Sangho Lee

Christopher Clark, Jieyu Zhang, Zixian Ma et al. · gatech

Today's strongest video-language models (VLMs) remain proprietary. The strongest open-weight models either rely on synthetic data from proprietary VLMs, effectively distilling from them, or do not disclose their training data or recipe. As a result, the open-source community lacks the foundations needed to improve on the state-of-the-art video (and image) language models. Crucially, many downstream applications require more than just high-level video understanding; they require grounding -- either by pointing or by tracking in pixels. Even proprietary models lack this capability. We present Molmo2, a new family of VLMs that are state-of-the-art among open-source models and demonstrate exceptional new capabilities in point-driven grounding in single image, multi-image, and video tasks. Our key contribution is a collection of 7 new video datasets and 2 multi-image datasets, including a dataset of highly detailed video captions for pre-training, a free-form video Q&A dataset for fine-tuning, a new object tracking dataset with complex queries, and an innovative new video pointing dataset, all collected without the use of closed VLMs. We also present a training recipe for this data utilizing an efficient packing and message-tree encoding scheme, and show bi-directional attention on vision tokens and a novel token-weight strategy improves performance. Our best-in-class 8B model outperforms others in the class of open weight and data models on short videos, counting, and captioning, and is competitive on long-videos. On video-grounding Molmo2 significantly outperforms existing open-weight models like Qwen3-VL (35.5 vs 29.6 accuracy on video counting) and surpasses proprietary models like Gemini 3 Pro on some tasks (38.4 vs 20.0 F1 on video pointing and 56.2 vs 41.1 J&F on video tracking).

Matt Deitke, Christopher Clark, Sangho Lee et al. · allen-ai

Today's most advanced vision-language models (VLMs) remain proprietary. The strongest open-weight models rely heavily on synthetic data from proprietary VLMs to achieve good performance, effectively distilling these closed VLMs into open ones. As a result, the community has been missing foundational knowledge about how to build performant VLMs from scratch. We present Molmo, a new family of VLMs that are state-of-the-art in their class of openness. Our key contribution is a collection of new datasets called PixMo, including a dataset of highly detailed image captions for pre-training, a free-form image Q&A dataset for fine-tuning, and an innovative 2D pointing dataset, all collected without the use of external VLMs. The success of our approach relies on careful modeling choices, a well-tuned training pipeline, and, most critically, the quality of our newly collected datasets. Our best-in-class 72B model not only outperforms others in the class of open weight and data models, but also outperforms larger proprietary models including Claude 3.5 Sonnet, and Gemini 1.5 Pro and Flash, second only to GPT-4o based on both academic benchmarks and on a large human evaluation. Our model weights, new datasets, and source code are available at https://molmo.allenai.org/blog.

Dayoon Ko, Sangho Lee, Gunhee Kim · allen-ai

As short-form funny videos on social networks are gaining popularity, it becomes demanding for AI models to understand them for better communication with humans. Unfortunately, previous video humor datasets target specific domains, such as speeches or sitcoms, and mostly focus on verbal cues. We curate a user-generated dataset of 10K multimodal funny videos from YouTube, called ExFunTube. Using a video filtering pipeline with GPT-3.5, we verify both verbal and visual elements contributing to humor. After filtering, we annotate each video with timestamps and text explanations for funny moments. Our ExFunTube is unique over existing datasets in that our videos cover a wide range of domains with various types of humor that necessitate a multimodal understanding of the content. Also, we develop a zero-shot video-to-text prompting to maximize video humor understanding of large language models (LLMs). With three different evaluation methods using automatic scores, rationale quality experiments, and human evaluations, we show that our prompting significantly improves LLMs' ability for humor explanation.

