Gunhee Kim

Hyunwoo Kim, Jack Hessel, Liwei Jiang et al. · allen-ai, cmu

Data scarcity has been a long standing issue in the field of open-domain social dialogue. To quench this thirst, we present SODA: the first publicly available, million-scale high-quality social dialogue dataset. By contextualizing social commonsense knowledge from a knowledge graph, we are able to distill an exceptionally broad spectrum of social interactions from a large language model. Human evaluation shows that conversations in SODA are more consistent, specific, and (surprisingly) natural than those in prior human-authored datasets. Using SODA, we train COSMO: a generalizable conversation model that is significantly more natural and consistent on unseen datasets than best-performing conversation models (e.g., GODEL, BlenderBot-1, Koala, Vicuna). Experiments reveal COSMO is sometimes even preferred to the original human-written gold responses. Additionally, our results shed light on the distinction between knowledge-enriched conversations and natural social chitchats. We plan to make our data, model, and code public.

Hyunwoo Kim, Youngjae Yu, Liwei Jiang et al. · allen-ai, cmu

Most existing dialogue systems fail to respond properly to potentially unsafe user utterances by either ignoring or passively agreeing with them. To address this issue, we introduce ProsocialDialog, the first large-scale multi-turn dialogue dataset to teach conversational agents to respond to problematic content following social norms. Covering diverse unethical, problematic, biased, and toxic situations, ProsocialDialog contains responses that encourage prosocial behavior, grounded in commonsense social rules (i.e., rules-of-thumb, RoTs). Created via a human-AI collaborative framework, ProsocialDialog consists of 58K dialogues, with 331K utterances, 160K unique RoTs, and 497K dialogue safety labels accompanied by free-form rationales. With this dataset, we introduce a dialogue safety detection module, Canary, capable of generating RoTs given conversational context, and a socially-informed dialogue agent, Prost. Empirical results show that Prost generates more socially acceptable dialogues compared to other state-of-the-art language and dialogue models in both in-domain and out-of-domain settings. Additionally, Canary effectively guides conversational agents and off-the-shelf language models to generate significantly more prosocial responses. Our work highlights the promise and importance of creating and steering conversational AI to be socially responsible.

Hyunwoo Kim, Melanie Sclar, Xuhui Zhou et al. · allen-ai, cmu

Theory of mind (ToM) evaluations currently focus on testing models using passive narratives that inherently lack interactivity. We introduce FANToM, a new benchmark designed to stress-test ToM within information-asymmetric conversational contexts via question answering. Our benchmark draws upon important theoretical requisites from psychology and necessary empirical considerations when evaluating large language models (LLMs). In particular, we formulate multiple types of questions that demand the same underlying reasoning to identify illusory or false sense of ToM capabilities in LLMs. We show that FANToM is challenging for state-of-the-art LLMs, which perform significantly worse than humans even with chain-of-thought reasoning or fine-tuning.

Youngjae Yu, Jiwan Chung, Heeseung Yun et al. · allen-ai, uw

Large language models readily adapt to novel settings, even without task-specific training data. Can their zero-shot capacity be extended to multimodal inputs? In this work, we propose ESPER which extends language-only zero-shot models to unseen multimodal tasks, like image and audio captioning. Our key novelty is to use reinforcement learning to align multimodal inputs to language model generations without direct supervision: for example, in the image case our reward optimization relies only on cosine similarity derived from CLIP, and thus requires no additional explicitly paired (image, caption) data. Because the parameters of the language model are left unchanged, the model maintains its capacity for zero-shot generalization. Experiments demonstrate that ESPER outperforms baselines and prior work on a variety of zero-shot tasks; these include a new benchmark we collect+release, ESP dataset, which tasks models with generating several diversely-styled captions for each image.

Heeseung Yun, Sehun Lee, Gunhee Kim

360$^\circ$ video saliency detection is one of the challenging benchmarks for 360$^\circ$ video understanding since non-negligible distortion and discontinuity occur in the projection of any format of 360$^\circ$ videos, and capture-worthy viewpoint in the omnidirectional sphere is ambiguous by nature. We present a new framework named Panoramic Vision Transformer (PAVER). We design the encoder using Vision Transformer with deformable convolution, which enables us not only to plug pretrained models from normal videos into our architecture without additional modules or finetuning but also to perform geometric approximation only once, unlike previous deep CNN-based approaches. Thanks to its powerful encoder, PAVER can learn the saliency from three simple relative relations among local patch features, outperforming state-of-the-art models for the Wild360 benchmark by large margins without supervision or auxiliary information like class activation. We demonstrate the utility of our saliency prediction model with the omnidirectional video quality assessment task in VQA-ODV, where we consistently improve performance without any form of supervision, including head movement.

Fatemeh Pesaran zadeh, Seyeon Choi, Xing Han Lù et al.

Large language models (LLMs) have enabled web agents that follow natural language goals through multi-step browser interactions. However, agents fine-tuned on specific trajectories and domain often struggle to generalize out of domain, and offline training can be compute-inefficient due to noisy, redundant trajectories and long accessibility-tree (AXTree) states. To address both issues, we propose Weasel, a trajectory selection method for offline training of web agents. Weasel selects a fixed-budget subset of trajectory steps by optimizing an objective that balances unary importance with pairwise diversity over states, websites, and interaction patterns, solving efficiently with a greedy algorithm. We further improve efficiency with target-centered AXTree pruning that keeps only content around the ground-truth action target, and we mitigate style mismatch for reasoning-native models by replacing expert traces with model-generated, style-consistent rationales. Across AgentTrek and NNetNav training datasets, evaluations in WebArena, WorkArena, and MiniWob, and experiments with Qwen2.5-7B, Gemma3-4B, and Qwen3-8B, Weasel improves out-of-domain performance while reducing training cost, producing roughly 9.7-12.5$\times$ training speedups over standard fine-tuning. We make the code available at https://github.com/fatemehpesaran310/weasel.

