Jiayuan Mao

Po-Chen Ko, Jiayuan Mao, Yilun Du et al. · mit

In this work, we present an approach to construct a video-based robot policy capable of reliably executing diverse tasks across different robots and environments from few video demonstrations without using any action annotations. Our method leverages images as a task-agnostic representation, encoding both the state and action information, and text as a general representation for specifying robot goals. By synthesizing videos that ``hallucinate'' robot executing actions and in combination with dense correspondences between frames, our approach can infer the closed-formed action to execute to an environment without the need of any explicit action labels. This unique capability allows us to train the policy solely based on RGB videos and deploy learned policies to various robotic tasks. We demonstrate the efficacy of our approach in learning policies on table-top manipulation and navigation tasks. Additionally, we contribute an open-source framework for efficient video modeling, enabling the training of high-fidelity policy models with four GPUs within a single day.

Joy Hsu, Jiayuan Mao, Jiajun Wu · stanford

Grounding object properties and relations in 3D scenes is a prerequisite for a wide range of artificial intelligence tasks, such as visually grounded dialogues and embodied manipulation. However, the variability of the 3D domain induces two fundamental challenges: 1) the expense of labeling and 2) the complexity of 3D grounded language. Hence, essential desiderata for models are to be data-efficient, generalize to different data distributions and tasks with unseen semantic forms, as well as ground complex language semantics (e.g., view-point anchoring and multi-object reference). To address these challenges, we propose NS3D, a neuro-symbolic framework for 3D grounding. NS3D translates language into programs with hierarchical structures by leveraging large language-to-code models. Different functional modules in the programs are implemented as neural networks. Notably, NS3D extends prior neuro-symbolic visual reasoning methods by introducing functional modules that effectively reason about high-arity relations (i.e., relations among more than two objects), key in disambiguating objects in complex 3D scenes. Modular and compositional architecture enables NS3D to achieve state-of-the-art results on the ReferIt3D view-dependence task, a 3D referring expression comprehension benchmark. Importantly, NS3D shows significantly improved performance on settings of data-efficiency and generalization, and demonstrate zero-shot transfer to an unseen 3D question-answering task.

Joy Hsu, Jiayuan Mao, Joshua B. Tenenbaum et al. · stanford

Recent works such as VisProg and ViperGPT have smartly composed foundation models for visual reasoning-using large language models (LLMs) to produce programs that can be executed by pre-trained vision-language models. However, they operate in limited domains, such as 2D images, not fully exploiting the generalization of language: abstract concepts like "left" can also be grounded in 3D, temporal, and action data, as in moving to your left. This limited generalization stems from these inference-only methods' inability to learn or adapt pre-trained models to a new domain. We propose the Logic-Enhanced Foundation Model (LEFT), a unified framework that learns to ground and reason with concepts across domains with a differentiable, domain-independent, first-order logic-based program executor. LEFT has an LLM interpreter that outputs a program represented in a general, logic-based reasoning language, which is shared across all domains and tasks. LEFT's executor then executes the program with trainable domain-specific grounding modules. We show that LEFT flexibly learns concepts in four domains: 2D images, 3D scenes, human motions, and robotic manipulation. It exhibits strong reasoning ability in a wide variety of tasks, including those that are complex and not seen during training, and can be easily applied to new domains.

Renhao Wang, Jiayuan Mao, Joy Hsu et al. · stanford

Robots operating in the real world require both rich manipulation skills as well as the ability to semantically reason about when to apply those skills. Towards this goal, recent works have integrated semantic representations from large-scale pretrained vision-language (VL) models into manipulation models, imparting them with more general reasoning capabilities. However, we show that the conventional pretraining-finetuning pipeline for integrating such representations entangles the learning of domain-specific action information and domain-general visual information, leading to less data-efficient training and poor generalization to unseen objects and tasks. To this end, we propose ProgramPort, a modular approach to better leverage pretrained VL models by exploiting the syntactic and semantic structures of language instructions. Our framework uses a semantic parser to recover an executable program, composed of functional modules grounded on vision and action across different modalities. Each functional module is realized as a combination of deterministic computation and learnable neural networks. Program execution produces parameters to general manipulation primitives for a robotic end-effector. The entire modular network can be trained with end-to-end imitation learning objectives. Experiments show that our model successfully disentangles action and perception, translating to improved zero-shot and compositional generalization in a variety of manipulation behaviors. Project webpage at: \url{https://progport.github.io}.

