Jaehong Yoon

Yiyang Zhou, Chenhang Cui, Jaehong Yoon et al.

Large vision-language models (LVLMs) have shown remarkable abilities in understanding visual information with human languages. However, LVLMs still suffer from object hallucination, which is the problem of generating descriptions that include objects that do not actually exist in the images. This can negatively impact many vision-language tasks, such as visual summarization and reasoning. To address this issue, we propose a simple yet powerful algorithm, LVLM Hallucination Revisor (LURE), to post-hoc rectify object hallucination in LVLMs by reconstructing less hallucinatory descriptions. LURE is grounded in a rigorous statistical analysis of the key factors underlying object hallucination, including co-occurrence (the frequent appearance of certain objects alongside others in images), uncertainty (objects with higher uncertainty during LVLM decoding), and object position (hallucination often appears in the later part of the generated text). LURE can also be seamlessly integrated with any LVLMs. We evaluate LURE on six open-source LVLMs, achieving a 23% improvement in general object hallucination evaluation metrics over the previous best approach. In both GPT and human evaluations, LURE consistently ranks at the top. Our data and code are available at https://github.com/YiyangZhou/LURE.

Xiaohui Zhang, Jaehong Yoon, Mohit Bansal et al.

Multimodal learning, which integrates data from diverse sensory modes, plays a pivotal role in artificial intelligence. However, existing multimodal learning methods often struggle with challenges where some modalities appear more dominant than others during multimodal learning, resulting in suboptimal performance. To address this challenge, we propose MLA (Multimodal Learning with Alternating Unimodal Adaptation). MLA reframes the conventional joint multimodal learning process by transforming it into an alternating unimodal learning process, thereby minimizing interference between modalities. Simultaneously, it captures cross-modal interactions through a shared head, which undergoes continuous optimization across different modalities. This optimization process is controlled by a gradient modification mechanism to prevent the shared head from losing previously acquired information. During the inference phase, MLA utilizes a test-time uncertainty-based model fusion mechanism to integrate multimodal information. Extensive experiments are conducted on five diverse datasets, encompassing scenarios with complete modalities and scenarios with missing modalities. These experiments demonstrate the superiority of MLA over competing prior approaches. Our code is available at https://github.com/Cecile-hi/Multimodal-Learning-with-Alternating-Unimodal-Adaptation.

Yi-Lin Sung, Jaehong Yoon, Mohit Bansal

Large Vision-Language Models (LVLMs) can understand the world comprehensively by integrating rich information from different modalities, achieving remarkable advancements on various multimodal downstream tasks. However, deploying LVLMs is often problematic due to their massive computational/energy costs and carbon consumption. Such issues make it infeasible to adopt conventional iterative global pruning, which is costly due to computing the Hessian matrix of the entire large model for sparsification. Alternatively, several studies have recently proposed layer-wise pruning approaches to avoid the expensive computation of global pruning and efficiently compress model weights according to their importance within a layer. However, they often suffer from suboptimal model compression due to their lack of a global perspective. To address this limitation in recent efficient pruning methods for large models, we propose Efficient Coarse-to-Fine LayerWise Pruning (ECoFLaP), a two-stage coarse-to-fine weight pruning approach for LVLMs. We first determine the sparsity ratios of different layers or blocks by leveraging the global importance score, which is efficiently computed based on the zeroth-order approximation of the global model gradients. Then, the model performs local layer-wise unstructured weight pruning based on globally-informed sparsity ratios. We validate our proposed method across various multimodal and unimodal models and datasets, demonstrating significant performance improvements over prevalent pruning techniques in the high-sparsity regime.

Jaewoo Lee, Jaehong Yoon, Wonjae Kim et al.

Continuously learning a variety of audio-video semantics over time is crucial for audio-related reasoning tasks in our ever-evolving world. However, this is a nontrivial problem and poses two critical challenges: sparse spatio-temporal correlation between audio-video pairs and multimodal correlation overwriting that forgets audio-video relations. To tackle this problem, we propose a new continual audio-video pre-training method with two novel ideas: (1) Localized Patch Importance Scoring: we introduce a multimodal encoder to determine the importance score for each patch, emphasizing semantically intertwined audio-video patches. (2) Replay-guided Correlation Assessment: to reduce the corruption of previously learned audiovisual knowledge due to drift, we propose to assess the correlation of the current patches on the past steps to identify the patches exhibiting high correlations with the past steps. Based on the results from the two ideas, we perform probabilistic patch selection for effective continual audio-video pre-training. Experimental validation on multiple benchmarks shows that our method achieves a 3.69%p of relative performance gain in zero-shot retrieval tasks compared to strong continual learning baselines, while reducing memory consumption by ~45%.