Yookoon Park, Chris Dongjoo Kim, Gunhee Kim

Variational autoencoders employ an amortized inference model to approximate the posterior of latent variables. However, such amortized variational inference faces two challenges: (1) the limited posterior expressiveness of fully-factorized Gaussian assumption and (2) the amortization error of the inference model. We present a novel approach that addresses both challenges. First, we focus on ReLU networks with Gaussian output and illustrate their connection to probabilistic PCA. Building on this observation, we derive an iterative algorithm that finds the mode of the posterior and apply full-covariance Gaussian posterior approximation centered on the mode. Subsequently, we present a general framework named Variational Laplace Autoencoders (VLAEs) for training deep generative models. Based on the Laplace approximation of the latent variable posterior, VLAEs enhance the expressiveness of the posterior while reducing the amortization error. Empirical results on MNIST, Omniglot, Fashion-MNIST, SVHN and CIFAR10 show that the proposed approach significantly outperforms other recent amortized or iterative methods on the ReLU networks.

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.

Jaehyeon Son, Soochan Lee, Gunhee Kim · gatech

Over the past decade, deep neural networks have demonstrated significant success using the training scheme that involves mini-batch stochastic gradient descent on extensive datasets. Expanding upon this accomplishment, there has been a surge in research exploring the application of neural networks in other learning scenarios. One notable framework that has garnered significant attention is meta-learning. Often described as "learning to learn," meta-learning is a data-driven approach to optimize the learning algorithm. Other branches of interest are continual learning and online learning, both of which involve incrementally updating a model with streaming data. While these frameworks were initially developed independently, recent works have started investigating their combinations, proposing novel problem settings and learning algorithms. However, due to the elevated complexity and lack of unified terminology, discerning differences between the learning frameworks can be challenging even for experienced researchers. To facilitate a clear understanding, this paper provides a comprehensive survey that organizes various problem settings using consistent terminology and formal descriptions. By offering an overview of these learning paradigms, our work aims to foster further advancements in this promising area of research.

Soochan Lee, Gunhee Kim

Generating intermediate steps, or Chain of Thought (CoT), is an effective way to significantly improve language models' (LM) multi-step reasoning capability. However, the CoT lengths can grow rapidly with the problem complexity, easily exceeding the maximum context size. Instead of increasing the context limit, which has already been heavily investigated, we explore an orthogonal direction: making LMs divide a problem into multiple contexts. We propose a new inference framework, called Recursion of Thought (RoT), which introduces several special tokens that the models can output to trigger context-related operations. Extensive experiments with multiple architectures including GPT-3 show that RoT dramatically improves LMs' inference capability to solve problems, whose solution consists of hundreds of thousands of tokens.

Sangwon Baik, Gunhee Kim, Mingi Choi et al.

Vision-Language Models (VLMs) exhibit strong visual reasoning capabilities, yet they still struggle with 3D understanding. In particular, VLMs often fail to infer a text-consistent goal 6D pose of a target object in a 3D scene. However, we find that with some inference-time techniques and iterative reasoning, VLMs can achieve dramatic performance gains. Concretely, given a 3D scene represented by an RGB-D image (or a compositional scene of 3D meshes) and a text instruction specifying a desired state change, we repeat the following loop: observe the current scene; evaluate whether it is faithful to the instruction; propose a pose update for the target object; apply the update; and render the updated scene. Through this closed-loop interaction, the VLM effectively acts as an agent. We further introduce three inference-time techniques that are essential to this closed-loop process: (i) multi-view reasoning with supporting view selection, (ii) object-centered coordinate system visualization, and (iii) single-axis rotation prediction. Without any additional fine-tuning or new modules, our approach surpasses prior methods at predicting the text-guided goal 6D pose of the target object. It works consistently across both closed-source and open-source VLMs. Moreover, when combining our 6D pose prediction with simple robot motion planning, it enables more successful robot manipulation than existing methods. Finally, we conduct an ablation study to demonstrate the necessity of each proposed technique.

Soochan Lee, Jaehyeon Son, Gunhee Kim · gatech

In this work, we aim to establish a strong connection between two significant bodies of machine learning research: continual learning and sequence modeling. That is, we propose to formulate continual learning as a sequence modeling problem, allowing advanced sequence models to be utilized for continual learning. Under this formulation, the continual learning process becomes the forward pass of a sequence model. By adopting the meta-continual learning (MCL) framework, we can train the sequence model at the meta-level, on multiple continual learning episodes. As a specific example of our new formulation, we demonstrate the application of Transformers and their efficient variants as MCL methods. Our experiments on seven benchmarks, covering both classification and regression, show that sequence models can be an attractive solution for general MCL.

Heeseung Yun, Ruohan Gao, Ishwarya Ananthabhotla et al.

Egocentric videos provide comprehensive contexts for user and scene understanding, spanning multisensory perception to behavioral interaction. We propose Spherical World-Locking (SWL) as a general framework for egocentric scene representation, which implicitly transforms multisensory streams with respect to measurements of head orientation. Compared to conventional head-locked egocentric representations with a 2D planar field-of-view, SWL effectively offsets challenges posed by self-motion, allowing for improved spatial synchronization between input modalities. Using a set of multisensory embeddings on a worldlocked sphere, we design a unified encoder-decoder transformer architecture that preserves the spherical structure of the scene representation, without requiring expensive projections between image and world coordinate systems. We evaluate the effectiveness of the proposed framework on multiple benchmark tasks for egocentric video understanding, including audio-visual active speaker localization, auditory spherical source localization, and behavior anticipation in everyday activities.

Eunkyu Park, Wesley Hanwen Deng, Cheyon Jin et al.

Vision-Language Models (VLMs) continue to struggle to make morally salient judgments in multimodal and socially ambiguous contexts. Prior works typically rely on binary or pairwise supervision, which often fail to capture the continuous and pluralistic nature of human moral reasoning. We present MM-SCALE (Multimodal Moral Scale), a large-scale dataset for aligning VLMs with human moral preferences through 5-point scalar ratings and explicit modality grounding. Each image-scenario pair is annotated with moral acceptability scores and grounded reasoning labels by humans using an interface we tailored for data collection, enabling listwise preference optimization over ranked scenario sets. By moving from discrete to scalar supervision, our framework provides richer alignment signals and finer calibration of multimodal moral reasoning. Experiments show that VLMs fine-tuned on MM-SCALE achieve higher ranking fidelity and more stable safety calibration than those trained with binary signals.

