Yen-Ling Kuo

Xi Wang, Gen Li, Yen-Ling Kuo et al. · eth-zurich

We present a method for inferring diverse 3D models of human-object interactions from images. Reasoning about how humans interact with objects in complex scenes from a single 2D image is a challenging task given ambiguities arising from the loss of information through projection. In addition, modeling 3D interactions requires the generalization ability towards diverse object categories and interaction types. We propose an action-conditioned modeling of interactions that allows us to infer diverse 3D arrangements of humans and objects without supervision on contact regions or 3D scene geometry. Our method extracts high-level commonsense knowledge from large language models (such as GPT-3), and applies them to perform 3D reasoning of human-object interactions. Our key insight is priors extracted from large language models can help in reasoning about human-object contacts from textural prompts only. We quantitatively evaluate the inferred 3D models on a large human-object interaction dataset and show how our method leads to better 3D reconstructions. We further qualitatively evaluate the effectiveness of our method on real images and demonstrate its generalizability towards interaction types and object categories.

Razvan-George Pasca, Alexey Gavryushin, Muhammad Hamza et al. · eth-zurich

We study object interaction anticipation in egocentric videos. This task requires an understanding of the spatio-temporal context formed by past actions on objects, coined action context. We propose TransFusion, a multimodal transformer-based architecture. It exploits the representational power of language by summarizing the action context. TransFusion leverages pre-trained image captioning and vision-language models to extract the action context from past video frames. This action context together with the next video frame is processed by the multimodal fusion module to forecast the next object interaction. Our model enables more efficient end-to-end learning. The large pre-trained language models add common sense and a generalisation capability. Experiments on Ego4D and EPIC-KITCHENS-100 show the effectiveness of our multimodal fusion model. They also highlight the benefits of using language-based context summaries in a task where vision seems to suffice. Our method outperforms state-of-the-art approaches by 40.4% in relative terms in overall mAP on the Ego4D test set. We validate the effectiveness of TransFusion via experiments on EPIC-KITCHENS-100. Video and code are available at https://eth-ait.github.io/transfusion-proj/.

Kunal Jha, Tuan Anh Le, Chuanyang Jin et al.

Multi-agent interactions, such as communication, teaching, and bluffing, often rely on higher-order social inference, i.e., understanding how others infer oneself. Such intricate reasoning can be effectively modeled through nested multi-agent reasoning. Nonetheless, the computational complexity escalates exponentially with each level of reasoning, posing a significant challenge. However, humans effortlessly perform complex social inferences as part of their daily lives. To bridge the gap between human-like inference capabilities and computational limitations, we propose a novel approach: leveraging neural networks to amortize high-order social inference, thereby expediting nested multi-agent reasoning. We evaluate our method in two challenging multi-agent interaction domains. The experimental results demonstrate that our method is computationally efficient while exhibiting minimal degradation in accuracy.

Nathan Cloos, Meagan Jens, Michelangelo Naim et al.

Humans solve problems by following existing rules and procedures, and also by leaps of creativity to redefine those rules and objectives. To probe these abilities, we developed a new benchmark based on the game Baba Is You where an agent manipulates both objects in the environment and rules, represented by movable tiles with words written on them, to reach a specified goal and win the game. We test three state-of-the-art multi-modal large language models (OpenAI GPT-4o, Google Gemini-1.5-Pro and Gemini-1.5-Flash) and find that they fail dramatically when generalization requires that the rules of the game must be manipulated and combined.

Haojun Shi, Suyu Ye, Xinyu Fang et al.

Understanding people's social interactions in complex real-world scenarios often relies on intricate mental reasoning. To truly understand how and why people interact with one another, we must infer the underlying mental states that give rise to the social interactions, i.e., Theory of Mind reasoning in multi-agent interactions. Additionally, social interactions are often multi-modal -- we can watch people's actions, hear their conversations, and/or read about their past behaviors. For AI systems to successfully and safely interact with people in real-world environments, they also need to understand people's mental states as well as their inferences about each other's mental states based on multi-modal information about their interactions. For this, we introduce MuMA-ToM, a Multi-modal Multi-Agent Theory of Mind benchmark. MuMA-ToM is the first multi-modal Theory of Mind benchmark that evaluates mental reasoning in embodied multi-agent interactions. In MuMA-ToM, we provide video and text descriptions of people's multi-modal behavior in realistic household environments. Based on the context, we then ask questions about people's goals, beliefs, and beliefs about others' goals. We validated MuMA-ToM in a human experiment and provided a human baseline. We also proposed a novel multi-modal, multi-agent ToM model, LIMP (Language model-based Inverse Multi-agent Planning). Our experimental results show that LIMP significantly outperforms state-of-the-art methods, including large multi-modal models (e.g., GPT-4o, Gemini-1.5 Pro) and a recent multi-modal ToM model, BIP-ALM.