Taekyung Ki, Sangwon Jang, Jaehyeong Jo et al.

Talking head generation creates lifelike avatars from static portraits for virtual communication and content creation. However, current models do not yet convey the feeling of truly interactive communication, often generating one-way responses that lack emotional engagement. We identify two key challenges toward truly interactive avatars: generating motion in real-time under causal constraints and learning expressive, vibrant reactions without additional labeled data. To address these challenges, we propose Avatar Forcing, a new framework for interactive head avatar generation that models real-time user-avatar interactions through diffusion forcing. This design allows the avatar to process real-time multimodal inputs, including the user's audio and motion, with low latency for instant reactions to both verbal and non-verbal cues such as speech, nods, and laughter. Furthermore, we introduce a direct preference optimization method that leverages synthetic losing samples constructed by dropping user conditions, enabling label-free learning of expressive interaction. Experimental results demonstrate that our framework enables real-time interaction with low latency (approximately 500ms), achieving 6.8X speedup compared to the baseline, and produces reactive and expressive avatar motion, which is preferred over 80% against the baseline.

Fengli Wu, Vaidehi Patil, Jaehong Yoon et al.

Pretrained Multimodal Large Language Models (MLLMs) are increasingly deployed in medical AI systems for clinical reasoning, diagnosis support, and report generation. However, their training on sensitive patient data raises critical privacy and compliance challenges under regulations such as HIPAA and GDPR, which enforce the "right to be forgotten". Unlearning, the process of tuning models to selectively remove the influence of specific training data points, offers a potential solution, yet its effectiveness in complex medical settings remains underexplored. To systematically study this, we introduce MedForget, a Hierarchy-Aware Multimodal Unlearning Testbed with explicit retain and forget splits and evaluation sets containing rephrased variants. MedForget models hospital data as a nested hierarchy (Institution -> Patient -> Study -> Section), enabling fine-grained assessment across eight organizational levels. The benchmark contains 3840 multimodal (image, question, answer) instances, each hierarchy level having a dedicated unlearning target, reflecting distinct unlearning challenges. Experiments with four SOTA unlearning methods on three tasks (generation, classification, cloze) show that existing methods struggle to achieve complete, hierarchy-aware forgetting without reducing diagnostic performance. To test whether unlearning truly deletes hierarchical pathways, we introduce a reconstruction attack that progressively adds hierarchical level context to prompts. Models unlearned at a coarse granularity show strong resistance, while fine-grained unlearning leaves models vulnerable to such reconstruction. MedForget provides a practical, HIPAA-aligned testbed for building compliant medical AI systems.

Xiyao Wang, Yuhang Zhou, Xiaoyu Liu et al.

Multimodal Large Language Models (MLLMs) have demonstrated proficiency in handling a variety of visual-language tasks. However, current MLLM benchmarks are predominantly designed to evaluate reasoning based on static information about a single image, and the ability of modern MLLMs to extrapolate from image sequences, which is essential for understanding our ever-changing world, has been less investigated. To address this challenge, this paper introduces Mementos, a new benchmark designed to assess MLLMs' sequential image reasoning abilities. Mementos features 4,761 diverse image sequences with varying lengths. We also employ a GPT-4 assisted method to evaluate MLLM reasoning performance. Through a careful evaluation of nine recent MLLMs on Mementos, including GPT-4V and Gemini, we find that they struggle to accurately describe dynamic information about given image sequences, often leading to hallucinations/misrepresentations of objects and their corresponding behaviors. Our quantitative analysis and case studies identify three key factors impacting MLLMs' sequential image reasoning: the correlation between object and behavioral hallucinations, the influence of cooccurring behaviors, and the compounding impact of behavioral hallucinations. Our dataset is available at https://github.com/umd-huang-lab/Mementos.

Wonyong Jeong, Jaehong Yoon, Eunho Yang et al.