Eunkyu Park, Wesley Hanwen Deng, Vasudha Varadarajan et al.

Explanations are often promoted as tools for transparency, but they can also foster confirmation bias; users may assume reasoning is correct whenever outputs appear acceptable. We study this double-edged role of Chain-of-Thought (CoT) explanations in multimodal moral scenarios by systematically perturbing reasoning chains and manipulating delivery tones. Specifically, we analyze reasoning errors in vision language models (VLMs) and how they impact user trust and the ability to detect errors. Our findings reveal two key effects: (1) users often equate trust with outcome agreement, sustaining reliance even when reasoning is flawed, and (2) the confident tone suppresses error detection while maintaining reliance, showing that delivery styles can override correctness. These results highlight how CoT explanations can simultaneously clarify and mislead, underscoring the need for NLP systems to provide explanations that encourage scrutiny and critical thinking rather than blind trust. All code will be released publicly.

Jongbin Won, Minhyuk Song, Gunhee Kim et al.

Displacement is an important measurement for the assessment of structural conditions, but its field measurement is often hindered by difficulties associated with sensor installation and measurement accuracy. To overcome the disadvantages of conventional displacement measurement, computer vision (CV)-based methods have been implemented due to their remote sensing capabilities and accuracy. This paper presents a strategy for non-target structural displacement measurement that makes use of CV to avoid the need to install a target on the structure while calibrating the displacement using structured light. The proposed system called as LAVOLUTION calculates the relative position of the camera with regard to the structure using four equally spaced beams of structured light and obtains a scale factor to convert pixel movement into structural displacement. A jig for the four beams of structured light is designed and a corresponding alignment process is proposed. A method for calculating the scale factor using the designed jig for tunable structured-light is proposed and validated via numerical simulations and lab-scale experiments. To confirm the feasibility of the proposed displacement measurement process, experiments on a shaking table and a full-scale bridge are conducted and the accuracy of the proposed method is compared with that of a reference laser doppler vibrometer.

Minjung Kim, Junseo Koo, Gunhee Kim

Visual localization is the task of estimating a 6-DoF camera pose of a query image within a provided 3D reference map. Thanks to recent advances in various 3D sensors, 3D point clouds are becoming a more accurate and affordable option for building the reference map, but research to match the points of 3D point clouds with pixels in 2D images for visual localization remains challenging. Existing approaches that jointly learn 2D-3D feature matching suffer from low inliers due to representational differences between the two modalities, and the methods that bypass this problem into classification have an issue of poor refinement. In this work, we propose EP2P-Loc, a novel large-scale visual localization method that mitigates such appearance discrepancy and enables end-to-end training for pose estimation. To increase the number of inliers, we propose a simple algorithm to remove invisible 3D points in the image, and find all 2D-3D correspondences without keypoint detection. To reduce memory usage and search complexity, we take a coarse-to-fine approach where we extract patch-level features from 2D images, then perform 2D patch classification on each 3D point, and obtain the exact corresponding 2D pixel coordinates through positional encoding. Finally, for the first time in this task, we employ a differentiable PnP for end-to-end training. In the experiments on newly curated large-scale indoor and outdoor benchmarks based on 2D-3D-S and KITTI, we show that our method achieves the state-of-the-art performance compared to existing visual localization and image-to-point cloud registration methods.

Insu Jeon, Minui Hong, Junhyeog Yun et al.

Federated Learning (FL) aims to train a global inference model from remotely distributed clients, gaining popularity due to its benefit of improving data privacy. However, traditional FL often faces challenges in practical applications, including model overfitting and divergent local models due to limited and non-IID data among clients. To address these issues, we introduce a novel Bayesian meta-learning approach called meta-variational dropout (MetaVD). MetaVD learns to predict client-dependent dropout rates via a shared hypernetwork, enabling effective model personalization of FL algorithms in limited non-IID data settings. We also emphasize the posterior adaptation view of meta-learning and the posterior aggregation view of Bayesian FL via the conditional dropout posterior. We conducted extensive experiments on various sparse and non-IID FL datasets. MetaVD demonstrated excellent classification accuracy and uncertainty calibration performance, especially for out-of-distribution (OOD) clients. MetaVD compresses the local model parameters needed for each client, mitigating model overfitting and reducing communication costs. Code is available at https://github.com/insujeon/MetaVD.

Seokwon Song, Minsu Park, Gunhee Kim

Source attribution aims to enhance the reliability of AI-generated answers by including references for each statement, helping users validate the provided answers. However, existing work has primarily focused on text-only scenario and largely overlooked the role of multimodality. We introduce MAVIS, the first benchmark designed to evaluate multimodal source attribution systems that understand user intent behind visual questions, retrieve multimodal evidence, and generate long-form answers with citations. Our dataset comprises 157K visual QA instances, where each answer is annotated with fact-level citations referring to multimodal documents. We develop fine-grained automatic metrics along three dimensions of informativeness, groundedness, and fluency, and demonstrate their strong correlation with human judgments. Our key findings are threefold: (1) LVLMs with multimodal RAG generate more informative and fluent answers than unimodal RAG, but they exhibit weaker groundedness for image documents than for text documents, a gap amplified in multimodal settings. (2) Given the same multimodal documents, there is a trade-off between informativeness and groundedness across different prompting methods. (3) Our proposed method highlights mitigating contextual bias in interpreting image documents as a crucial direction for future research. The dataset and experimental code are available at https://github.com/seokwon99/MAVIS

Minjung Kim, Hyung Suk Lim, Soonyoung Lee et al.

