sub-optimal for maximizing true user utility. Finally, using real user feedback from the Baidu-ULTR dataset, RewardRank establishes a new state of the art in offline relevance performance. Overall, our results show that learning-to-rank can be reformulated as direct optimization of counterfactual utility, achieved in a purely data-driven manner without relying on explicit modeling assumptions such as position bias. Our code is available at: $https://github.com/GauravBh1010tt/RewardRank$

Jordan Vice, Naveed Akhtar, Leonid Sigal et al.

The rapid proliferation of multimodal generative models has sparked critical discussions on their reliability, fairness and potential for misuse. While text-to-image models excel at producing high-fidelity, user-guided content, they often exhibit unpredictable behaviors and vulnerabilities that can be exploited to manipulate class or concept representations. To address this, we propose an evaluation framework to assess model reliability by analyzing responses to global and local perturbations in the embedding space, enabling the identification of inputs that trigger unreliable or biased behavior. Beyond social implications, fairness and diversity are fundamental to defining robust and trustworthy model behavior. Our approach offers deeper insights into these essential aspects by evaluating: (i) generative diversity, measuring the breadth of visual representations for learned concepts, and (ii) generative fairness, which examines the impact that removing concepts from input prompts has on control, under a low guidance setup. Beyond these evaluations, our method lays the groundwork for detecting unreliable, bias-injected models and tracing the provenance of embedded biases. Our code is publicly available at https://github.com/JJ-Vice/T2I_Fairness_Diversity_Reliability. Keywords: Fairness, Reliability, AI Ethics, Bias, Text-to-Image Models

Weidong Yin, Yanwei Fu, Leonid Sigal et al.

Generating and manipulating human facial images using high-level attributal controls are important and interesting problems. The models proposed in previous work can solve one of these two problems (generation or manipulation), but not both coherently. This paper proposes a novel model that learns how to both generate and modify the facial image from high-level semantic attributes. Our key idea is to formulate a Semi-Latent Facial Attribute Space (SL-FAS) to systematically learn relationship between user-defined and latent attributes, as well as between those attributes and RGB imagery. As part of this newly formulated space, we propose a new model --- SL-GAN which is a specific form of Generative Adversarial Network. Finally, we present an iterative training algorithm for SL-GAN. The experiments on recent CelebA and CASIA-WebFace datasets validate the effectiveness of our proposed framework. We will also make data, pre-trained models and code available.

Gaurav Bhatt, Aditya Chinchure, Jiawei Zhou et al.

Fine-tuning is the primary mechanism for adapting foundation models to downstream tasks; however, standard approaches largely optimize task objectives in isolation and do not account for secondary yet critical alignment objectives (e.g., safety and hallucination avoidance). As a result, downstream fine-tuning can degrade alignment and fail to correct pre-existing misaligned behavior. We propose an alignment-aware fine-tuning framework that integrates feedback from an external alignment signal through policy-gradient-based regularization. Our method introduces an adaptive gating mechanism that dynamically balances supervised and alignment-driven gradients on a per-sample basis, prioritizing uncertain or misaligned cases while allowing well-aligned examples to follow standard supervised updates. The framework further learns abstention behavior for fully misaligned inputs, incorporating conservative responses directly into the fine-tuned model. Experiments on general and domain-specific instruction-tuning benchmarks demonstrate consistent reductions in harmful and hallucinated outputs without sacrificing downstream task performance. Additional analyses show robustness to adversarial fine-tuning, prompt-based attacks, and unsafe initializations, establishing adaptively gated alignment optimization as an effective approach for alignment-preserving and alignment-recovering model adaptation.

Ali Salamatian, Anthony Fuller, Pritam Sarkar et al.

Transformers dominate video recognition. They split videos into tokens, and processing them has expensive superlinear computational cost. Yet videos are filled with redundancy, so we can question the need for this expense. We introduce LookWhen, a selector-extractor framework that factorizes video recognition into learning when, where, and what to compute. Our shallow selector gets a scaled-down video and quickly scores all tokens across space-time, while our deep extractor gets the top-K selected tokens to approximate full-video representations without actually processing all the tokens. A key challenge is defining effective supervision for selection and extraction. For selection pre-training, we introduce a score on representations that ranks tokens by uniqueness using a simple nearest-neighbor distance. For extraction pre-training, we distill both a video teacher and an image teacher, for which we normalize its frame-wise representations to learn what changes within videos. Through these strategies, our selector-extractor learns general and efficient representations for feature extraction or fine-tuning to a task. Through experiments on Kinetics-400, SSv2, Epic-Kitchens, Diving48, Jester, and Charades, we show that LookWhen achieves a better accuracy-computation trade-off than efficient models and upgraded baselines of similar size. LookWhen Pareto-dominates in accuracy-FLOPs on 9 of 12 cases (6 tasks x 2 settings) and roughly matches on 3. In accuracy-throughput, measuring time in practice, LookWhen is more efficient still at 6.7x faster than InternVideo2-B at equal accuracy.