Jitesh Jain, Jialuo Li, Zixian Ma et al. · gatech

As humans, we are natural any-horizon reasoners, i.e., we can decide whether to iteratively skim long videos or watch short ones in full when necessary for a given task. With this in mind, one would expect video reasoning models to reason flexibly across different durations. However, SOTA models are still trained to predict answers in a single turn while processing a large number of frames, akin to watching an entire long video, requiring significant resources. This raises the question: Is it possible to develop performant any-horizon video reasoning systems? Inspired by human behavior, we first propose SAGE, an agent system that performs multi-turn reasoning on long videos while handling simpler problems in a single turn. Secondly, we introduce an easy synthetic data generation pipeline using Gemini-2.5-Flash to train the orchestrator, SAGE-MM, which lies at the core of SAGE. We further propose an effective RL post-training recipe essential for instilling any-horizon reasoning ability in SAGE-MM. Thirdly, we curate SAGE-Bench with an average duration of greater than 700 seconds for evaluating video reasoning ability in real-world entertainment use cases. Lastly, we empirically validate the effectiveness of our system, data, and RL recipe, observing notable improvements of up to 6.1% on open-ended video reasoning tasks, as well as an impressive 8.2% improvement on videos longer than 10 minutes.

Christopher Clark, Yue Yang, Jae Sung Park et al.

Grounding has become a fundamental capability of vision-language models (VLMs). Most existing VLMs point by generating coordinates as part of their text output, which requires learning a complicated coordinate system and results in a high token count. Instead, we propose a more intuitive pointing mechanism that directly selects the visual tokens that contain the target concept. Our model generates a special pointing token that cross-attends to the input image or video tokens and selects the appropriate one. To make this model more fine-grained, we follow these pointing tokens with an additional special token that selects a fine-grained subpatch within the initially selected region, and then a third token that specifies a location within that subpatch. We further show that performance improves by generating points sequentially in a consistent order, encoding the relative position of the previously selected point, and including a special no-more-points class when selecting visual tokens. Using this method, we set a new state-of-the-art on image pointing (70.7% on PointBench), set a new state-of-the-art among fully open models on GUI pointing (61.1% on ScreenSpotPro), and improve video pointing (59.1% human preference win rate vs. a text coordinate baseline) and tracking (+6.3% gain on Molmo2Track). We additionally show that our method achieves much higher sample efficiency and discuss the qualitative differences that emerge from this design change.

Jianrui Zhang, Yue Yang, Rohun Tripathi et al.

Token pruning is essential for enhancing the computational efficiency of vision-language models (VLMs), particularly for video-based tasks where temporal redundancy is prevalent. Prior approaches typically prune tokens either (1) within the vision transformer (ViT) exclusively for unimodal perception tasks such as action recognition and object segmentation, without adapting to downstream vision-language tasks; or (2) only within the LLM while leaving the ViT output intact, often requiring complex text-conditioned token selection mechanisms. In this paper, we introduce Spatio-Temporal Token Scoring (STTS), a simple and lightweight module that prunes vision tokens across both the ViT and the LLM without text conditioning or token merging, and is fully compatible with end-to-end training. By learning how to score temporally via an auxiliary loss and spatially via LLM downstream gradients, aided by our efficient packing algorithm, STTS prunes 50% of vision tokens throughout the entire architecture, resulting in a 62% improvement in efficiency during both training and inference with only a 0.7% drop in average performance across 13 short and long video QA tasks. Efficiency gains increase with more sampled frames per video. Applying test-time scaling for long-video QA further yields performance gains of 0.5-1% compared to the baseline. Overall, STTS represents a novel, simple yet effective technique for unified, architecture-wide vision token pruning.

Duong H. Le, Tuan Pham, Sangho Lee et al. · allen-ai

We introduce OneDiffusion, a versatile, large-scale diffusion model that seamlessly supports bidirectional image synthesis and understanding across diverse tasks. It enables conditional generation from inputs such as text, depth, pose, layout, and semantic maps, while also handling tasks like image deblurring, upscaling, and reverse processes such as depth estimation and segmentation. Additionally, OneDiffusion allows for multi-view generation, camera pose estimation, and instant personalization using sequential image inputs. Our model takes a straightforward yet effective approach by treating all tasks as frame sequences with varying noise scales during training, allowing any frame to act as a conditioning image at inference time. Our unified training framework removes the need for specialized architectures, supports scalable multi-task training, and adapts smoothly to any resolution, enhancing both generalization and scalability. Experimental results demonstrate competitive performance across tasks in both generation and prediction such as text-to-image, multiview generation, ID preservation, depth estimation and camera pose estimation despite relatively small training dataset. Our code and checkpoint are freely available at https://github.com/lehduong/OneDiffusion