While existing federated learning approaches mostly require that clients have fully-labeled data to train on, in realistic settings, data obtained at the client-side often comes without any accompanying labels. Such deficiency of labels may result from either high labeling cost, or difficulty of annotation due to the requirement of expert knowledge. Thus the private data at each client may be either partly labeled, or completely unlabeled with labeled data being available only at the server, which leads us to a new practical federated learning problem, namely Federated Semi-Supervised Learning (FSSL). In this work, we study two essential scenarios of FSSL based on the location of the labeled data. The first scenario considers a conventional case where clients have both labeled and unlabeled data (labels-at-client), and the second scenario considers a more challenging case, where the labeled data is only available at the server (labels-at-server). We then propose a novel method to tackle the problems, which we refer to as Federated Matching (FedMatch). FedMatch improves upon naive combinations of federated learning and semi-supervised learning approaches with a new inter-client consistency loss and decomposition of the parameters for disjoint learning on labeled and unlabeled data. Through extensive experimental validation of our method in the two different scenarios, we show that our method outperforms both local semi-supervised learning and baselines which naively combine federated learning with semi-supervised learning. The code is available at https://github.com/wyjeong/FedMatch.

Jaehong Yoon, Wonyong Jeong, Giwoong Lee et al.

There has been a surge of interest in continual learning and federated learning, both of which are important in deep neural networks in real-world scenarios. Yet little research has been done regarding the scenario where each client learns on a sequence of tasks from a private local data stream. This problem of federated continual learning poses new challenges to continual learning, such as utilizing knowledge from other clients, while preventing interference from irrelevant knowledge. To resolve these issues, we propose a novel federated continual learning framework, Federated Weighted Inter-client Transfer (FedWeIT), which decomposes the network weights into global federated parameters and sparse task-specific parameters, and each client receives selective knowledge from other clients by taking a weighted combination of their task-specific parameters. FedWeIT minimizes interference between incompatible tasks, and also allows positive knowledge transfer across clients during learning. We validate our FedWeIT against existing federated learning and continual learning methods under varying degrees of task similarity across clients, and our model significantly outperforms them with a large reduction in the communication cost. Code is available at https://github.com/wyjeong/FedWeIT

Jaehong Yoon, Shoubin Yu, Vaidehi Patil et al.

Recent advances in diffusion models have significantly enhanced their ability to generate high-quality images and videos, but they have also increased the risk of producing unsafe content. Existing unlearning/editing-based methods for safe generation remove harmful concepts from models but face several challenges: (1) They cannot instantly remove harmful concepts without training. (2) Their safe generation capabilities depend on collected training data. (3) They alter model weights, risking degradation in quality for content unrelated to toxic concepts. To address these, we propose SAFREE, a novel, training-free approach for safe T2I and T2V, that does not alter the model's weights. Specifically, we detect a subspace corresponding to a set of toxic concepts in the text embedding space and steer prompt embeddings away from this subspace, thereby filtering out harmful content while preserving intended semantics. To balance the trade-off between filtering toxicity and preserving safe concepts, SAFREE incorporates a novel self-validating filtering mechanism that dynamically adjusts the denoising steps when applying the filtered embeddings. Additionally, we incorporate adaptive re-attention mechanisms within the diffusion latent space to selectively diminish the influence of features related to toxic concepts at the pixel level. In the end, SAFREE ensures coherent safety checking, preserving the fidelity, quality, and safety of the output. SAFREE achieves SOTA performance in suppressing unsafe content in T2I generation compared to training-free baselines and effectively filters targeted concepts while maintaining high-quality images. It also shows competitive results against training-based methods. We extend SAFREE to various T2I backbones and T2V tasks, showcasing its flexibility and generalization. SAFREE provides a robust and adaptable safeguard for ensuring safe visual generation.

Abhay Zala, Jaemin Cho, Han Lin et al. · allen-ai

Recent SOTA approaches for embodied learning via interaction directly employ large language models (LLMs) as agents to determine the next steps in an environment. Due to their world knowledge and reasoning capabilities, LLM agents achieve stronger performance than previous smaller agents based on reinforcement learning (RL); however, frequently calling LLMs is slow and expensive. Instead of directly employing LLMs as agents, can we use LLMs' reasoning capabilities to adaptively create training environments to help smaller RL agents learn useful skills that they are weak at? We propose EnvGen, a novel framework to address this question. We first prompt an LLM to generate training environments by giving it the task description and simulator objectives that the agents should learn and then asking it to generate a set of environment configurations (e.g., different terrains, items initially given to agents, etc.). Next, we train a small RL agent in a mixture of the original and LLM-generated environments. Then, we enable the LLM to continuously adapt the generated environments to progressively improve the skills that the agent is weak at, by providing feedback to the LLM in the form of the agent's performance. We demonstrate the usefulness of EnvGen with comprehensive experiments in Crafter and Heist environments. We find that a small RL agent trained with EnvGen can outperform SOTA methods, including a GPT-4 agent, and learns long-horizon tasks significantly faster. We also show that using an LLM to adapt environments dynamically outperforms curriculum learning approaches and how the environments are adapted to help improve RL agents' weaker skills over time. Additionally, EnvGen is substantially more efficient as it only uses a small number of LLM calls (e.g., 4 in total), whereas LLM agents require thousands of calls. Lastly, we present detailed ablation studies for EnvGen design choices.