Mir Rayat Imtiaz Hossain, Leonid Sigal, James J. Little

Recent advances in pixel-level tasks (e.g. segmentation) illustrate the benefit of of long-range interactions between aggregated region-based representations that can enhance local features. However, such aggregated representations, often in the form of attention, fail to model the underlying semantics of the scene (e.g. individual objects and, by extension, their interactions). In this work, we address the issue by proposing a component that learns to project image features into latent representations and reason between them using a transformer encoder to generate contextualized and scene-consistent representations which are fused with original image features. Our design encourages the latent regions to represent semantic concepts by ensuring that the activated regions are spatially disjoint and the union of such regions corresponds to a connected object segment. The proposed semantic global reasoning (SGR) component is end-to-end trainable and can be easily added to a wide variety of backbones (CNN or transformer-based) and segmentation heads (per-pixel or mask classification) to consistently improve the segmentation results on different datasets. In addition, our latent tokens are semantically interpretable and diverse and provide a rich set of features that can be transferred to downstream tasks like object detection and segmentation, with improved performance. Furthermore, we also proposed metrics to quantify the semantics of latent tokens at both class \& instance level.

Shweta Mahajan, Tanzila Rahman, Kwang Moo Yi et al.

The quality of the prompts provided to text-to-image diffusion models determines how faithful the generated content is to the user's intent, often requiring `prompt engineering'. To harness visual concepts from target images without prompt engineering, current approaches largely rely on embedding inversion by optimizing and then mapping them to pseudo-tokens. However, working with such high-dimensional vector representations is challenging because they lack semantics and interpretability, and only allow simple vector operations when using them. Instead, this work focuses on inverting the diffusion model to obtain interpretable language prompts directly. The challenge of doing this lies in the fact that the resulting optimization problem is fundamentally discrete and the space of prompts is exponentially large; this makes using standard optimization techniques, such as stochastic gradient descent, difficult. To this end, we utilize a delayed projection scheme to optimize for prompts representative of the vocabulary space in the model. Further, we leverage the findings that different timesteps of the diffusion process cater to different levels of detail in an image. The later, noisy, timesteps of the forward diffusion process correspond to the semantic information, and therefore, prompt inversion in this range provides tokens representative of the image semantics. We show that our approach can identify semantically interpretable and meaningful prompts for a target image which can be used to synthesize diverse images with similar content. We further illustrate the application of the optimized prompts in evolutionary image generation and concept removal.

Mir Rayat Imtiaz Hossain, Mennatullah Siam, Leonid Sigal et al.

The emergence of attention-based transformer models has led to their extensive use in various tasks, due to their superior generalization and transfer properties. Recent research has demonstrated that such models, when prompted appropriately, are excellent for few-shot inference. However, such techniques are under-explored for dense prediction tasks like semantic segmentation. In this work, we examine the effectiveness of prompting a transformer-decoder with learned visual prompts for the generalized few-shot segmentation (GFSS) task. Our goal is to achieve strong performance not only on novel categories with limited examples, but also to retain performance on base categories. We propose an approach to learn visual prompts with limited examples. These learned visual prompts are used to prompt a multiscale transformer decoder to facilitate accurate dense predictions. Additionally, we introduce a unidirectional causal attention mechanism between the novel prompts, learned with limited examples, and the base prompts, learned with abundant data. This mechanism enriches the novel prompts without deteriorating the base class performance. Overall, this form of prompting helps us achieve state-of-the-art performance for GFSS on two different benchmark datasets: COCO-$20^i$ and Pascal-$5^i$, without the need for test-time optimization (or transduction). Furthermore, test-time optimization leveraging unlabelled test data can be used to improve the prompts, which we refer to as transductive prompt tuning.