Alex Ning, Yen-Ling Kuo, Gabe Gomes

Latent reasoning represents a new development in Transformer language models that has shown potential in compressing reasoning lengths compared to chain-of-thought reasoning. By directly passing the information-rich previous final latent state into the next sequence, latent reasoning removes the restriction to human language tokens as the medium for reasoning. We develop adaptive-length latent reasoning models and introduce a post-SFT reinforcement-learning methodology to optimize latent reasoning length by minimizing reasoning length while maintaining accuracy. This, in turn, further reduces compute usage and raises the bar on the compressive capabilities of latent reasoning models. Experiments on the Llama 3.2 1B model and the GSM8K-Aug dataset show a $52\%$ drop in total reasoning length with no penalty to accuracy. In future work, we plan to extend to additional models and datasets, analyze relationships between training coefficients, experiment with architecture variations, and continue our knowledge distillation for latent reasoning SFT efforts. We make our code and pretrained weights available at https://github.com/apning/adaptive-latent-reasoning.

Alex Ning, Vainateya Rangaraju, Yen-Ling Kuo

Large language models (LLMs) achieve state-of-the-art results across many natural language tasks, but their internal mechanisms remain difficult to interpret. In this work, we extract, process, and visualize latent state geometries in Transformer-based language models through dimensionality reduction. We capture layerwise activations at multiple points within Transformer blocks and enable systematic analysis through Principal Component Analysis (PCA) and Uniform Manifold Approximation and Projection (UMAP). We demonstrate experiments on GPT-2 and LLaMa models, where we uncover interesting geometric patterns in latent space. Notably, we identify a clear separation between attention and MLP component outputs across intermediate layers, a pattern not documented in prior work to our knowledge. We also characterize the high norm of latent states at the initial sequence position and visualize the layerwise evolution of latent states. Additionally, we demonstrate the high-dimensional helical structure of GPT-2's positional embeddings and the sequence-wise geometric patterns in LLaMa. We make our code available at https://github.com/Vainateya/Feature_Geometry_Visualization.

Bo Wang, Dingwei Tan, Yen-Ling Kuo et al.

Imagine searching a collection of coins for quarters ($0.25$), dimes ($0.10$), nickels ($0.05$), and pennies ($0.01$)-a hybrid foraging task where observers look for multiple instances of multiple target types. In such tasks, how do target values and their prevalence influence foraging and eye movement behaviors (e.g., should you prioritize rare quarters or common nickels)? To explore this, we conducted human psychophysics experiments, revealing that humans are proficient reward foragers. Their eye fixations are drawn to regions with higher average rewards, fixation durations are longer on more valuable targets, and their cumulative rewards exceed chance, approaching the upper bound of optimal foragers. To probe these decision-making processes of humans, we developed a transformer-based Visual Forager (VF) model trained via reinforcement learning. Our VF model takes a series of targets, their corresponding values, and the search image as inputs, processes the images using foveated vision, and produces a sequence of eye movements along with decisions on whether to collect each fixated item. Our model outperforms all baselines, achieves cumulative rewards comparable to those of humans, and approximates human foraging behavior in eye movements and foraging biases within time-limited environments. Furthermore, stress tests on out-of-distribution tasks with novel targets, unseen values, and varying set sizes demonstrate the VF model's effective generalization. Our work offers valuable insights into the relationship between eye movements and decision-making, with our model serving as a powerful tool for further exploration of this connection. All data, code, and models are available at https://github.com/ZhangLab-DeepNeuroCogLab/visual-forager.