Daeun Lee, Jaehong Yoon, Sung Ju Hwang

Continual Test Time Adaptation (CTTA) is required to adapt efficiently to continuous unseen domains while retaining previously learned knowledge. However, despite the progress of CTTA, it is still challenging to deploy the model with improved forgetting-adaptation trade-offs and efficiency. In addition, current CTTA scenarios assume only the disjoint situation, even though real-world domains are seamlessly changed. To address these challenges, this paper proposes BECoTTA, an input-dependent and efficient modular framework for CTTA. We propose Mixture-of Domain Low-rank Experts (MoDE) that contains two core components: (i) Domain-Adaptive Routing, which helps to selectively capture the domain adaptive knowledge with multiple domain routers, and (ii) Domain-Expert Synergy Loss to maximize the dependency between each domain and expert. We validate that our method outperforms multiple CTTA scenarios, including disjoint and gradual domain shits, while only requiring ~98% fewer trainable parameters. We also provide analyses of our method, including the construction of experts, the effect of domain-adaptive experts, and visualizations.

Jialu Li, Jaemin Cho, Yi-Lin Sung et al. · allen-ai

Recent text-to-image (T2I) generation models have demonstrated impressive capabilities in creating images from text descriptions. However, these T2I generation models often fall short of generating images that precisely match the details of the text inputs, such as incorrect spatial relationship or missing objects. In this paper, we introduce SELMA: Skill-Specific Expert Learning and Merging with Auto-Generated Data, a novel paradigm to improve the faithfulness of T2I models by fine-tuning models on automatically generated, multi-skill image-text datasets, with skill-specific expert learning and merging. First, SELMA leverages an LLM's in-context learning capability to generate multiple datasets of text prompts that can teach different skills, and then generates the images with a T2I model based on the prompts. Next, SELMA adapts the T2I model to the new skills by learning multiple single-skill LoRA (low-rank adaptation) experts followed by expert merging. Our independent expert fine-tuning specializes multiple models for different skills, and expert merging helps build a joint multi-skill T2I model that can generate faithful images given diverse text prompts, while mitigating the knowledge conflict from different datasets. We empirically demonstrate that SELMA significantly improves the semantic alignment and text faithfulness of state-of-the-art T2I diffusion models on multiple benchmarks (+2.1% on TIFA and +6.9% on DSG), human preference metrics (PickScore, ImageReward, and HPS), as well as human evaluation. Moreover, fine-tuning with image-text pairs auto-collected via SELMA shows comparable performance to fine-tuning with ground truth data. Lastly, we show that fine-tuning with images from a weaker T2I model can help improve the generation quality of a stronger T2I model, suggesting promising weak-to-strong generalization in T2I models.

Shoubin Yu, Jaehong Yoon, Mohit Bansal

Despite impressive advancements in recent multimodal reasoning approaches, they are still limited in flexibility and efficiency, as these models typically process only a few fixed modality inputs and require updates to numerous parameters. This paper tackles these critical challenges and proposes CREMA, a generalizable, highly efficient, and modular modality-fusion framework that can incorporate any new modality to enhance video reasoning. We first augment multiple informative modalities (such as optical flow, 3D point cloud, audio, thermal heatmap, and touch map) from given videos without extra human annotation by leveraging sensors or existing pre-trained models. Next, we introduce a query transformer with multiple parameter-efficient modules associated with each accessible modality. It projects diverse modality features to the LLM token embedding space, allowing the model to integrate different data types for response generation. Furthermore, we propose a novel progressive multimodal fusion design supported by a lightweight fusion module and modality-sequential training strategy. It helps compress information across various assisting modalities, maintaining computational efficiency in the LLM while improving performance. We validate our method on 7 video-language reasoning tasks assisted by diverse modalities, including conventional VideoQA and Video-Audio/3D/Touch/Thermal QA, and achieve better/equivalent performance against strong multimodal LLMs, including OneLLM, BLIP-2, and SeViLA while reducing over 90% trainable parameters. We provide extensive analyses of CREMA, including the impact of each modality on reasoning domains, the design of the fusion module, and example visualizations.