Gaurav Bhatt, James Ross, Leonid Sigal

Modern pre-trained architectures struggle to retain previous information while undergoing continuous fine-tuning on new tasks. Despite notable progress in continual classification, systems designed for complex vision tasks such as detection or segmentation still struggle to attain satisfactory performance. In this work, we introduce a memory-based detection transformer architecture to adapt a pre-trained DETR-style detector to new tasks while preserving knowledge from previous tasks. We propose a novel localized query function for efficient information retrieval from memory units, aiming to minimize forgetting. Furthermore, we identify a fundamental challenge in continual detection referred to as background relegation. This arises when object categories from earlier tasks reappear in future tasks, potentially without labels, leading them to be implicitly treated as background. This is an inevitable issue in continual detection or segmentation. The introduced continual optimization technique effectively tackles this challenge. Finally, we assess the performance of our proposed system on continual detection benchmarks and demonstrate that our approach surpasses the performance of existing state-of-the-art resulting in 5-7% improvements on MS-COCO and PASCAL-VOC on the task of continual detection.

Yixin Cheng, Hongcheng Guo, Yangming Li et al.

Text watermarking aims to subtly embed statistical signals into text by controlling the Large Language Model (LLM)'s sampling process, enabling watermark detectors to verify that the output was generated by the specified model. The robustness of these watermarking algorithms has become a key factor in evaluating their effectiveness. Current text watermarking algorithms embed watermarks in high-entropy tokens to ensure text quality. In this paper, we reveal that this seemingly benign design can be exploited by attackers, posing a significant risk to the robustness of the watermark. We introduce a generic efficient paraphrasing attack, the Self-Information Rewrite Attack (SIRA), which leverages the vulnerability by calculating the self-information of each token to identify potential pattern tokens and perform targeted attack. Our work exposes a widely prevalent vulnerability in current watermarking algorithms. The experimental results show SIRA achieves nearly 100% attack success rates on seven recent watermarking methods with only 0.88 USD per million tokens cost. Our approach does not require any access to the watermark algorithms or the watermarked LLM and can seamlessly transfer to any LLM as the attack model, even mobile-level models. Our findings highlight the urgent need for more robust watermarking.

Raghav Goyal, Wan-Cyuan Fan, Mennatullah Siam et al.

Video Object Segmentation (VOS) has emerged as an increasingly important problem with availability of larger datasets and more complex and realistic settings, which involve long videos with global motion (e.g, in egocentric settings), depicting small objects undergoing both rigid and non-rigid (including state) deformations. While a number of recent approaches have been explored for this task, these data characteristics still present challenges. In this work we propose a novel, clip-based DETR-style encoder-decoder architecture, which focuses on systematically analyzing and addressing aforementioned challenges. Specifically, we propose a novel transformation-aware loss that focuses learning on portions of the video where an object undergoes significant deformations -- a form of "soft" hard examples mining. Further, we propose a multiplicative time-coded memory, beyond vanilla additive positional encoding, which helps propagate context across long videos. Finally, we incorporate these in our proposed holistic multi-scale video transformer for tracking via multi-scale memory matching and decoding to ensure sensitivity and accuracy for long videos and small objects. Our model enables on-line inference with long videos in a windowed fashion, by breaking the video into clips and propagating context among them. We illustrate that short clip length and longer memory with learned time-coding are important design choices for improved performance. Collectively, these technical contributions enable our model to achieve new state-of-the-art (SoTA) performance on two complex egocentric datasets -- VISOR and VOST, while achieving comparable to SoTA results on the conventional VOS benchmark, DAVIS'17. A series of detailed ablations validate our design choices as well as provide insights into the importance of parameter choices and their impact on performance.

Wan-Cyuan Fan, Tanzila Rahman, Leonid Sigal

With advances in foundational and vision-language models, and effective fine-tuning techniques, a large number of both general and special-purpose models have been developed for a variety of visual tasks. Despite the flexibility and accessibility of these models, no single model is able to handle all tasks and/or applications that may be envisioned by potential users. Recent approaches, such as visual programming and multimodal LLMs with integrated tools aim to tackle complex visual tasks, by way of program synthesis. However, such approaches overlook user constraints (e.g., performance / computational needs), produce test-time sample-specific solutions that are difficult to deploy, and, sometimes, require low-level instructions that maybe beyond the abilities of a naive user. To address these limitations, we introduce MMFactory, a universal framework that includes model and metrics routing components, acting like a solution search engine across various available models. Based on a task description and few sample input-output pairs and (optionally) resource and/or performance constraints, MMFactory can suggest a diverse pool of programmatic solutions by instantiating and combining visio-lingual tools from its model repository. In addition to synthesizing these solutions, MMFactory also proposes metrics and benchmarks performance / resource characteristics, allowing users to pick a solution that meets their unique design constraints. From the technical perspective, we also introduced a committee-based solution proposer that leverages multi-agent LLM conversation to generate executable, diverse, universal, and robust solutions for the user. Experimental results show that MMFactory outperforms existing methods by delivering state-of-the-art solutions tailored to user problem specifications. Project page is available at https://davidhalladay.github.io/mmfactory_demo.