Jiayuan Mao, Xuelin Yang, Xikun Zhang et al. · stanford

Building machines that can reason about physical events and their causal relationships is crucial for flexible interaction with the physical world. However, most existing physical and causal reasoning benchmarks are exclusively based on synthetically generated events and synthetic natural language descriptions of causal relationships. This design brings up two issues. First, there is a lack of diversity in both event types and natural language descriptions; second, causal relationships based on manually-defined heuristics are different from human judgments. To address both shortcomings, we present the CLEVRER-Humans benchmark, a video reasoning dataset for causal judgment of physical events with human labels. We employ two techniques to improve data collection efficiency: first, a novel iterative event cloze task to elicit a new representation of events in videos, which we term Causal Event Graphs (CEGs); second, a data augmentation technique based on neural language generative models. We convert the collected CEGs into questions and answers to be consistent with prior work. Finally, we study a collection of baseline approaches for CLEVRER-Humans question-answering, highlighting the great challenges set forth by our benchmark.

Ruocheng Wang, Yunzhi Zhang, Jiayuan Mao et al. · stanford

We study the problem of translating an image-based, step-by-step assembly manual created by human designers into machine-interpretable instructions. We formulate this problem as a sequential prediction task: at each step, our model reads the manual, locates the components to be added to the current shape, and infers their 3D poses. This task poses the challenge of establishing a 2D-3D correspondence between the manual image and the real 3D object, and 3D pose estimation for unseen 3D objects, since a new component to be added in a step can be an object built from previous steps. To address these two challenges, we present a novel learning-based framework, the Manual-to-Executable-Plan Network (MEPNet), which reconstructs the assembly steps from a sequence of manual images. The key idea is to integrate neural 2D keypoint detection modules and 2D-3D projection algorithms for high-precision prediction and strong generalization to unseen components. The MEPNet outperforms existing methods on three newly collected LEGO manual datasets and a Minecraft house dataset.

Jiayuan Mao, Tomás Lozano-Pérez, Joshua B. Tenenbaum et al. · mit

This paper studies a model learning and online planning approach towards building flexible and general robots. Specifically, we investigate how to exploit the locality and sparsity structures in the underlying environmental transition model to improve model generalization, data-efficiency, and runtime-efficiency. We present a new domain definition language, named PDSketch. It allows users to flexibly define high-level structures in the transition models, such as object and feature dependencies, in a way similar to how programmers use TensorFlow or PyTorch to specify kernel sizes and hidden dimensions of a convolutional neural network. The details of the transition model will be filled in by trainable neural networks. Based on the defined structures and learned parameters, PDSketch automatically generates domain-independent planning heuristics without additional training. The derived heuristics accelerate the performance-time planning for novel goals.

Zhezheng Luo, Jiayuan Mao, Jiajun Wu et al. · mit, stanford

We present a framework for learning useful subgoals that support efficient long-term planning to achieve novel goals. At the core of our framework is a collection of rational subgoals (RSGs), which are essentially binary classifiers over the environmental states. RSGs can be learned from weakly-annotated data, in the form of unsegmented demonstration trajectories, paired with abstract task descriptions, which are composed of terms initially unknown to the agent (e.g., collect-wood then craft-boat then go-across-river). Our framework also discovers dependencies between RSGs, e.g., the task collect-wood is a helpful subgoal for the task craft-boat. Given a goal description, the learned subgoals and the derived dependencies facilitate off-the-shelf planning algorithms, such as A* and RRT, by setting helpful subgoals as waypoints to the planner, which significantly improves performance-time efficiency.