Jiasen Lu, Christopher Clark, Sangho Lee et al. · allen-ai

We present Unified-IO 2, the first autoregressive multimodal model that is capable of understanding and generating image, text, audio, and action. To unify different modalities, we tokenize inputs and outputs -- images, text, audio, action, bounding boxes, etc., into a shared semantic space and then process them with a single encoder-decoder transformer model. Since training with such diverse modalities is challenging, we propose various architectural improvements to stabilize model training. We train our model from scratch on a large multimodal pre-training corpus from diverse sources with a multimodal mixture of denoisers objective. To learn an expansive set of skills, such as following multimodal instructions, we construct and finetune on an ensemble of 120 datasets with prompts and augmentations. With a single unified model, Unified-IO 2 achieves state-of-the-art performance on the GRIT benchmark and strong results in more than 35 benchmarks, including image generation and understanding, natural language understanding, video and audio understanding, and robotic manipulation. We release all our models to the research community.

Chris Dongjoo Kim, Jihwan Moon, Sangwoo Moon et al. · allen-ai

The rapid growth of video-text data presents challenges in storage and computation during training. Online learning, which processes streaming data in real-time, offers a promising solution to these issues while also allowing swift adaptations in scenarios demanding real-time responsiveness. One strategy to enhance the efficiency and effectiveness of learning involves identifying and prioritizing data that enhances performance on target downstream tasks. We propose Relevance and Specificity-based online filtering framework (ReSpec) that selects data based on four criteria: (i) modality alignment for clean data, (ii) task relevance for target focused data, (iii) specificity for informative and detailed data, and (iv) efficiency for low-latency processing. Relevance is determined by the probabilistic alignment of incoming data with downstream tasks, while specificity employs the distance to a root embedding representing the least specific data as an efficient proxy for informativeness. By establishing reference points from target task data, ReSpec filters incoming data in real-time, eliminating the need for extensive storage and compute. Evaluating on large-scale datasets WebVid2M and VideoCC3M, ReSpec attains state-of-the-art performance on five zeroshot video retrieval tasks, using as little as 5% of the data while incurring minimal compute. The source code is available at https://github.com/cdjkim/ReSpec.

Jonghwan Mun, Minchul Shin, Gunsoo Han et al.

Self-supervised learning has drawn attention through its effectiveness in learning in-domain representations with no ground-truth annotations; in particular, it is shown that properly designed pretext tasks (e.g., contrastive prediction task) bring significant performance gains for downstream tasks (e.g., classification task). Inspired from this, we tackle video scene segmentation, which is a task of temporally localizing scene boundaries in a video, with a self-supervised learning framework where we mainly focus on designing effective pretext tasks. In our framework, we discover a pseudo-boundary from a sequence of shots by splitting it into two continuous, non-overlapping sub-sequences and leverage the pseudo-boundary to facilitate the pre-training. Based on this, we introduce three novel boundary-aware pretext tasks: 1) Shot-Scene Matching (SSM), 2) Contextual Group Matching (CGM) and 3) Pseudo-boundary Prediction (PP); SSM and CGM guide the model to maximize intra-scene similarity and inter-scene discrimination while PP encourages the model to identify transitional moments. Through comprehensive analysis, we empirically show that pre-training and transferring contextual representation are both critical to improving the video scene segmentation performance. Lastly, we achieve the new state-of-the-art on the MovieNet-SSeg benchmark. The code is available at https://github.com/kakaobrain/bassl.

Yookoon Park, Sangho Lee, Gunhee Kim et al.