Daeun Lee, Jaehong Yoon, Jaemin Cho et al. · allen-ai

Recent text-to-video (T2V) diffusion models have demonstrated impressive generation capabilities across various domains. However, these models often generate videos that have misalignments with text prompts, especially when the prompts describe complex scenes with multiple objects and attributes. To address this, we introduce VideoRepair, a novel model-agnostic, training-free video refinement framework that automatically identifies fine-grained text-video misalignments and generates explicit spatial and textual feedback, enabling a T2V diffusion model to perform targeted, localized refinements. VideoRepair consists of two stages: In (1) video refinement planning, we first detect misalignments by generating fine-grained evaluation questions and answering them using an MLLM. Based on video evaluation outputs, we identify accurately generated objects and construct localized prompts to precisely refine misaligned regions. In (2) localized refinement, we enhance video alignment by 'repairing' the misaligned regions from the original video while preserving the correctly generated areas. This is achieved by frame-wise region decomposition using our Region-Preserving Segmentation (RPS) module. On two popular video generation benchmarks (EvalCrafter and T2V-CompBench), VideoRepair substantially outperforms recent baselines across various text-video alignment metrics. We provide a comprehensive analysis of VideoRepair components and qualitative examples.

Ziyang Wang, Jaehong Yoon, Shoubin Yu et al.

Despite advances in reinforcement learning (RL)-based video reasoning with large language models (LLMs), data collection and fine-tuning remain significant challenges. These methods often rely on large-scale supervised fine-tuning (SFT) with extensive video data and long Chain-of-Thought (CoT) annotations, making them costly and hard to scale. To address this, we present Video-RTS, a new approach to improve video reasoning capability with drastically improved data efficiency by combining data-efficient RL with a video-adaptive test-time scaling (TTS) strategy. Building on observations about the data scaling, we skip the resource-intensive SFT step and employ efficient pure-RL training with output-based rewards, requiring no additional annotations or extensive fine-tuning. Furthermore, to utilize computational resources more efficiently, we introduce a sparse-to-dense video TTS strategy that improves inference by iteratively adding frames based on output consistency. We validate our approach on multiple video reasoning benchmarks, showing that Video-RTS surpasses existing video reasoning models by 2.4% in accuracy using only 3.6% training samples. Specifically, Video-RTS achieves a 4.2% improvement on Video-Holmes, a recent and challenging video reasoning benchmark. Notably, our pure RL training and adaptive video TTS offer complementary strengths, enabling Video-RTS's strong reasoning performance.

Jialu Li, Shoubin Yu, Han Lin et al. · allen-ai

Recent advancements in text-to-video (T2V) diffusion models have significantly enhanced the visual quality of the generated videos. However, even recent T2V models find it challenging to follow text descriptions accurately, especially when the prompt requires accurate control of spatial layouts or object trajectories. A recent line of research uses layout guidance for T2V models that require fine-tuning or iterative manipulation of the attention map during inference time. This significantly increases the memory requirement, making it difficult to adopt a large T2V model as a backbone. To address this, we introduce Video-MSG, a training-free Guidance method for T2V generation based on Multimodal planning and Structured noise initialization. Video-MSG consists of three steps, where in the first two steps, Video-MSG creates Video Sketch, a fine-grained spatio-temporal plan for the final video, specifying background, foreground, and object trajectories, in the form of draft video frames. In the last step, Video-MSG guides a downstream T2V diffusion model with Video Sketch through noise inversion and denoising. Notably, Video-MSG does not need fine-tuning or attention manipulation with additional memory during inference time, making it easier to adopt large T2V models. Video-MSG demonstrates its effectiveness in enhancing text alignment with multiple T2V backbones (VideoCrafter2 and CogVideoX-5B) on popular T2V generation benchmarks (T2VCompBench and VBench). We provide comprehensive ablation studies about noise inversion ratio, different background generators, background object detection, and foreground object segmentation.

Adyasha Maharana, Jaehong Yoon, Tianlong Chen et al.