3D dense captioning is a task involving the localization of objects and the generation of descriptions for each object in a 3D scene. Recent approaches have attempted to incorporate contextual information by modeling relationships with object pairs or aggregating the nearest neighbor features of an object. However, the contextual information constructed in these scenarios is limited in two aspects: first, objects have multiple positional relationships that exist across the entire global scene, not only near the object itself. Second, it faces with contradicting objectives--where localization and attribute descriptions are generated better with tight localization, while descriptions involving global positional relations are generated better with contextualized features of the global scene. To overcome this challenge, we introduce BiCA, a transformer encoder-decoder pipeline that engages in 3D dense captioning for each object with Bi-directional Contextual Attention. Leveraging parallelly decoded instance queries for objects and context queries for non-object contexts, BiCA generates object-aware contexts, where the contexts relevant to each object is summarized, and context-aware objects, where the objects relevant to the summarized object-aware contexts are aggregated. This extension relieves previous methods from the contradicting objectives, enhancing both localization performance and enabling the aggregation of contextual features throughout the global scene; thus improving caption generation performance simultaneously. Extensive experiments on two of the most widely-used 3D dense captioning datasets demonstrate that our proposed method achieves a significant improvement over prior methods.

Minjung Kim, Hyung Suk Lim, Seung Hwan Kim et al.

3D dense captioning is a task to localize objects in a 3D scene and generate descriptive sentences for each object. Recent approaches in 3D dense captioning have adopted transformer encoder-decoder frameworks from object detection to build an end-to-end pipeline without hand-crafted components. However, these approaches struggle with contradicting objectives where a single query attention has to simultaneously view both the tightly localized object regions and contextual environment. To overcome this challenge, we introduce SIA (See-It-All), a transformer pipeline that engages in 3D dense captioning with a novel paradigm called late aggregation. SIA simultaneously decodes two sets of queries-context query and instance query. The instance query focuses on localization and object attribute descriptions, while the context query versatilely captures the region-of-interest of relationships between multiple objects or with the global scene, then aggregated afterwards (i.e., late aggregation) via simple distance-based measures. To further enhance the quality of contextualized caption generation, we design a novel aggregator to generate a fully informed caption based on the surrounding context, the global environment, and object instances. Extensive experiments on two of the most widely-used 3D dense captioning datasets demonstrate that our proposed method achieves a significant improvement over prior methods.

Fatemeh Pesaran Zadeh, Yoojin Oh, Gunhee Kim

Aligning large VLMs with human preferences is a challenging task, as methods like RLHF and DPO often overfit to textual information or exacerbate hallucinations. Although augmenting negative image samples partially addresses these pitfalls, no prior work has employed listwise preference optimization for VLMs, due to the complexity and cost of constructing listwise image samples. In this work, we propose LPOI, the first object-aware listwise preference optimization developed for reducing hallucinations in VLMs. LPOI identifies and masks a critical object in the image, and then interpolates the masked region between the positive and negative images to form a sequence of incrementally more complete images. The model is trained to rank these images in ascending order of object visibility, effectively reducing hallucinations while retaining visual fidelity. LPOI requires no extra annotations beyond standard pairwise preference data, as it automatically constructs the ranked lists through object masking and interpolation. Comprehensive experiments on MMHalBench, AMBER, and Object HalBench confirm that LPOI outperforms existing preference optimization methods in reducing hallucinations and enhancing VLM performance. We make the code available at https://github.com/fatemehpesaran310/lpoi.

Junyoung Lim, Jaewoo Ahn, Gunhee Kim

Generating accurate, informative, and hallucination-free captions for charts remains challenging for vision language models, primarily due to the lack of large-scale, high-quality datasets of real-world charts. However, existing real-world chart datasets suffer from the inclusion of extraneous information that cannot be inferred from the chart and failure to sufficiently capture structural elements and key insights. Therefore, we introduce ChartCap, a large-scale dataset of 565K real-world chart images paired with type-specific, dense captions that exclude extraneous information and highlight both structural elements and key insights in detail. To build ChartCap, we design a four-stage pipeline that generates captions using only the discernible data from the chart and employ a cycle consistency-based human verification, which accelerates quality control without sacrificing accuracy. Additionally, we propose a novel metric, the Visual Consistency Score, which evaluates caption quality by measuring the similarity between the chart regenerated from a caption and the original chart, independent of reference captions. Extensive experiments confirms that models fine-tuned on ChartCap consistently generate more accurate and informative captions with reduced hallucinations, surpassing both open-source and proprietary models and even human-annotated captions.

Eunkyu Park, Minyeong Kim, Gunhee Kim

Hallucinations pose a significant challenge to the reliability of large vision-language models, making their detection essential for ensuring accuracy in critical applications. Current detection methods often rely on computationally intensive models, leading to high latency and resource demands. Their definitive outcomes also fail to account for real-world scenarios where the line between hallucinated and truthful information is unclear. To address these issues, we propose HalLoc, a dataset designed for efficient, probabilistic hallucination detection. It features 150K token-level annotated samples, including hallucination types, across Visual Question Answering (VQA), instruction-following, and image captioning tasks. This dataset facilitates the development of models that detect hallucinations with graded confidence, enabling more informed user interactions. Additionally, we introduce a baseline model trained on HalLoc, offering low-overhead, concurrent hallucination detection during generation. The model can be seamlessly integrated into existing VLMs, improving reliability while preserving efficiency. The prospect of a robust plug-and-play hallucination detection module opens new avenues for enhancing the trustworthiness of vision-language models in real-world applications. The HalLoc dataset and code are publicly available at: https://github.com/dbsltm/cvpr25_halloc.

Seokwon Song, Taehyun Lee, Jaewoo Ahn et al.

Conceptual combination is a cognitive process that merges basic concepts, enabling the creation of complex expressions. During this process, the properties of combination (e.g., the whiteness of a peeled apple) can be inherited from basic concepts, newly emerge, or be canceled. However, previous studies have evaluated a limited set of properties and have not examined the generative process. To address this gap, we introduce the Conceptual Combination with Property Type dataset (CCPT), which consists of 12.3K annotated triplets of noun phrases, properties, and property types. Using CCPT, we establish three types of tasks to evaluate LLMs for conceptual combination thoroughly. Our key findings are threefold: (1) Our automatic metric grading property emergence and cancellation closely corresponds with human judgments. (2) LLMs, including OpenAI's o1, struggle to generate noun phrases which possess given emergent properties. (3) Our proposed method, inspired by cognitive psychology model that explains how relationships between concepts are formed, improves performances in all generative tasks. The dataset and experimental code are available at https://github.com/seokwon99/CCPT.git.