We present neural activation coding (NAC) as a novel approach for learning deep representations from unlabeled data for downstream applications. We argue that the deep encoder should maximize its nonlinear expressivity on the data for downstream predictors to take full advantage of its representation power. To this end, NAC maximizes the mutual information between activation patterns of the encoder and the data over a noisy communication channel. We show that learning for a noise-robust activation code increases the number of distinct linear regions of ReLU encoders, hence the maximum nonlinear expressivity. More interestingly, NAC learns both continuous and discrete representations of data, which we respectively evaluate on two downstream tasks: (i) linear classification on CIFAR-10 and ImageNet-1K and (ii) nearest neighbor retrieval on CIFAR-10 and FLICKR-25K. Empirical results show that NAC attains better or comparable performance on both tasks over recent baselines including SimCLR and DistillHash. In addition, NAC pretraining provides significant benefits to the training of deep generative models. Our code is available at https://github.com/yookoon/nac.

Soyeon Park, Sangho Lee, Wen Xu et al.

Intel memory protection keys (MPK) is a new hardware feature to support thread-local permission control on groups of pages without requiring modification of page tables. Unfortunately, its current hardware implementation and software supports suffer from security, scalability, and semantic-gap problems: (1) MPK is vulnerable to protection-key-use-after-free and protection-key corruption; (2) MPK does not scale due to hardware limitations; and (3) MPK is not perfectly compatible with mprotect() because it does not support permission synchronization across threads. In this paper, we propose libmpk, a software abstraction for MPK. libmpk virtualizes protection keys to eliminate the protection-key-use-after-free and protection-key corruption problems while supporting a tremendous number of memory page groups. libmpk also prevents unauthorized writes to its metadata and supports inter-thread key synchronization. We apply libmpk to three real-world applications: OpenSSL, JavaScript JIT compiler, and Memcached for memory protection and isolation. An evaluation shows that libmpk introduces negligible performance overhead (<1%) compared with insecure versions, and improves their performance by 8.1x over secure equivalents using mprotect(). The source code of libmpk will be publicly available and maintained as an open source project.

Seongsu Ha, Chaeyun Kim, Donghwa Kim et al.

Referring Image Segmentation is a comprehensive task to segment an object referred by a textual query from an image. In nature, the level of difficulty in this task is affected by the existence of similar objects and the complexity of the referring expression. Recent RIS models still show a significant performance gap between easy and hard scenarios. We pose that the bottleneck exists in the data, and propose a simple but powerful data augmentation method, Negative-mined Mosaic Augmentation (NeMo). This method augments a training image into a mosaic with three other negative images carefully curated by a pretrained multimodal alignment model, e.g., CLIP, to make the sample more challenging. We discover that it is critical to properly adjust the difficulty level, neither too ambiguous nor too trivial. The augmented training data encourages the RIS model to recognize subtle differences and relationships between similar visual entities and to concretely understand the whole expression to locate the right target better. Our approach shows consistent improvements on various datasets and models, verified by extensive experiments.

Sangho Lee, Il Yong Chun, Hogun Park

Multi-modal transformers are rapidly gaining attention in video captioning tasks. Existing multi-modal video captioning methods typically extract a fixed number of frames, which raises critical challenges. When a limited number of frames are extracted, important frames with essential information for caption generation may be missed. Conversely, extracting an excessive number of frames includes consecutive frames, potentially causing redundancy in visual tokens extracted from consecutive video frames. To extract an appropriate number of frames for each video, this paper proposes the first model-agnostic module selection framework in video captioning that has two main functions: (1) selecting a caption generation module with an appropriate size based on visual tokens extracted from video frames, and (2) constructing subsets of visual tokens for the selected caption generation module. Furthermore, we propose a new adaptive attention masking scheme that enhances attention on important visual tokens. Our experiments on three different benchmark datasets demonstrate that the proposed framework significantly improves the performance of three recent video captioning models.

Sangho Lee, Jiwan Chung, Youngjae Yu et al.