Visual instruction datasets from various distributors are released at different times and often contain a significant number of semantically redundant text-image pairs, depending on their task compositions (i.e., skills) or reference sources. This redundancy greatly limits the efficient deployment of continually adaptable multimodal large language models, hindering their ability to refine existing skills and acquire new competencies over time. We reframe the problem of lifelong Instruction Tuning (LiIT) via data selection, where the model automatically selects beneficial samples to learn from earlier and new datasets based on the current state of acquired knowledge in the model. We propose Adapt-$\infty$, a new multi-way and adaptive data selection approach that dynamically balances sample efficiency and effectiveness during LiIT. We first construct pseudo-skill clusters by grouping gradient-based sample vectors. Next, we select the best-performing data selector for each skill cluster from a pool of selector experts, including our newly proposed scoring function, Image Grounding score. This data selector samples a subset of the most important samples from each skill cluster for training. To prevent the continuous increase in the size of the dataset pool during LiIT, we introduce a cluster-wise permanent data pruning strategy to remove the most semantically redundant samples from each cluster, keeping computational requirements manageable. We validate the effectiveness and efficiency of Adapt-$\infty$ over a sequence of multimodal instruction tuning datasets with various tasks, including (Knowledge) VQA, multilingual, grounding, reasoning, language-only, and multi-image comprehension. Training with samples selected by Adapt-$\infty$ alleviates catastrophic forgetting, especially for rare tasks, and promotes forward transfer across the continuum using only a fraction of the original data.

Haeyong Kang, Jaehong Yoon, Sung Ju Hwang et al.

Inspired by the Lottery Ticket Hypothesis (LTH), which highlights the existence of efficient subnetworks within larger, dense networks, a high-performing Winning Subnetwork (WSN) in terms of task performance under appropriate sparsity conditions is considered for various continual learning tasks. It leverages pre-existing weights from dense networks to achieve efficient learning in Task Incremental Learning (TIL) and Task-agnostic Incremental Learning (TaIL) scenarios. In Few-Shot Class Incremental Learning (FSCIL), a variation of WSN referred to as the Soft subnetwork (SoftNet) is designed to prevent overfitting when the data samples are scarce. Furthermore, the sparse reuse of WSN weights is considered for Video Incremental Learning (VIL). The use of Fourier Subneural Operator (FSO) within WSN is considered. It enables compact encoding of videos and identifies reusable subnetworks across varying bandwidths. We have integrated FSO into different architectural frameworks for continual learning, including VIL, TIL, and FSCIL. Our comprehensive experiments demonstrate FSO's effectiveness, significantly improving task performance at various convolutional representational levels. Specifically, FSO enhances higher-layer performance in TIL and FSCIL and lower-layer performance in VIL.

Daeun Lee, Jaehong Yoon, Jaemin Cho et al. · allen-ai

Recent advances in Chain-of-Thought (CoT) reasoning have improved complex video understanding, but existing methods often struggle to adapt to domain-specific skills (e.g., event detection, spatial relation understanding, emotion understanding) over various video content. To address this, we propose Video-Skill-CoT (a.k.a. Video-SKoT), a framework that automatically constructs and leverages skill-aware CoT supervisions for domain-adaptive video reasoning. First, we construct skill-based CoT annotations: we extract domain-relevant reasoning skills from training questions, cluster them into a shared skill taxonomy, and create detailed multi-step CoT rationale tailored to each video-question pair for training. Second, we introduce a skill-specific expert learning framework. Each expert module specializes in a subset of reasoning skills and is trained with lightweight adapters using the collected CoT supervision. We demonstrate the effectiveness of the proposed approach on three video understanding benchmarks, where Video-SKoT consistently outperforms strong baselines. We also provide in-depth analyses on comparing different CoT annotation pipelines and learned skills over multiple video domains.

Yi-Lin Sung, Prateek Yadav, Jialu Li et al.

Layer-wise quantization is a key technique for efficiently compressing large models without expensive retraining. Previous methods typically quantize the weights of each layer by "uniformly" optimizing the layer reconstruction loss across all output tokens. However, in this paper, we demonstrate that better-quantized models can be obtained by prioritizing learning from important tokens (e.g. which have large attention scores). Building on this finding, we propose RSQ (Rotate, Scale, then Quantize), which (1) applies rotations (orthogonal transformation) to the model to mitigate outliers (those with exceptionally large magnitude), (2) scales the token feature based on its importance, and (3) quantizes the model using the GPTQ framework with the second-order statistics computed by scaled tokens. To compute token importance, we explore both heuristic and dynamic strategies. Based on a thorough analysis of all approaches, we adopt attention concentration, which uses attention scores of each token as its importance, as the best approach. We demonstrate that RSQ consistently outperforms baseline methods across multiple downstream tasks and three model families: LLaMA3, Mistral, and Qwen2.5. Additionally, models quantized with RSQ achieve superior performance on long-context tasks, further highlighting its effectiveness. Lastly, RSQ demonstrates generalizability across various setups, including different model sizes, calibration datasets, bit precisions, and quantization methods.