Yeda Song, Dongwook Lee, Gunhee Kim

Offline reinforcement learning (RL) is a compelling framework for learning optimal policies from past experiences without additional interaction with the environment. Nevertheless, offline RL inevitably faces the problem of distributional shifts, where the states and actions encountered during policy execution may not be in the training dataset distribution. A common solution involves incorporating conservatism into the policy or the value function to safeguard against uncertainties and unknowns. In this work, we focus on achieving the same objectives of conservatism but from a different perspective. We propose COmpositional COnservatism with Anchor-seeking (COCOA) for offline RL, an approach that pursues conservatism in a compositional manner on top of the transductive reparameterization (Netanyahu et al., 2023), which decomposes the input variable (the state in our case) into an anchor and its difference from the original input. Our COCOA seeks both in-distribution anchors and differences by utilizing the learned reverse dynamics model, encouraging conservatism in the compositional input space for the policy or value function. Such compositional conservatism is independent of and agnostic to the prevalent behavioral conservatism in offline RL. We apply COCOA to four state-of-the-art offline RL algorithms and evaluate them on the D4RL benchmark, where COCOA generally improves the performance of each algorithm. The code is available at https://github.com/runamu/compositional-conservatism.

Heeseung Yun, Joonil Na, Gunhee Kim

Sound can convey significant information for spatial reasoning in our daily lives. To endow deep networks with such ability, we address the challenge of dense indoor prediction with sound in both 2D and 3D via cross-modal knowledge distillation. In this work, we propose a Spatial Alignment via Matching (SAM) distillation framework that elicits local correspondence between the two modalities in vision-to-audio knowledge transfer. SAM integrates audio features with visually coherent learnable spatial embeddings to resolve inconsistencies in multiple layers of a student model. Our approach does not rely on a specific input representation, allowing for flexibility in the input shapes or dimensions without performance degradation. With a newly curated benchmark named Dense Auditory Prediction of Surroundings (DAPS), we are the first to tackle dense indoor prediction of omnidirectional surroundings in both 2D and 3D with audio observations. Specifically, for audio-based depth estimation, semantic segmentation, and challenging 3D scene reconstruction, the proposed distillation framework consistently achieves state-of-the-art performance across various metrics and backbone architectures.

Sang Hoon Woo, Sehun Lee, Kang-wook Kim et al.

Spoken dialogue systems increasingly employ large language models (LLMs) to leverage their advanced reasoning capabilities. However, direct application of LLMs in spoken communication often yield suboptimal results due to mismatches between optimal textual and verbal delivery. While existing approaches adapt LLMs to produce speech-friendly outputs, their impact on reasoning performance remains underexplored. In this work, we propose Think-Verbalize-Speak, a framework that decouples reasoning from spoken delivery to preserve the full reasoning capacity of LLMs. Central to our method is verbalizing, an intermediate step that translates thoughts into natural, speech-ready text. We also introduce ReVerT, a latency-efficient verbalizer based on incremental and asynchronous summarization. Experiments across multiple benchmarks show that our method enhances speech naturalness and conciseness with minimal impact on reasoning. The project page with the dataset and the source code is available at https://yhytoto12.github.io/TVS-ReVerT

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.

Taehyun Lee, Seokhee Hong, Jaewoo Ahn et al.

Since the remarkable generation performance of large language models raised ethical and legal concerns, approaches to detect machine-generated text by embedding watermarks are being developed. However, we discover that the existing works fail to function appropriately in code generation tasks due to the task's nature of having low entropy. Extending a logit-modifying watermark method, we propose Selective WatErmarking via Entropy Thresholding (SWEET), which enhances detection ability and mitigates code quality degeneration by removing low-entropy segments at generating and detecting watermarks. Our experiments show that SWEET significantly improves code quality preservation while outperforming all baselines, including post-hoc detection methods, in detecting machine-generated code text. Our code is available in https://github.com/hongcheki/sweet-watermark.

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.

Seohong Park, Jaekyeom Kim, Gunhee Kim

In reinforcement learning, continuous time is often discretized by a time scale $δ$, to which the resulting performance is known to be highly sensitive. In this work, we seek to find a $δ$-invariant algorithm for policy gradient (PG) methods, which performs well regardless of the value of $δ$. We first identify the underlying reasons that cause PG methods to fail as $δ\to 0$, proving that the variance of the PG estimator can diverge to infinity in stochastic environments under a certain assumption of stochasticity. While durative actions or action repetition can be employed to have $δ$-invariance, previous action repetition methods cannot immediately react to unexpected situations in stochastic environments. We thus propose a novel $δ$-invariant method named Safe Action Repetition (SAR) applicable to any existing PG algorithm. SAR can handle the stochasticity of environments by adaptively reacting to changes in states during action repetition. We empirically show that our method is not only $δ$-invariant but also robust to stochasticity, outperforming previous $δ$-invariant approaches on eight MuJoCo environments with both deterministic and stochastic settings. Our code is available at https://vision.snu.ac.kr/projects/sar.

Chris Dongjoo Kim, Jinseo Jeong, Sangwoo Moon et al.

Continually learning in the real world must overcome many challenges, among which noisy labels are a common and inevitable issue. In this work, we present a repla-ybased continual learning framework that simultaneously addresses both catastrophic forgetting and noisy labels for the first time. Our solution is based on two observations; (i) forgetting can be mitigated even with noisy labels via self-supervised learning, and (ii) the purity of the replay buffer is crucial. Building on this regard, we propose two key components of our method: (i) a self-supervised replay technique named Self-Replay which can circumvent erroneous training signals arising from noisy labeled data, and (ii) the Self-Centered filter that maintains a purified replay buffer via centrality-based stochastic graph ensembles. The empirical results on MNIST, CIFAR-10, CIFAR-100, and WebVision with real-world noise demonstrate that our framework can maintain a highly pure replay buffer amidst noisy streamed data while greatly outperforming the combinations of the state-of-the-art continual learning and noisy label learning methods. The source code is available at http://vision.snu.ac.kr/projects/SPR

Insu Jeon, Wonkwang Lee, Myeongjang Pyeon et al.