The natural association between visual observations and their corresponding sound provides powerful self-supervisory signals for learning video representations, which makes the ever-growing amount of online videos an attractive source of training data. However, large portions of online videos contain irrelevant audio-visual signals because of edited/overdubbed audio, and models trained on such uncurated videos have shown to learn suboptimal representations. Therefore, existing approaches rely almost exclusively on datasets with predetermined taxonomies of semantic concepts, where there is a high chance of audio-visual correspondence. Unfortunately, constructing such datasets require labor intensive manual annotation and/or verification, which severely limits the utility of online videos for large-scale learning. In this work, we present an automatic dataset curation approach based on subset optimization where the objective is to maximize the mutual information between audio and visual channels in videos. We demonstrate that our approach finds videos with high audio-visual correspondence and show that self-supervised models trained on our data achieve competitive performances compared to models trained on existing manually curated datasets. The most significant benefit of our approach is scalability: We release ACAV100M that contains 100 million videos with high audio-visual correspondence, ideal for self-supervised video representation learning.

Sangho Lee, Youngjae Yu, Gunhee Kim et al.

The recent success of Transformers in the language domain has motivated adapting it to a multimodal setting, where a new visual model is trained in tandem with an already pretrained language model. However, due to the excessive memory requirements from Transformers, existing work typically fixes the language model and train only the vision module, which limits its ability to learn cross-modal information in an end-to-end manner. In this work, we focus on reducing the parameters of multimodal Transformers in the context of audio-visual video representation learning. We alleviate the high memory requirement by sharing the parameters of Transformers across layers and modalities; we decompose the Transformer into modality-specific and modality-shared parts so that the model learns the dynamics of each modality both individually and together, and propose a novel parameter sharing scheme based on low-rank approximation. We show that our approach reduces parameters of the Transformers up to 97$\%$, allowing us to train our model end-to-end from scratch. We also propose a negative sampling approach based on an instance similarity measured on the CNN embedding space that our model learns together with the Transformers. To demonstrate our approach, we pretrain our model on 30-second clips (480 frames) from Kinetics-700 and transfer it to audio-visual classification tasks.

Sangho Lee, Kiyoon Yoo, Nojun Kwak

Federated learning (FL), which utilizes communication between the server (core) and local devices (edges) to indirectly learn from more data, is an emerging field in deep learning research. Recently, Knowledge Distillation-based FL methods with notable performance and high applicability have been suggested. In this paper, we choose knowledge distillation-based FL method as our baseline and tackle a challenging problem that ensues from using these methods. Especially, we focus on the problem incurred in the server model that tries to mimic different datasets, each of which is unique to an individual edge device. We dub the problem 'edge bias', which occurs when multiple teacher models trained on different datasets are used individually to distill knowledge. We introduce this nuisance that occurs in certain scenarios of FL, and to alleviate it, we propose a simple yet effective distillation scheme named 'buffered distillation'. In addition, we also experimentally show that this scheme is effective in mitigating the straggler problem caused by delayed edges.

Sangho Lee, Jinyoung Sung, Youngjae Yu et al.

We address the problem of story-based temporal summarization of long 360° videos. We propose a novel memory network model named Past-Future Memory Network (PFMN), in which we first compute the scores of 81 normal field of view (NFOV) region proposals cropped from the input 360° video, and then recover a latent, collective summary using the network with two external memories that store the embeddings of previously selected subshots and future candidate subshots. Our major contributions are two-fold. First, our work is the first to address story-based temporal summarization of 360° videos. Second, our model is the first attempt to leverage memory networks for video summarization tasks. For evaluation, we perform three sets of experiments. First, we investigate the view selection capability of our model on the Pano2Vid dataset. Second, we evaluate the temporal summarization with a newly collected 360° video dataset. Finally, we experiment our model's performance in another domain, with image-based storytelling VIST dataset. We verify that our model achieves state-of-the-art performance on all the tasks.

Youngjae Yu, Sangho Lee, Joonil Na et al.