Sumith Kulal, Jiayuan Mao, Alex Aiken et al. · stanford

We introduce Programmatic Motion Concepts, a hierarchical motion representation for human actions that captures both low-level motion and high-level description as motion concepts. This representation enables human motion description, interactive editing, and controlled synthesis of novel video sequences within a single framework. We present an architecture that learns this concept representation from paired video and action sequences in a semi-supervised manner. The compactness of our representation also allows us to present a low-resource training recipe for data-efficient learning. By outperforming established baselines, especially in the small data regime, we demonstrate the efficiency and effectiveness of our framework for multiple applications.

Guangxuan Xiao, Leslie Pack Kaelbling, Jiajun Wu et al. · stanford

Reasoning about the relationships between entities from input facts (e.g., whether Ari is a grandparent of Charlie) generally requires explicit consideration of other entities that are not mentioned in the query (e.g., the parents of Charlie). In this paper, we present an approach for learning to solve problems of this kind in large, real-world domains, using sparse and local hypergraph neural networks (SpaLoc). SpaLoc is motivated by two observations from traditional logic-based reasoning: relational inferences usually apply locally (i.e., involve only a small number of individuals), and relations are usually sparse (i.e., only hold for a small percentage of tuples in a domain). We exploit these properties to make learning and inference efficient in very large domains by (1) using a sparse tensor representation for hypergraph neural networks, (2) applying a sparsification loss during training to encourage sparse representations, and (3) subsampling based on a novel information sufficiency-based sampling process during training. SpaLoc achieves state-of-the-art performance on several real-world, large-scale knowledge graph reasoning benchmarks, and is the first framework for applying hypergraph neural networks on real-world knowledge graphs with more than 10k nodes.

Yanming Wan, Jiayuan Mao, Joshua B. Tenenbaum

We introduce HandMeThat, a benchmark for a holistic evaluation of instruction understanding and following in physical and social environments. While previous datasets primarily focused on language grounding and planning, HandMeThat considers the resolution of human instructions with ambiguities based on the physical (object states and relations) and social (human actions and goals) information. HandMeThat contains 10,000 episodes of human-robot interactions. In each episode, the robot first observes a trajectory of human actions towards her internal goal. Next, the robot receives a human instruction and should take actions to accomplish the subgoal set through the instruction. In this paper, we present a textual interface for our benchmark, where the robot interacts with a virtual environment through textual commands. We evaluate several baseline models on HandMeThat, and show that both offline and online reinforcement learning algorithms perform poorly on HandMeThat, suggesting significant room for future work on physical and social human-robot communications and interactions.

Ruocheng Wang, Yunzhi Zhang, Jiayuan Mao et al. · stanford

Human-designed visual manuals are crucial components in shape assembly activities. They provide step-by-step guidance on how we should move and connect different parts in a convenient and physically-realizable way. While there has been an ongoing effort in building agents that perform assembly tasks, the information in human-design manuals has been largely overlooked. We identify that this is due to 1) a lack of realistic 3D assembly objects that have paired manuals and 2) the difficulty of extracting structured information from purely image-based manuals. Motivated by this observation, we present IKEA-Manual, a dataset consisting of 102 IKEA objects paired with assembly manuals. We provide fine-grained annotations on the IKEA objects and assembly manuals, including decomposed assembly parts, assembly plans, manual segmentation, and 2D-3D correspondence between 3D parts and visual manuals. We illustrate the broad application of our dataset on four tasks related to shape assembly: assembly plan generation, part segmentation, pose estimation, and 3D part assembly.

Jiayuan Mao, Joshua B. Tenenbaum, Tomás Lozano-Pérez et al.

Humans demonstrate an impressive ability to acquire and generalize manipulation "tricks." Even from a single demonstration, such as using soup ladles to reach for distant objects, we can apply this skill to new scenarios involving different object positions, sizes, and categories (e.g., forks and hammers). Additionally, we can flexibly combine various skills to devise long-term plans. In this paper, we present a framework that enables machines to acquire such manipulation skills, referred to as "mechanisms," through a single demonstration and self-play. Our key insight lies in interpreting each demonstration as a sequence of changes in robot-object and object-object contact modes, which provides a scaffold for learning detailed samplers for continuous parameters. These learned mechanisms and samplers can be seamlessly integrated into standard task and motion planners, enabling their compositional use.