Zun Wang, Jaemin Cho, Jialu Li et al. · allen-ai

Recent approaches on 3D camera control in video diffusion models (VDMs) often create anchor videos to guide diffusion models as a structured prior by rendering from estimated point clouds following annotated camera trajectories. However, errors inherent in point cloud estimation often lead to inaccurate anchor videos. Moreover, the requirement for extensive camera trajectory annotations further increases resource demands. To address these limitations, we introduce EPiC, an efficient and precise camera control learning framework that automatically constructs high-quality anchor videos without expensive camera trajectory annotations. Concretely, we create highly precise anchor videos for training by masking source videos based on first-frame visibility. This approach ensures high alignment, eliminates the need for camera trajectory annotations, and thus can be readily applied to any in-the-wild video to generate image-to-video (I2V) training pairs. Furthermore, we introduce Anchor-ControlNet, a lightweight conditioning module that integrates anchor video guidance in visible regions to pretrained VDMs, with less than 1% of backbone model parameters. By combining the proposed anchor video data and ControlNet module, EPiC achieves efficient training with substantially fewer parameters, training steps, and less data, without requiring modifications to the diffusion model backbone typically needed to mitigate rendering misalignments. Although being trained on masking-based anchor videos, our method generalizes robustly to anchor videos made with point clouds during inference, enabling precise 3D-informed camera control. EPiC achieves SOTA performance on RealEstate10K and MiraData for I2V camera control task, demonstrating precise and robust camera control ability both quantitatively and qualitatively. Notably, EPiC also exhibits strong zero-shot generalization to video-to-video scenarios.

Zun Wang, Jialu Li, Han Lin et al.

Storytelling video generation (SVG) aims to produce coherent and visually rich multi-scene videos that follow a structured narrative. Existing methods primarily employ LLM for high-level planning to decompose a story into scene-level descriptions, which are then independently generated and stitched together. However, these approaches struggle with generating high-quality videos aligned with the complex single-scene description, as visualizing such complex description involves coherent composition of multiple characters and events, complex motion synthesis and multi-character customization. To address these challenges, we propose DREAMRUNNER, a novel story-to-video generation method: First, we structure the input script using a large language model (LLM) to facilitate both coarse-grained scene planning as well as fine-grained object-level layout planning. Next, DREAMRUNNER presents retrieval-augmented test-time adaptation to capture target motion priors for objects in each scene, supporting diverse motion customization based on retrieved videos, thus facilitating the generation of new videos with complex, scripted motions. Lastly, we propose a novel spatial-temporal region-based 3D attention and prior injection module SR3AI for fine-grained object-motion binding and frame-by-frame spatial-temporal semantic control. We compare DREAMRUNNER with various SVG baselines, demonstrating state-of-the-art performance in character consistency, text alignment, and smooth transitions. Additionally, DREAMRUNNER exhibits strong fine-grained condition-following ability in compositional text-to-video generation, significantly outperforming baselines on T2V-ComBench. Finally, we validate DREAMRUNNER's robust ability to generate multi-object interactions with qualitative examples.

Yidong Huang, Zun Wang, Han Lin et al.

Recent video generation approaches increasingly rely on planning intermediate control signals such as object trajectories to improve temporal coherence and motion fidelity. However, these methods mostly employ single-shot plans that are typically limited to simple motions, or iterative refinement which requires multiple calls to the video generator, incuring high computational cost. To overcome these limitations, we propose SketchVerify, a training-free, sketch-verification-based planning framework that improves motion planning quality with more dynamically coherent trajectories (i.e., physically plausible and instruction-consistent motions) prior to full video generation by introducing a test-time sampling and verification loop. Given a prompt and a reference image, our method predicts multiple candidate motion plans and ranks them using a vision-language verifier that jointly evaluates semantic alignment with the instruction and physical plausibility. To efficiently score candidate motion plans, we render each trajectory as a lightweight video sketch by compositing objects over a static background, which bypasses the need for expensive, repeated diffusion-based synthesis while achieving comparable performance. We iteratively refine the motion plan until a satisfactory one is identified, which is then passed to the trajectory-conditioned generator for final synthesis. Experiments on WorldModelBench and PhyWorldBench demonstrate that our method significantly improves motion quality, physical realism, and long-term consistency compared to competitive baselines while being substantially more efficient. Our ablation study further shows that scaling up the number of trajectory candidates consistently enhances overall performance.