We address the problem of highlight detection from a 360 degree video by summarizing it both spatially and temporally. Given a long 360 degree video, we spatially select pleasantly-looking normal field-of-view (NFOV) segments from unlimited field of views (FOV) of the 360 degree video, and temporally summarize it into a concise and informative highlight as a selected subset of subshots. We propose a novel deep ranking model named as Composition View Score (CVS) model, which produces a spherical score map of composition per video segment, and determines which view is suitable for highlight via a sliding window kernel at inference. To evaluate the proposed framework, we perform experiments on the Pano2Vid benchmark dataset and our newly collected 360 degree video highlight dataset from YouTube and Vimeo. Through evaluation using both quantitative summarization metrics and user studies via Amazon Mechanical Turk, we demonstrate that our approach outperforms several state-of-the-art highlight detection methods. We also show that our model is 16 times faster at inference than AutoCam, which is one of the first summarization algorithms of 360 degree videos

Sangho Lee, Taesoo Kim

Intel software guard extensions (SGX) aims to provide an isolated execution environment, known as an enclave, for a user-level process to maximize its confidentiality and integrity. In this paper, we study how uninitialized data inside a secure enclave can be leaked via structure padding. We found that, during ECALL and OCALL, proxy functions that are automatically generated by the Intel SGX Software Development Kit (SDK) fully copy structure variables from an enclave to the normal memory to return the result of an ECALL function and to pass input parameters to an OCALL function. If the structure variables contain padding bytes, uninitialized enclave memory, which might contain confidential data like a private key, can be copied to the normal memory through the padding bytes. We also consider potential countermeasures against these security threats.

Seil Na, Sangho Lee, Jisung Kim et al.

We propose a novel memory network model named Read-Write Memory Network (RWMN) to perform question and answering tasks for large-scale, multimodal movie story understanding. The key focus of our RWMN model is to design the read network and the write network that consist of multiple convolutional layers, which enable memory read and write operations to have high capacity and flexibility. While existing memory-augmented network models treat each memory slot as an independent block, our use of multi-layered CNNs allows the model to read and write sequential memory cells as chunks, which is more reasonable to represent a sequential story because adjacent memory blocks often have strong correlations. For evaluation, we apply our model to all the six tasks of the MovieQA benchmark, and achieve the best accuracies on several tasks, especially on the visual QA task. Our model shows a potential to better understand not only the content in the story, but also more abstract information, such as relationships between characters and the reasons for their actions.

Seil Na, Youngjae Yu, Sangho Lee et al.

YouTube-8M is the largest video dataset for multi-label video classification. In order to tackle the multi-label classification on this challenging dataset, it is necessary to solve several issues such as temporal modeling of videos, label imbalances, and correlations between labels. We develop a deep neural network model, which consists of four components: the frame encoder, the classification layer, the label processing layer, and the loss function. We introduce our newly proposed methods and discusses how existing models operate in the YouTube-8M Classification Task, what insights they have, and why they succeed (or fail) to achieve good performance. Most of the models we proposed are very high compared to the baseline models, and the ensemble of the models we used is 8th in the Kaggle Competition.

Sangho Lee, Ming-Wei Shih, Prasun Gera et al.

In this paper, we explore a new, yet critical, side-channel attack against Intel Software Guard Extension (SGX), called a branch shadowing attack, which can reveal fine-grained control flows (i.e., each branch) of an enclave program running on real SGX hardware. The root cause of this attack is that Intel SGX does not clear the branch history when switching from enclave mode to non-enclave mode, leaving the fine-grained traces to the outside world through a branch-prediction side channel. However, exploiting the channel is not so straightforward in practice because 1) measuring branch prediction/misprediction penalties based on timing is too inaccurate to distinguish fine-grained control-flow changes and 2) it requires sophisticated control over the enclave execution to force its execution to the interesting code blocks. To overcome these challenges, we developed two novel exploitation techniques: 1) Intel PT- and LBR-based history-inferring techniques and 2) APIC-based technique to control the execution of enclave programs in a fine-grained manner. As a result, we could demonstrate our attack by breaking recent security constructs, including ORAM schemes, Sanctum, SGX-Shield, and T-SGX. Not limiting our work to the attack itself, we thoroughly studied the feasibility of hardware-based solutions (e.g., branch history clearing) and also proposed a software-based countermeasure, called Zigzagger, to mitigate the branch shadowing attack in practice.