We propose a new GAN-based unsupervised model for disentangled representation learning. The new model is discovered in an attempt to utilize the Information Bottleneck (IB) framework to the optimization of GAN, thereby named IB-GAN. The architecture of IB-GAN is partially similar to that of InfoGAN but has a critical difference; an intermediate layer of the generator is leveraged to constrain the mutual information between the input and the generated output. The intermediate stochastic layer can serve as a learnable latent distribution that is trained with the generator jointly in an end-to-end fashion. As a result, the generator of IB-GAN can harness the latent space in a disentangled and interpretable manner. With the experiments on dSprites and Color-dSprites dataset, we demonstrate that IB-GAN achieves competitive disentanglement scores to those of state-of-the-art \b{eta}-VAEs and outperforms InfoGAN. Moreover, the visual quality and the diversity of samples generated by IB-GAN are often better than those by \b{eta}-VAEs and Info-GAN in terms of FID score on CelebA and 3D Chairs dataset.

Jaeyeon Kim, Heeseung Yun, Gunhee Kim

Spatial audio is essential for enhancing the immersiveness of audio-visual experiences, yet its production typically demands complex recording systems and specialized expertise. In this work, we address a novel problem of generating first-order ambisonics, a widely used spatial audio format, directly from silent videos. To support this task, we introduce YT-Ambigen, a dataset comprising 102K 5-second YouTube video clips paired with corresponding first-order ambisonics. We also propose new evaluation metrics to assess the spatial aspect of generated audio based on audio energy maps and saliency metrics. Furthermore, we present Video-to-Spatial Audio Generation (ViSAGe), an end-to-end framework that generates first-order ambisonics from silent video frames by leveraging CLIP visual features, autoregressive neural audio codec modeling with both directional and visual guidance. Experimental results demonstrate that ViSAGe produces plausible and coherent first-order ambisonics, outperforming two-stage approaches consisting of video-to-audio generation and audio spatialization. Qualitative examples further illustrate that ViSAGe generates temporally aligned high-quality spatial audio that adapts to viewpoint changes.

Jaehyeon Son, Soochan Lee, Gunhee Kim · gatech

Recent studies have shown that Transformers can perform in-context reinforcement learning (RL) by imitating existing RL algorithms, enabling sample-efficient adaptation to unseen tasks without parameter updates. However, these models also inherit the suboptimal behaviors of the RL algorithms they imitate. This issue primarily arises due to the gradual update rule employed by those algorithms. Model-based planning offers a promising solution to this limitation by allowing the models to simulate potential outcomes before taking action, providing an additional mechanism to deviate from the suboptimal behavior. Rather than learning a separate dynamics model, we propose Distillation for In-Context Planning (DICP), an in-context model-based RL framework where Transformers simultaneously learn environment dynamics and improve policy in-context. We evaluate DICP across a range of discrete and continuous environments, including Darkroom variants and Meta-World. Our results show that DICP achieves state-of-the-art performance while requiring significantly fewer environment interactions than baselines, which include both model-free counterparts and existing meta-RL methods.

Dayoon Ko, Jihyuk Kim, Haeju Park et al.

Large reasoning models (LRMs) have demonstrated strong performance in complex, multi-step reasoning tasks. Existing methods enhance LRMs by sequentially integrating external knowledge retrieval; models iteratively generate queries, retrieve external information, and progressively reason over this information. However, purely sequential querying increases inference latency and context length, diminishing coherence and potentially reducing accuracy. To address these limitations, we introduce HDS-QA (Hybrid Deep Search QA), a synthetic dataset automatically generated from Natural Questions, explicitly designed to train LRMs to distinguish parallelizable from sequential queries. HDS-QA comprises hybrid-hop questions that combine parallelizable independent subqueries (executable simultaneously) and sequentially dependent subqueries (requiring step-by-step resolution), along with synthetic reasoning-querying-retrieval paths involving parallel queries. We fine-tune an LRM using HDS-QA, naming the model HybridDeepSearcher, which outperforms state-of-the-art baselines across multiple benchmarks, notably achieving +15.9 and +11.5 F1 on FanOutQA and a subset of BrowseComp, respectively, both requiring comprehensive and exhaustive search. Experimental results highlight two key advantages: HybridDeepSearcher reaches comparable accuracy with fewer search turns, significantly reducing inference latency, and it effectively scales as more turns are permitted. These results demonstrate the efficiency, scalability, and effectiveness of explicitly training LRMs to leverage hybrid parallel and sequential querying.

Chao-Han Huck Yang, Sreyan Ghosh, Qing Wang et al.

We present Task 5 of the DCASE 2025 Challenge: an Audio Question Answering (AQA) benchmark spanning multiple domains of sound understanding. This task defines three QA subsets (Bioacoustics, Temporal Soundscapes, and Complex QA) to test audio-language models on interactive question-answering over diverse acoustic scenes. We describe the dataset composition (from marine mammal calls to soundscapes and complex real-world clips), the evaluation protocol (top-1 accuracy with answer-shuffling robustness), and baseline systems (Qwen2-Audio-7B, AudioFlamingo 2, Gemini-2-Flash). Preliminary results on the development set are compared, showing strong variation across models and subsets. This challenge aims to advance the audio understanding and reasoning capabilities of audio-language models toward human-level acuity, which are crucial for enabling AI agents to perceive and interact about the world effectively.