Jing Wang, Wonho Bae, Jiahong Chen et al.

Source-free domain adaptation (SFDA) involves adapting a model originally trained using a labeled dataset ({\em source domain}) to perform effectively on an unlabeled dataset ({\em target domain}) without relying on any source data during adaptation. This adaptation is especially crucial when significant disparities in data distributions exist between the two domains and when there are privacy concerns regarding the source model's training data. The absence of access to source data during adaptation makes it challenging to analytically estimate the domain gap. To tackle this issue, various techniques have been proposed, such as unsupervised clustering, contrastive learning, and continual learning. In this paper, we first conduct an extensive theoretical analysis of SFDA based on contrastive learning, primarily because it has demonstrated superior performance compared to other techniques. Motivated by the obtained insights, we then introduce a straightforward yet highly effective latent augmentation method tailored for contrastive SFDA. This augmentation method leverages the dispersion of latent features within the neighborhood of the query sample, guided by the source pre-trained model, to enhance the informativeness of positive keys. Our approach, based on a single InfoNCE-based contrastive loss, outperforms state-of-the-art SFDA methods on widely recognized benchmark datasets.

Tanzila Rahman, Shweta Mahajan, Hsin-Ying Lee et al.

Text-to-image (TTI) diffusion models have demonstrated impressive results in generating high-resolution images of complex and imaginative scenes. Recent approaches have further extended these methods with personalization techniques that allow them to integrate user-illustrated concepts (e.g., the user him/herself) using a few sample image illustrations. However, the ability to generate images with multiple interacting concepts, such as human subjects, as well as concepts that may be entangled in one, or across multiple, image illustrations remains illusive. In this work, we propose a concept-driven TTI personalization framework that addresses these core challenges. We build on existing works that learn custom tokens for user-illustrated concepts, allowing those to interact with existing text tokens in the TTI model. However, importantly, to disentangle and better learn the concepts in question, we jointly learn (latent) segmentation masks that disentangle these concepts in user-provided image illustrations. We do so by introducing an Expectation Maximization (EM)-like optimization procedure where we alternate between learning the custom tokens and estimating (latent) masks encompassing corresponding concepts in user-supplied images. We obtain these masks based on cross-attention, from within the U-Net parameterized latent diffusion model and subsequent DenseCRF optimization. We illustrate that such joint alternating refinement leads to the learning of better tokens for concepts and, as a by-product, latent masks. We illustrate the benefits of the proposed approach qualitatively and quantitatively with several examples and use cases that can combine three or more entangled concepts.

Mir Rayat Imtiaz Hossain, Mennatullah Siam, Leonid Sigal et al.

Large-scale vision-language models (VLMs), trained on extensive datasets of image-text pairs, exhibit strong multimodal understanding capabilities by implicitly learning associations between textual descriptions and image regions. This emergent ability enables zero-shot object detection and segmentation, using techniques that rely on text-image attention maps, without necessarily training on abundant labeled segmentation datasets. However, performance of such methods depends heavily on prompt engineering and manually selected layers or head choices for the attention layers. In this work, we demonstrate that, rather than relying solely on textual prompts, providing a single visual example for each category and fine-tuning the text-to-image attention layers and embeddings significantly improves the performance. Additionally, we propose learning an ensemble through few-shot fine-tuning across multiple layers and/or prompts. An entropy-based ranking and selection mechanism for text-to-image attention layers is proposed to identify the top-performing layers without the need for segmentation labels. This eliminates the need for hyper-parameter selection of text-to-image attention layers, providing a more flexible and scalable solution for open-vocabulary segmentation. We show that this approach yields strong zero-shot performance, further enhanced through fine-tuning with a single visual example. Moreover, we demonstrate that our method and findings are general and can be applied across various vision-language models (VLMs).

Nitesh Bharadwaj Gundavarapu, Luke Friedman, Raghav Goyal et al.