Suleyman Ozdel, Yao Rong, Berat Mert Albaba et al. · eth-zurich

Humans utilize their gaze to concentrate on essential information while perceiving and interpreting intentions in videos. Incorporating human gaze into computational algorithms can significantly enhance model performance in video understanding tasks. In this work, we address a challenging and innovative task in video understanding: predicting the actions of an agent in a video based on a partial video. We introduce the Gaze-guided Action Anticipation algorithm, which establishes a visual-semantic graph from the video input. Our method utilizes a Graph Neural Network to recognize the agent's intention and predict the action sequence to fulfill this intention. To assess the efficiency of our approach, we collect a dataset containing household activities generated in the VirtualHome environment, accompanied by human gaze data of viewing videos. Our method outperforms state-of-the-art techniques, achieving a 7\% improvement in accuracy for 18-class intention recognition. This highlights the efficiency of our method in learning important features from human gaze data.

Suleyman Ozdel, Yao Rong, Berat Mert Albaba et al. · eth-zurich

Eye-tracking applications that utilize the human gaze in video understanding tasks have become increasingly important. To effectively automate the process of video analysis based on eye-tracking data, it is important to accurately replicate human gaze behavior. However, this task presents significant challenges due to the inherent complexity and ambiguity of human gaze patterns. In this work, we introduce a novel method for simulating human gaze behavior. Our approach uses a transformer-based reinforcement learning algorithm to train an agent that acts as a human observer, with the primary role of watching videos and simulating human gaze behavior. We employed an eye-tracking dataset gathered from videos generated by the VirtualHome simulator, with a primary focus on activity recognition. Our experimental results demonstrate the effectiveness of our gaze prediction method by highlighting its capability to replicate human gaze behavior and its applicability for downstream tasks where real human-gaze is used as input.

Tongxuan Tian, Xuhui Kang, Yen-Ling Kuo

Grounding object affordance is fundamental to robotic manipulation as it establishes the critical link between perception and action among interacting objects. However, prior works predominantly focus on predicting single-object affordance, overlooking the fact that most real-world interactions involve relationships between pairs of objects. In this work, we address the challenge of object-to-object affordance grounding under limited data contraints. Inspired by recent advances in few-shot learning with 2D vision foundation models, we propose a novel one-shot 3D object-to-object affordance learning approach for robotic manipulation. Semantic features from vision foundation models combined with point cloud representation for geometric understanding enable our one-shot learning pipeline to generalize effectively to novel objects and categories. We further integrate our 3D affordance representation with large language models (LLMs) for robotics manipulation, significantly enhancing LLMs' capability to comprehend and reason about object interactions when generating task-specific constraint functions. Our experiments on 3D object-to-object affordance grounding and robotic manipulation demonstrate that our O$^3$Afford significantly outperforms existing baselines in terms of both accuracy and generalization capability.

Xuhui Kang, Sung-Wook Lee, Haolin Liu et al.

The ability to adapt to physical actions and constraints in an environment is crucial for embodied agents (e.g., robots) to effectively collaborate with humans. Such physically grounded human-AI collaboration must account for the increased complexity of the continuous state-action space and constrained dynamics caused by physical constraints. In this paper, we introduce Moving Out, a new human-AI collaboration benchmark that resembles a wide range of collaboration modes affected by physical attributes and constraints, such as moving heavy items together and maintaining consistent actions to move a big item around a corner. Using Moving Out, we designed two tasks and collected human-human interaction data to evaluate models' abilities to adapt to diverse human behaviors and unseen physical attributes. To address the challenges in physical environments, we propose a novel method, BASS (Behavior Augmentation, Simulation, and Selection), to enhance the diversity of agents and their understanding of the outcome of actions. Our experiments show that BASS outperforms state-of-the-art models in AI-AI and human-AI collaboration. The project page is available at https://live-robotics-uva.github.io/movingout_ai/.