Insu Jeon, Youngjin Park, Gunhee Kim

Learning to infer the conditional posterior model is a key step for robust meta-learning. This paper presents a new Bayesian meta-learning approach called Neural Variational Dropout Processes (NVDPs). NVDPs model the conditional posterior distribution based on a task-specific dropout; a low-rank product of Bernoulli experts meta-model is utilized for a memory-efficient mapping of dropout rates from a few observed contexts. It allows for a quick reconfiguration of a globally learned and shared neural network for new tasks in multi-task few-shot learning. In addition, NVDPs utilize a novel prior conditioned on the whole task data to optimize the conditional \textit{dropout} posterior in the amortized variational inference. Surprisingly, this enables the robust approximation of task-specific dropout rates that can deal with a wide range of functional ambiguities and uncertainties. We compared the proposed method with other meta-learning approaches in the few-shot learning tasks such as 1D stochastic regression, image inpainting, and classification. The results show the excellent performance of NVDPs.

Seungyoon Woo, Junhyeog Yun, Gunhee Kim

Neural Fields (NF) have gained prominence as a versatile framework for complex data representation. This work unveils a new problem setting termed \emph{Meta-Continual Learning of Neural Fields} (MCL-NF) and introduces a novel strategy that employs a modular architecture combined with optimization-based meta-learning. Focused on overcoming the limitations of existing methods for continual learning of neural fields, such as catastrophic forgetting and slow convergence, our strategy achieves high-quality reconstruction with significantly improved learning speed. We further introduce Fisher Information Maximization loss for neural radiance fields (FIM-NeRF), which maximizes information gains at the sample level to enhance learning generalization, with proved convergence guarantee and generalization bound. We perform extensive evaluations across image, audio, video reconstruction, and view synthesis tasks on six diverse datasets, demonstrating our method's superiority in reconstruction quality and speed over existing MCL and CL-NF approaches. Notably, our approach attains rapid adaptation of neural fields for city-scale NeRF rendering with reduced parameter requirement.

Jinyoung Kim, Dayoon Ko, Gunhee Kim

In the rapidly evolving landscape of language, resolving new linguistic expressions in continuously updating knowledge bases remains a formidable challenge. This challenge becomes critical in retrieval-augmented generation (RAG) with knowledge bases, as emerging expressions hinder the retrieval of relevant documents, leading to generator hallucinations. To address this issue, we introduce a novel task aimed at resolving emerging mentions to dynamic entities and present DynamicER benchmark. Our benchmark includes dynamic entity mention resolution and entity-centric knowledge-intensive QA task, evaluating entity linking and RAG model's adaptability to new expressions, respectively. We discovered that current entity linking models struggle to link these new expressions to entities. Therefore, we propose a temporal segmented clustering method with continual adaptation, effectively managing the temporal dynamics of evolving entities and emerging mentions. Extensive experiments demonstrate that our method outperforms existing baselines, enhancing RAG model performance on QA task with resolved mentions.

Jinseo Jeong, Junseo Koo, Qimeng Zhang et al.

Existing NeRF-based inverse rendering methods suppose that scenes are exclusively illuminated by distant light sources, neglecting the potential influence of emissive sources within a scene. In this work, we confront this limitation using LDR multi-view images captured with emissive sources turned on and off. Two key issues must be addressed: 1) ambiguity arising from the limited dynamic range along with unknown lighting details, and 2) the expensive computational cost in volume rendering to backtrace the paths leading to final object colors. We present a novel approach, ESR-NeRF, leveraging neural networks as learnable functions to represent ray-traced fields. By training networks to satisfy light transport segments, we regulate outgoing radiances, progressively identifying emissive sources while being aware of reflection areas. The results on scenes encompassing emissive sources with various properties demonstrate the superiority of ESR-NeRF in qualitative and quantitative ways. Our approach also extends its applicability to the scenes devoid of emissive sources, achieving lower CD metrics on the DTU dataset.

Jaewoo Ahn, Junseo Kim, Heeseung Yun et al. · gatech

GUI agents powered by LLMs show promise in interacting with diverse digital environments. Among these, video games offer a valuable testbed due to their varied interfaces, with adventure games posing additional challenges through complex, narrative-driven interactions. Existing game benchmarks, however, lack diversity and rarely evaluate agents on completing entire storylines. To address this, we introduce FlashAdventure, a benchmark of 34 Flash-based adventure games designed to test full story arc completion and tackle the observation-behavior gap: the challenge of remembering and acting on earlier gameplay information. We also propose CUA-as-a-Judge, an automated gameplay evaluator, and COAST, an agentic framework leveraging long-term clue memory to better plan and solve sequential tasks. Experiments show current GUI agents struggle with full story arcs, while COAST improves milestone completion by bridging the observation-behavior gap. Nonetheless, a marked discrepancy between humans and best-performing agents warrants continued research efforts to narrow this divide.

Eunkyu Park, Wesley Hanwen Deng, Gunhee Kim et al. · allen-ai, cmu

Chain-of-Thought (CoT) prompting helps models think step by step. But what happens when they must see, understand, and judge-all at once? In visual tasks grounded in social context, where bridging perception with norm-grounded judgments is essential, flat CoT often breaks down. We introduce Cognitive Chain-of-Thought (CoCoT), a prompting strategy that scaffolds VLM reasoning through three cognitively inspired stages: perception, situation, and norm. Our experiments show that, across multiple multimodal benchmarks (including intent disambiguation, commonsense reasoning, and safety), CoCoT consistently outperforms CoT and direct prompting (+8\% on average). Our findings demonstrate that cognitively grounded reasoning stages enhance interpretability and social awareness in VLMs, paving the way for safer and more reliable multimodal systems.

Dayoon Ko, Jinyoung Kim, Sohyeon Kim et al.

Dense retrievers encode texts into embeddings to efficiently retrieve relevant documents from large databases in response to user queries. However, real-world corpora continually evolve, leading to a shift from the original training distribution of the retriever. Without timely updates or retraining, indexing newly emerging documents can degrade retrieval performance for future queries. Thus, identifying when a dense retriever requires an update is critical for maintaining robust retrieval systems. In this paper, we propose a novel task of predicting whether a corpus is out-of-distribution (OOD) relative to a dense retriever before indexing. Addressing this task allows us to proactively manage retriever updates, preventing potential retrieval failures. We introduce GradNormIR, an unsupervised approach that leverages gradient norms to detect OOD corpora effectively. Experiments on the BEIR benchmark demonstrate that GradNormIR enables timely updates of dense retrievers in evolving document collections, significantly enhancing retrieval robustness and efficiency.