Zhezheng Luo, Jiayuan Mao, Joshua B. Tenenbaum et al.

This extended abstract describes a framework for analyzing the expressiveness, learning, and (structural) generalization of hypergraph neural networks (HyperGNNs). Specifically, we focus on how HyperGNNs can learn from finite datasets and generalize structurally to graph reasoning problems of arbitrary input sizes. Our first contribution is a fine-grained analysis of the expressiveness of HyperGNNs, that is, the set of functions that they can realize. Our result is a hierarchy of problems they can solve, defined in terms of various hyperparameters such as depths and edge arities. Next, we analyze the learning properties of these neural networks, especially focusing on how they can be trained on a finite set of small graphs and generalize to larger graphs, which we term structural generalization. Our theoretical results are further supported by the empirical results.

Zora Zhiruo Wang, Jiayuan Mao, Daniel Fried et al.

Despite the potential of language model-based agents to solve real-world tasks such as web navigation, current methods still struggle with long-horizon tasks with complex action trajectories. In contrast, humans can flexibly solve complex tasks by learning reusable task workflows from past experiences and using them to guide future actions. To build agents that can similarly benefit from this process, we introduce Agent Workflow Memory (AWM), a method for inducing commonly reused routines, i.e., workflows, and selectively providing workflows to the agent to guide subsequent generations. AWM flexibly applies to both offline and online scenarios, where agents induce workflows from training examples beforehand or from test queries on the fly. We experiment on two major web navigation benchmarks -- Mind2Web and WebArena -- that collectively cover 1000+ tasks from 200+ domains across travel, shopping, and social media, among others. AWM substantially improves the baseline results by 24.6% and 51.1% relative success rate on Mind2Web and WebArena while reducing the number of steps taken to solve WebArena tasks successfully. Furthermore, online AWM robustly generalizes in cross-task, website, and domain evaluations, surpassing baselines from 8.9 to 14.0 absolute points as train-test task distribution gaps widen.

Joy Hsu, Jiayuan Mao, Joshua B. Tenenbaum et al.

A unique aspect of human visual understanding is the ability to flexibly interpret abstract concepts: acquiring lifted rules explaining what they symbolize, grounding them across familiar and unfamiliar contexts, and making predictions or reasoning about them. While off-the-shelf vision-language models excel at making literal interpretations of images (e.g., recognizing object categories such as tree branches), they still struggle to make sense of such visual abstractions (e.g., how an arrangement of tree branches may form the walls of a maze). To address this challenge, we introduce Deep Schema Grounding (DSG), a framework that leverages explicit structured representations of visual abstractions for grounding and reasoning. At the core of DSG are schemas--dependency graph descriptions of abstract concepts that decompose them into more primitive-level symbols. DSG uses large language models to extract schemas, then hierarchically grounds concrete to abstract components of the schema onto images with vision-language models. The grounded schema is used to augment visual abstraction understanding. We systematically evaluate DSG and different methods in reasoning on our new Visual Abstractions Dataset, which consists of diverse, real-world images of abstract concepts and corresponding question-answer pairs labeled by humans. We show that DSG significantly improves the abstract visual reasoning performance of vision-language models, and is a step toward human-aligned understanding of visual abstractions.

Yanming Wan, Yue Wu, Yiping Wang et al.

For AI agents to be helpful to humans, they should be able to follow natural language instructions to complete everyday cooperative tasks in human environments. However, real human instructions inherently possess ambiguity, because the human speakers assume sufficient prior knowledge about their hidden goals and intentions. Standard language grounding and planning methods fail to address such ambiguities because they do not model human internal goals as additional partially observable factors in the environment. We propose a new framework, Follow Instructions with Social and Embodied Reasoning (FISER), aiming for better natural language instruction following in collaborative embodied tasks. Our framework makes explicit inferences about human goals and intentions as intermediate reasoning steps. We implement a set of Transformer-based models and evaluate them over a challenging benchmark, HandMeThat. We empirically demonstrate that using social reasoning to explicitly infer human intentions before making action plans surpasses purely end-to-end approaches. We also compare our implementation with strong baselines, including Chain of Thought prompting on the largest available pre-trained language models, and find that FISER provides better performance on the embodied social reasoning tasks under investigation, reaching the state-of-the-art on HandMeThat.