Varun Reddy, Yen-Ling Kuo

Large language models (LLMs) exhibit strong performance on factual recall and general reasoning but struggle to adapt to user-specific, commonsense knowledge, a challenge particularly acute in small-parameter settings where computational efficiency is prioritized. We introduce CaseEdit, a new dataset and generation pipeline for evaluating localized, personalized commonsense knowledge editing in small LLMs to address this. Built upon the ATOMIC20/20 commonsense graph, CaseEdit uses a multi-stage inference process to generate both typical and atypical contextual edits for household objects, paired with targeted evaluation questions across four axes: reliability, generalization, locality, and portability. We evaluate established knowledge editing methods using CaseEdit and demonstrate that AlphaEdit, a technique employing null-space projection to minimize interference with unrelated knowledge, consistently outperforms other methods when applied to an LLaMA 3.2 3B model, even in scalability tests, showing minimal ripple effects. Our results indicate that using CaseEdit with effective editing techniques like AlphaEdit allows small models to internalize high-quality, context-sensitive common-sense knowledge, paving the way for lightweight, personalized assistants.

Jie Jing, Yongjian Huang, Serena J. -W. Wang et al.

We introduce intra-class memorability, where certain images within the same class are more memorable than others despite shared category characteristics. To investigate what features make one object instance more memorable than others, we design and conduct human behavior experiments, where participants are shown a series of images, and they must identify when the current image matches the image presented a few steps back in the sequence. To quantify memorability, we propose the Intra-Class Memorability score (ICMscore), a novel metric that incorporates the temporal intervals between repeated image presentations into its calculation. Furthermore, we curate the Intra-Class Memorability Dataset (ICMD), comprising over 5,000 images across ten object classes with their ICMscores derived from 2,000 participants' responses. Subsequently, we demonstrate the usefulness of ICMD by training AI models on this dataset for various downstream tasks: memorability prediction, image recognition, continual learning, and memorability-controlled image editing. Surprisingly, high-ICMscore images impair AI performance in image recognition and continual learning tasks, while low-ICMscore images improve outcomes in these tasks. Additionally, we fine-tune a state-of-the-art image diffusion model on ICMD image pairs with and without masked semantic objects. The diffusion model can successfully manipulate image elements to enhance or reduce memorability. Our contributions open new pathways in understanding intra-class memorability by scrutinizing fine-grained visual features behind the most and least memorable images and laying the groundwork for real-world applications in computer vision. We will release all code, data, and models publicly.

Chuanyang Jin, Yutong Wu, Jing Cao et al.

Theory of Mind (ToM), the ability to understand people's mental states, is an essential ingredient for developing machines with human-level social intelligence. Recent machine learning models, particularly large language models, seem to show some aspects of ToM understanding. However, existing ToM benchmarks use unimodal datasets - either video or text. Human ToM, on the other hand, is more than video or text understanding. People can flexibly reason about another person's mind based on conceptual representations (e.g., goals, beliefs, plans) extracted from any available data. To address this, we introduce a multimodal Theory of Mind question answering (MMToM-QA) benchmark. MMToM-QA comprehensively evaluates machine ToM both on multimodal data and on different kinds of unimodal data about a person's activity in a household environment. To engineer multimodal ToM capacity, we propose a novel method, BIP-ALM (Bayesian Inverse Planning Accelerated by Language Models). BIP-ALM extracts unified representations from multimodal data and utilizes language models for scalable Bayesian inverse planning. We conducted a systematic comparison of human performance, BIP-ALM, and state-of-the-art models, including GPT-4. The experiments demonstrate that large language models and large multimodal models still lack robust ToM capacity. BIP-ALM, on the other hand, shows promising results, by leveraging the power of both model-based mental inference and language models.

Justin Lidard, Oswin So, Yanxia Zhang et al.

Interactions between road agents present a significant challenge in trajectory prediction, especially in cases involving multiple agents. Because existing diversity-aware predictors do not account for the interactive nature of multi-agent predictions, they may miss these important interaction outcomes. In this paper, we propose NashFormer, a framework for trajectory prediction that leverages game-theoretic inverse reinforcement learning to improve coverage of multi-modal predictions. We use a training-time game-theoretic analysis as an auxiliary loss resulting in improved coverage and accuracy without presuming a taxonomy of actions for the agents. We demonstrate our approach on the interactive split of the Waymo Open Motion Dataset, including four subsets involving scenarios with high interaction complexity. Experiment results show that our predictor produces accurate predictions while covering $33\%$ more potential interactions versus a baseline model.

Ravi Tejwani, Yen-Ling Kuo, Tianmin Shu et al.