Junseo Koo, Jinseo Jeong, Gunhee Kim

The recent introduction of 3D Gaussian Splatting (3DGS) has significantly advanced novel view synthesis. Several studies have further improved the rendering quality of 3DGS, yet they still exhibit noticeable visual discrepancies when synthesizing views at sampling rates unseen during training. Specifically, they suffer from (i) erosion-induced blurring artifacts when zooming in and (ii) dilation-induced staircase artifacts when zooming out. We speculate that these artifacts arise from the fundamental limitation of the alpha blending adopted in 3DGS methods. Instead of the conventional alpha blending that computes alpha and transmittance as scalar quantities over a pixel, we propose to replace it with our novel Gaussian Blending that treats alpha and transmittance as spatially varying distributions. Thus, transmittances can be updated considering the spatial distribution of alpha values across the pixel area, allowing nearby background splats to contribute to the final rendering. Our Gaussian Blending maintains real-time rendering speed and requires no additional memory cost, while being easily integrated as a drop-in replacement into existing 3DGS-based or other NVS frameworks. Extensive experiments demonstrate that Gaussian Blending effectively captures fine details at various sampling rates unseen during training, consistently outperforming existing novel view synthesis models across both unseen and seen sampling rates.

Heeseung Yun, Joonil Na, Jaeyeon Kim et al.

People continuously perceive and interact with their surroundings based on underlying intentions that drive their exploration and behaviors. While research in egocentric user and scene understanding has focused primarily on motion and contact-based interaction, forecasting human visual perception itself remains less explored despite its fundamental role in guiding human actions and its implications for AR/VR and assistive technologies. We address the challenge of egocentric 3D visual span forecasting, predicting where a person's visual perception will focus next within their three-dimensional environment. To this end, we propose EgoSpanLift, a novel method that transforms egocentric visual span forecasting from 2D image planes to 3D scenes. EgoSpanLift converts SLAM-derived keypoints into gaze-compatible geometry and extracts volumetric visual span regions. We further combine EgoSpanLift with 3D U-Net and unidirectional transformers, enabling spatio-temporal fusion to efficiently predict future visual span in the 3D grid. In addition, we curate a comprehensive benchmark from raw egocentric multisensory data, creating a testbed with 364.6K samples for 3D visual span forecasting. Our approach outperforms competitive baselines for egocentric 2D gaze anticipation and 3D localization while achieving comparable results even when projected back onto 2D image planes without additional 2D-specific training.

Jaeyeon Kim, Heeseung Yun, Sang Hoon Woo et al.

Large audio language models (LALMs) extend language understanding into the auditory domain, yet their ability to perform low-level listening, such as pitch and duration detection, remains underexplored. However, low-level listening is critical for real-world, out-of-distribution tasks where models must reason about unfamiliar sounds based on fine-grained acoustic cues. To address this gap, we introduce the World-of-Whale benchmark (WoW-Bench) to evaluate low-level auditory perception and cognition using marine mammal vocalizations. WoW-bench is composed of a Perception benchmark for categorizing novel sounds and a Cognition benchmark, inspired by Bloom's taxonomy, to assess the abilities to remember, understand, apply, and analyze sound events. For the Cognition benchmark, we additionally introduce distractor questions to evaluate whether models are truly solving problems through listening rather than relying on other heuristics. Experiments with state-of-the-art LALMs show performance far below human levels, indicating a need for stronger auditory grounding in LALMs.

Junhyeog Yun, Minui Hong, Gunhee Kim

Neural fields provide a memory-efficient representation of data, which can effectively handle diverse modalities and large-scale data. However, learning to map neural fields often requires large amounts of training data and computations, which can be limited to resource-constrained edge devices. One approach to tackle this limitation is to leverage Federated Meta-Learning (FML), but traditional FML approaches suffer from privacy leakage. To address these issues, we introduce a novel FML approach called FedMeNF. FedMeNF utilizes a new privacy-preserving loss function that regulates privacy leakage in the local meta-optimization. This enables the local meta-learner to optimize quickly and efficiently without retaining the client's private data. Our experiments demonstrate that FedMeNF achieves fast optimization speed and robust reconstruction performance, even with few-shot or non-IID data across diverse data modalities, while preserving client data privacy.

Mijeong Kim, Gunhee Kim, Jungyoon Choi et al.

We introduce PhysGaia, a novel physics-aware dataset specifically designed for Dynamic Novel View Synthesis (DyNVS), encompassing both structured objects and unstructured physical phenomena. Unlike existing datasets that primarily focus on photorealistic reconstruction, PhysGaia is created to actively support physics-aware dynamic scene modeling. Our dataset provides complex dynamic scenarios with rich interactions among multiple objects, where they realistically collide with each other and exchange forces. Furthermore, it contains a diverse range of physical materials, such as liquid, gas, viscoelastic substance, and textile, which moves beyond the rigid bodies prevalent in existing datasets. All scenes in PhysGaia are faithfully generated to strictly adhere to physical laws, leveraging carefully selected material-specific physics solvers. To enable quantitative evaluation of physical modeling, our dataset provides essential ground-truth information, including 3D particle trajectories and physics parameters, e.g., viscosity. To facilitate research adoption, we also provide essential integration pipelines for using state-of-the-art DyNVS models with our dataset and report their results. By addressing the critical lack of datasets for physics-aware modeling, PhysGaia will significantly advance research in dynamic view synthesis, physics-based scene understanding, and deep learning models integrated with physical simulation -- ultimately enabling more faithful reconstruction and interpretation of complex dynamic scenes. Our datasets and codes are available in the project website, http://cvlab.snu.ac.kr/research/PhysGaia.