Sangwon Jang, Taekyung Ki, Jaehyeong Jo et al.

Advancements in diffusion models have significantly improved video quality, directing attention to fine-grained controllability. However, many existing methods depend on fine-tuning large-scale video models for specific tasks, which becomes increasingly impractical as model sizes continue to grow. In this work, we present Frame Guidance, a training-free guidance for controllable video generation based on frame-level signals, such as keyframes, style reference images, sketches, or depth maps. For practical training-free guidance, we propose a simple latent processing method that dramatically reduces memory usage, and apply a novel latent optimization strategy designed for globally coherent video generation. Frame Guidance enables effective control across diverse tasks, including keyframe guidance, stylization, and looping, without any training, compatible with any video models. Experimental results show that Frame Guidance can produce high-quality controlled videos for a wide range of tasks and input signals.

Jaehong Yoon, Divyam Madaan, Eunho Yang et al.

A dataset is a shred of crucial evidence to describe a task. However, each data point in the dataset does not have the same potential, as some of the data points can be more representative or informative than others. This unequal importance among the data points may have a large impact in rehearsal-based continual learning, where we store a subset of the training examples (coreset) to be replayed later to alleviate catastrophic forgetting. In continual learning, the quality of the samples stored in the coreset directly affects the model's effectiveness and efficiency. The coreset selection problem becomes even more important under realistic settings, such as imbalanced continual learning or noisy data scenarios. To tackle this problem, we propose Online Coreset Selection (OCS), a simple yet effective method that selects the most representative and informative coreset at each iteration and trains them in an online manner. Our proposed method maximizes the model's adaptation to a current dataset while selecting high-affinity samples to past tasks, which directly inhibits catastrophic forgetting. We validate the effectiveness of our coreset selection mechanism over various standard, imbalanced, and noisy datasets against strong continual learning baselines, demonstrating that it improves task adaptation and prevents catastrophic forgetting in a sample-efficient manner.

Minyoung Song, Jaehong Yoon, Eunho Yang et al.

As deep neural networks are growing in size and being increasingly deployed to more resource-limited devices, there has been a recent surge of interest in network pruning methods, which aim to remove less important weights or activations of a given network. A common limitation of most existing pruning techniques, is that they require pre-training of the network at least once before pruning, and thus we can benefit from reduction in memory and computation only at the inference time. However, reducing the training cost of neural networks with rapid structural pruning may be beneficial either to minimize monetary cost with cloud computing or to enable on-device learning on a resource-limited device. Recently introduced random-weight pruning approaches can eliminate the needs of pretraining, but they often obtain suboptimal performance over conventional pruning techniques and also does not allow for faster training since they perform unstructured pruning. To overcome their limitations, we propose Set-based Task-Adaptive Meta Pruning (STAMP), which task-adaptively prunes a network pretrained on a large reference dataset by generating a pruning mask on it as a function of the target dataset. To ensure maximum performance improvements on the target task, we meta-learn the mask generator over different subsets of the reference dataset, such that it can generalize well to any unseen datasets within a few gradient steps of training. We validate STAMP against recent advanced pruning methods on benchmark datasets, on which it not only obtains significantly improved compression rates over the baselines at similar accuracy, but also orders of magnitude faster training speed.

Jaehong Yoon, Eunho Yang, Jeongtae Lee et al.

We propose a novel deep network architecture for lifelong learning which we refer to as Dynamically Expandable Network (DEN), that can dynamically decide its network capacity as it trains on a sequence of tasks, to learn a compact overlapping knowledge sharing structure among tasks. DEN is efficiently trained in an online manner by performing selective retraining, dynamically expands network capacity upon arrival of each task with only the necessary number of units, and effectively prevents semantic drift by splitting/duplicating units and timestamping them. We validate DEN on multiple public datasets under lifelong learning scenarios, on which it not only significantly outperforms existing lifelong learning methods for deep networks, but also achieves the same level of performance as the batch counterparts with substantially fewer number of parameters. Further, the obtained network fine-tuned on all tasks obtained significantly better performance over the batch models, which shows that it can be used to estimate the optimal network structure even when all tasks are available in the first place.