Much of what we do as humans is engage socially with other agents, a skill that robots must also eventually possess. We demonstrate that a rich theory of social interactions originating from microsociology and economics can be formalized by extending a nested MDP where agents reason about arbitrary functions of each other's hidden rewards. This extended Social MDP allows us to encode the five basic interactions that underlie microsociology: cooperation, conflict, coercion, competition, and exchange. The result is a robotic agent capable of executing social interactions zero-shot in new environments; like humans it can engage socially in novel ways even without a single example of that social interaction. Moreover, the judgments of these Social MDPs align closely with those of humans when considering which social interaction is taking place in an environment. This method both sheds light on the nature of social interactions, by providing concrete mathematical definitions, and brings rich social interactions into a mathematical framework that has proven to be natural for robotics, MDPs.

Yen-Ling Kuo, Xin Huang, Andrei Barbu et al.

Language allows humans to build mental models that interpret what is happening around them resulting in more accurate long-term predictions. We present a novel trajectory prediction model that uses linguistic intermediate representations to forecast trajectories, and is trained using trajectory samples with partially-annotated captions. The model learns the meaning of each of the words without direct per-word supervision. At inference time, it generates a linguistic description of trajectories which captures maneuvers and interactions over an extended time interval. This generated description is used to refine predictions of the trajectories of multiple agents. We train and validate our model on the Argoverse dataset, and demonstrate improved accuracy results in trajectory prediction. In addition, our model is more interpretable: it presents part of its reasoning in plain language as captions, which can aid model development and can aid in building confidence in the model before deploying it.

Christopher Wang, Candace Ross, Yen-Ling Kuo et al.

Children acquire their native language with apparent ease by observing how language is used in context and attempting to use it themselves. They do so without laborious annotations, negative examples, or even direct corrections. We take a step toward robots that can do the same by training a grounded semantic parser, which discovers latent linguistic representations that can be used for the execution of natural-language commands. In particular, we focus on the difficult domain of commands with a temporal aspect, whose semantics we capture with Linear Temporal Logic, LTL. Our parser is trained with pairs of sentences and executions as well as an executor. At training time, the parser hypothesizes a meaning representation for the input as a formula in LTL. Three competing pressures allow the parser to discover meaning from language. First, any hypothesized meaning for a sentence must be permissive enough to reflect all the annotated execution trajectories. Second, the executor -- a pretrained end-to-end LTL planner -- must find that the observe trajectories are likely executions of the meaning. Finally, a generator, which reconstructs the original input, encourages the model to find representations that conserve knowledge about the command. Together these ensure that the meaning is neither too general nor too specific. Our model generalizes well, being able to parse and execute both machine-generated and human-generated commands, with near-equal accuracy, despite the fact that the human-generated sentences are much more varied and complex with an open lexicon. The approach presented here is not specific to LTL: it can be applied to any domain where sentence meanings can be hypothesized and an executor can verify these meanings, thus opening the door to many applications for robotic agents.

Yen-Ling Kuo, Boris Katz, Andrei Barbu

Humans are remarkably flexible when understanding new sentences that include combinations of concepts they have never encountered before. Recent work has shown that while deep networks can mimic some human language abilities when presented with novel sentences, systematic variation uncovers the limitations in the language-understanding abilities of networks. We demonstrate that these limitations can be overcome by addressing the generalization challenges in the gSCAN dataset, which explicitly measures how well an agent is able to interpret novel linguistic commands grounded in vision, e.g., novel pairings of adjectives and nouns. The key principle we employ is compositionality: that the compositional structure of networks should reflect the compositional structure of the problem domain they address, while allowing other parameters to be learned end-to-end. We build a general-purpose mechanism that enables agents to generalize their language understanding to compositional domains. Crucially, our network has the same state-of-the-art performance as prior work while generalizing its knowledge when prior work does not. Our network also provides a level of interpretability that enables users to inspect what each part of networks learns. Robust grounded language understanding without dramatic failures and without corner cases is critical to building safe and fair robots; we demonstrate the significant role that compositionality can play in achieving that goal.