Haoyue Shi, Jiayuan Mao, Tete Xiao et al.

We study the problem of grounding distributional representations of texts on the visual domain, namely visual-semantic embeddings (VSE for short). Begin with an insightful adversarial attack on VSE embeddings, we show the limitation of current frameworks and image-text datasets (e.g., MS-COCO) both quantitatively and qualitatively. The large gap between the number of possible constitutions of real-world semantics and the size of parallel data, to a large extent, restricts the model to establish the link between textual semantics and visual concepts. We alleviate this problem by augmenting the MS-COCO image captioning datasets with textual contrastive adversarial samples. These samples are synthesized using linguistic rules and the WordNet knowledge base. The construction procedure is both syntax- and semantics-aware. The samples enforce the model to ground learned embeddings to concrete concepts within the image. This simple but powerful technique brings a noticeable improvement over the baselines on a diverse set of downstream tasks, in addition to defending known-type adversarial attacks. We release the codes at https://github.com/ExplorerFreda/VSE-C.

Lionel Wong, Jiayuan Mao, Pratyusha Sharma et al. · mit

Effective planning in the real world requires not only world knowledge, but the ability to leverage that knowledge to build the right representation of the task at hand. Decades of hierarchical planning techniques have used domain-specific temporal action abstractions to support efficient and accurate planning, almost always relying on human priors and domain knowledge to decompose hard tasks into smaller subproblems appropriate for a goal or set of goals. This paper describes Ada (Action Domain Acquisition), a framework for automatically constructing task-specific planning representations using task-general background knowledge from language models (LMs). Starting with a general-purpose hierarchical planner and a low-level goal-conditioned policy, Ada interactively learns a library of planner-compatible high-level action abstractions and low-level controllers adapted to a particular domain of planning tasks. On two language-guided interactive planning benchmarks (Mini Minecraft and ALFRED Household Tasks), Ada strongly outperforms other approaches that use LMs for sequential decision-making, offering more accurate plans and better generalization to complex tasks.

Xiaolin Fang, Bo-Ruei Huang, Jiayuan Mao et al.

Generalization to novel object configurations and instances across diverse tasks and environments is a critical challenge in robotics. Keypoint-based representations have been proven effective as a succinct representation for capturing essential object features, and for establishing a reference frame in action prediction, enabling data-efficient learning of robot skills. However, their manual design nature and reliance on additional human labels limit their scalability. In this paper, we propose KALM, a framework that leverages large pre-trained vision-language models (LMs) to automatically generate task-relevant and cross-instance consistent keypoints. KALM distills robust and consistent keypoints across views and objects by generating proposals using LMs and verifies them against a small set of robot demonstration data. Based on the generated keypoints, we can train keypoint-conditioned policy models that predict actions in keypoint-centric frames, enabling robots to generalize effectively across varying object poses, camera views, and object instances with similar functional shapes. Our method demonstrates strong performance in the real world, adapting to different tasks and environments from only a handful of demonstrations while requiring no additional labels. Website: https://kalm-il.github.io/

Weiyu Liu, Neil Nie, Ruohan Zhang et al.

We introduce Behavior from Language and Demonstration (BLADE), a framework for long-horizon robotic manipulation by integrating imitation learning and model-based planning. BLADE leverages language-annotated demonstrations, extracts abstract action knowledge from large language models (LLMs), and constructs a library of structured, high-level action representations. These representations include preconditions and effects grounded in visual perception for each high-level action, along with corresponding controllers implemented as neural network-based policies. BLADE can recover such structured representations automatically, without manually labeled states or symbolic definitions. BLADE shows significant capabilities in generalizing to novel situations, including novel initial states, external state perturbations, and novel goals. We validate the effectiveness of our approach both in simulation and on real robots with a diverse set of objects with articulated parts, partial observability, and geometric constraints.