Video understanding has witnessed significant progress with recent video foundation models demonstrating strong performance owing to self-supervised pre-training objectives; Masked Autoencoders (MAE) being the design of choice. Nevertheless, the majority of prior works that leverage MAE pre-training have focused on relatively short video representations (16 / 32 frames in length) largely due to hardware memory and compute limitations that scale poorly with video length due to the dense memory-intensive self-attention decoding. One natural strategy to address these challenges is to subsample tokens to reconstruct during decoding (or decoder masking). In this work, we propose an effective strategy for prioritizing tokens which allows training on longer video sequences (128 frames) and gets better performance than, more typical, random and uniform masking strategies. The core of our approach is an adaptive decoder masking strategy that prioritizes the most important tokens and uses quantized tokens as reconstruction objectives. Our adaptive strategy leverages a powerful MAGVIT-based tokenizer that jointly learns the tokens and their priority. We validate our design choices through exhaustive ablations and observe improved performance of the resulting long-video (128 frames) encoders over short-video (32 frames) counterparts. With our long-video masked autoencoder (LVMAE) strategy, we surpass state-of-the-art on Diving48 by 3.9 points and EPIC-Kitchens-100 verb classification by 2.5 points while relying on a simple core architecture and video-only pre-training (unlike some of the prior works that require millions of labeled video-text pairs or specialized encoders).

Bicheng Xu, Qi Yan, Renjie Liao et al.

We introduce a framework for joint grounded scene graph - image generation, a challenging task involving high-dimensional, multi-modal structured data. To effectively model this complex joint distribution, we adopt a factorized approach: first generating a grounded scene graph, followed by image generation conditioned on the generated grounded scene graph. While conditional image generation has been widely explored in the literature, our primary focus is on the generation of grounded scene graphs from noise, which provides efficient and interpretable control over the image generation process. This task requires generating plausible grounded scene graphs with heterogeneous attributes for both nodes (objects) and edges (relations among objects), encompassing continuous attributes (e.g., object bounding boxes) and discrete attributes (e.g., object and relation categories). To address this challenge, we introduce DiffuseSG, a novel diffusion model that jointly models the heterogeneous node and edge attributes. We explore different encoding strategies to effectively handle the categorical data. Leveraging a graph transformer as the denoiser, DiffuseSG progressively refines grounded scene graph representations in a continuous space before discretizing them to generate structured outputs. Additionally, we introduce an IoU-based regularization term to enhance empirical performance. Our model outperforms existing methods in grounded scene graph generation on the VG and COCO-Stuff datasets, excelling in both standard and newly introduced metrics that more accurately capture the task's complexity. Furthermore, we demonstrate the broader applicability of DiffuseSG in two important downstream tasks: 1) achieving superior results in a range of grounded scene graph completion tasks, and 2) enhancing grounded scene graph detection models by leveraging additional training samples generated by DiffuseSG.

Shivam Chandhok, Qian Yang, Oscar Manas et al.

Instruction tuning has been central to the success of recent vision-language models (VLMs), but it remains expensive-requiring large-scale datasets, high-quality annotations, and large compute budgets. We propose PRioritized cOncept learninG via Relative Error-driven Sample Selection (PROGRESS), a data- and compute-efficient framework that enables VLMs to dynamically select what to learn next based on their evolving needs during training. At each stage, the model tracks its learning progress across skills and selects the most informative samples-those it has not already mastered and that are not too difficult to learn at the current stage of training. This strategy effectively controls skill acquisition and the order in which skills are learned. Specifically, we sample from skills showing the highest learning progress, prioritizing those with the most rapid improvement. Unlike prior methods, PROGRESS requires no upfront answer annotations, queries answers only on a need basis, avoids reliance on additional supervision from auxiliary VLMs, and does not require compute-heavy gradient computations for data selection. Experiments across multiple instruction-tuning datasets of varying scales demonstrate that PROGRESS consistently outperforms state-of-the-art baselines with much less data and supervision. Additionally, we show strong cross-architecture generalization and transferability to larger models, validating PROGRESS as a scalable solution for efficient learning.

Pushkar Shukla, Aditya Chinchure, Emily Diana et al.