Yen-Ling Kuo, Boris Katz, Andrei Barbu

We demonstrate a reinforcement learning agent which uses a compositional recurrent neural network that takes as input an LTL formula and determines satisfying actions. The input LTL formulas have never been seen before, yet the network performs zero-shot generalization to satisfy them. This is a novel form of multi-task learning for RL agents where agents learn from one diverse set of tasks and generalize to a new set of diverse tasks. The formulation of the network enables this capacity to generalize. We demonstrate this ability in two domains. In a symbolic domain, the agent finds a sequence of letters that is accepted. In a Minecraft-like environment, the agent finds a sequence of actions that conform to the formula. While prior work could learn to execute one formula reliably given examples of that formula, we demonstrate how to encode all formulas reliably. This could form the basis of new multitask agents that discover sub-tasks and execute them without any additional training, as well as the agents which follow more complex linguistic commands. The structures required for this generalization are specific to LTL formulas, which opens up an interesting theoretical question: what structures are required in neural networks for zero-shot generalization to different logics?

Yu-Siang Wang, Yen-Ling Kuo, Boris Katz

We demonstrate how we can practically incorporate multi-step future information into a decoder of maximum likelihood sequence models. We propose a "k-step look-ahead" module to consider the likelihood information of a rollout up to k steps. Unlike other approaches that need to train another value network to evaluate the rollouts, we can directly apply this look-ahead module to improve the decoding of any sequence model trained in a maximum likelihood framework. We evaluate our look-ahead module on three datasets of varying difficulties: IM2LATEX-100k OCR image to LaTeX, WMT16 multimodal machine translation, and WMT14 machine translation. Our look-ahead module improves the performance of the simpler datasets such as IM2LATEX-100k and WMT16 multimodal machine translation. However, the improvement of the more difficult dataset (e.g., containing longer sequences), WMT14 machine translation, becomes marginal. Our further investigation using the k-step look-ahead suggests that the more difficult tasks suffer from the overestimated EOS (end-of-sentence) probability. We argue that the overestimated EOS probability also causes the decreased performance of beam search when increasing its beam width. We tackle the EOS problem by integrating an auxiliary EOS loss into the training to estimate if the model should emit EOS or other words. Our experiments show that improving EOS estimation not only increases the performance of our proposed look-ahead module but also the robustness of the beam search.

Yen-Ling Kuo, Boris Katz, Andrei Barbu

We demonstrate how a sampling-based robotic planner can be augmented to learn to understand a sequence of natural language commands in a continuous configuration space to move and manipulate objects. Our approach combines a deep network structured according to the parse of a complex command that includes objects, verbs, spatial relations, and attributes, with a sampling-based planner, RRT. A recurrent hierarchical deep network controls how the planner explores the environment, determines when a planned path is likely to achieve a goal, and estimates the confidence of each move to trade off exploitation and exploration between the network and the planner. Planners are designed to have near-optimal behavior when information about the task is missing, while networks learn to exploit observations which are available from the environment, making the two naturally complementary. Combining the two enables generalization to new maps, new kinds of obstacles, and more complex sentences that do not occur in the training set. Little data is required to train the model despite it jointly acquiring a CNN that extracts features from the environment as it learns the meanings of words. The model provides a level of interpretability through the use of attention maps allowing users to see its reasoning steps despite being an end-to-end model. This end-to-end model allows robots to learn to follow natural language commands in challenging continuous environments.

Yen-Ling Kuo, Andrei Barbu, Boris Katz

We demonstrate how a sequence model and a sampling-based planner can influence each other to produce efficient plans and how such a model can automatically learn to take advantage of observations of the environment. Sampling-based planners such as RRT generally know nothing of their environments even if they have traversed similar spaces many times. A sequence model, such as an HMM or LSTM, guides the search for good paths. The resulting model, called DeRRT*, observes the state of the planner and the local environment to bias the next move and next planner state. The neural-network-based models avoid manual feature engineering by co-training a convolutional network which processes map features and observations from sensors. We incorporate this sequence model in a manner that combines its likelihood with the existing bias for searching large unexplored Voronoi regions. This leads to more efficient trajectories with fewer rejected samples even in difficult domains such as when escaping bug traps. This model can also be used for dimensionality reduction in multi-agent environments with dynamic obstacles. Instead of planning in a high-dimensional space that includes the configurations of the other agents, we plan in a low-dimensional subspace relying on the sequence model to bias samples using the observed behavior of the other agents. The techniques presented here are general, include both graphical models and deep learning approaches, and can be adapted to a range of planners.