The biases exhibited by Text-to-Image (TTI) models are often treated as if they are independent, but in reality, they may be deeply interrelated. Addressing bias along one dimension, such as ethnicity or age, can inadvertently influence another dimension, like gender, either mitigating or exacerbating existing disparities. Understanding these interdependencies is crucial for designing fairer generative models, yet measuring such effects quantitatively remains a challenge. In this paper, we aim to address these questions by introducing BiasConnect, a novel tool designed to analyze and quantify bias interactions in TTI models. Our approach leverages a counterfactual-based framework to generate pairwise causal graphs that reveals the underlying structure of bias interactions for the given text prompt. Additionally, our method provides empirical estimates that indicate how other bias dimensions shift toward or away from an ideal distribution when a given bias is modified. Our estimates have a strong correlation (+0.69) with the interdependency observations post bias mitigation. We demonstrate the utility of BiasConnect for selecting optimal bias mitigation axes, comparing different TTI models on the dependencies they learn, and understanding the amplification of intersectional societal biases in TTI models.

Shih-Han Chou, Matthew Kowal, Yasmin Niknam et al.

While progress has been made in the domain of video-language understanding, current state-of-the-art algorithms are still limited in their ability to understand videos at high levels of abstraction, such as news-oriented videos. Alternatively, humans easily amalgamate information from video and language to infer information beyond what is visually observable in the pixels. An example of this is watching a news story, where the context of the event can play as big of a role in understanding the story as the event itself. Towards a solution for designing this ability in algorithms, we present a large-scale analysis on an in-house dataset collected by the Reuters News Agency, called Reuters Video-Language News (ReutersViLNews) dataset which focuses on high-level video-language understanding with an emphasis on long-form news. The ReutersViLNews Dataset consists of long-form news videos collected and labeled by news industry professionals over several years and contains prominent news reporting from around the world. Each video involves a single story and contains action shots of the actual event, interviews with people associated with the event, footage from nearby areas, and more. ReutersViLNews dataset contains videos from seven subject categories: disaster, finance, entertainment, health, politics, sports, and miscellaneous with annotations from high-level to low-level, title caption, visual video description, high-level story description, keywords, and location. We first present an analysis of the dataset statistics of ReutersViLNews compared to previous datasets. Then we benchmark state-of-the-art approaches for four different video-language tasks. The results suggest that news-oriented videos are a substantial challenge for current video-language understanding algorithms and we conclude by providing future directions in designing approaches to solve the ReutersViLNews dataset.

Apratim Bhattacharyya, Bicheng Xu, Sanjay Haresh et al.

Multi-modal Large Language Models (LLM) have advanced conversational abilities but struggle with providing live, interactive step-by-step guidance, a key capability for future AI assistants. Effective guidance requires not only delivering instructions but also detecting their successful execution, as well as identifying and alerting users to mistakes, all of which has to happen in real-time. This requires models that are not turn-based, but that can react asynchronously to a video stream, as well as video data showing users performing tasks including mistakes and their corrections. To this end, we introduce Qualcomm Interactive Cooking, a new benchmark and dataset built upon CaptainCook4D, which contains user mistakes during task execution. Our dataset and benchmark features densely annotated, timed instructions and feedback messages, specifically including mistake alerts precisely timestamped to their visual occurrence in the video. We evaluate state-of-the-art multi-modal LLMs on the Qualcomm Interactive Cooking benchmark and introduce LiveMamba, a streaming multi-modal LLM designed for interactive instructional guidance. This work provides the first dedicated benchmark and a strong baseline for developing and evaluating on live, situated coaching.

Aditya Chinchure, Sahithya Ravi, Pushkar Shukla et al.

Current text-to-video models (T2V) can generate high-quality, temporally coherent, and visually realistic videos. Nonetheless, errors still often occur, and are more nuanced and local compared to the previous generation of T2V models. While current evaluation paradigms assess video models across diverse dimensions, they typically evaluate videos holistically without identifying when specific errors occur or describing their nature. We address this gap by introducing Spotlight, a novel task aimed at localizing and explaining video-generation errors. We generate 600 videos using 200 diverse textual prompts and three state-of-the-art video generators (Veo 3, Seedance, and LTX-2), and annotate over 1600 fine-grained errors across six types, including motion, physics, and prompt adherence. We observe that adherence and physics errors are predominant and persist across longer segments, whereas appearance-disappearance and body pose errors manifest in shorter segments. We then evaluate current VLMs on Spotlight and find that VLMs lag significantly behind humans in error identification and localization in videos. We propose inference-time strategies to probe the limits of current VLMs on our task, improving performance by nearly 2x. Our task paves a way forward to building fine-grained evaluation tools and more sophisticated reward models for video generators.