Bingda Tang, Yuhui Zhang, Xiaohan Wang et al.

Aligning denoising generative models with human preferences or verifiable rewards remains a key challenge. While policy-gradient online reinforcement learning (RL) offers a principled post-training framework, its direct application is hindered by the intractable likelihoods of these models. Prior work therefore either optimizes an induced Markov decision process (MDP) over sampling trajectories, which is stable but inefficient, or uses likelihood surrogates based on the diffusion evidence lower bound (ELBO), which have so far underperformed on visual generation. Our key insight is that the ELBO-based approach can, in fact, be made both stable and efficient. By reducing surrogate variance and controlling gradient steps, we show that this approach can beat MDP-based methods. To this end, we introduce Variational GRPO (V-GRPO), a method that integrates ELBO-based surrogates with the Group Relative Policy Optimization (GRPO) algorithm, alongside a set of simple yet essential techniques. Our method is easy to implement, aligns with pretraining objectives, and avoids the limitations of MDP-based methods. V-GRPO achieves state-of-the-art performance in text-to-image synthesis, while delivering a $2\times$ speedup over MixGRPO and a $3\times$ speedup over DiffusionNFT.

Yanwei Wang, Tsun-Hsuan Wang, Jiayuan Mao et al.

Grounding the common-sense reasoning of Large Language Models (LLMs) in physical domains remains a pivotal yet unsolved problem for embodied AI. Whereas prior works have focused on leveraging LLMs directly for planning in symbolic spaces, this work uses LLMs to guide the search of task structures and constraints implicit in multi-step demonstrations. Specifically, we borrow from manipulation planning literature the concept of mode families, which group robot configurations by specific motion constraints, to serve as an abstraction layer between the high-level language representations of an LLM and the low-level physical trajectories of a robot. By replaying a few human demonstrations with synthetic perturbations, we generate coverage over the demonstrations' state space with additional successful executions as well as counterfactuals that fail the task. Our explanation-based learning framework trains an end-to-end differentiable neural network to predict successful trajectories from failures and as a by-product learns classifiers that ground low-level states and images in mode families without dense labeling. The learned grounding classifiers can further be used to translate language plans into reactive policies in the physical domain in an interpretable manner. We show our approach improves the interpretability and reactivity of imitation learning through 2D navigation and simulated and real robot manipulation tasks. Website: https://yanweiw.github.io/glide

Jiayuan Mao, Tomás Lozano-Pérez, Joshua B. Tenenbaum et al.

Goal-conditioned policies are generally understood to be "feed-forward" circuits, in the form of neural networks that map from the current state and the goal specification to the next action to take. However, under what circumstances such a policy can be learned and how efficient the policy will be are not well understood. In this paper, we present a circuit complexity analysis for relational neural networks (such as graph neural networks and transformers) representing policies for planning problems, by drawing connections with serialized goal regression search (S-GRS). We show that there are three general classes of planning problems, in terms of the growth of circuit width and depth as a function of the number of objects and planning horizon, providing constructive proofs. We also illustrate the utility of this analysis for designing neural networks for policy learning.

Yuyao Liu, Jiayuan Mao, Joshua Tenenbaum et al.

We present a novel approach, MAGIC (manipulation analogies for generalizable intelligent contacts), for one-shot learning of manipulation strategies with fast and extensive generalization to novel objects. By leveraging a reference action trajectory, MAGIC effectively identifies similar contact points and sequences of actions on novel objects to replicate a demonstrated strategy, such as using different hooks to retrieve distant objects of different shapes and sizes. Our method is based on a two-stage contact-point matching process that combines global shape matching using pretrained neural features with local curvature analysis to ensure precise and physically plausible contact points. We experiment with three tasks including scooping, hanging, and hooking objects. MAGIC demonstrates superior performance over existing methods, achieving significant improvements in runtime speed and generalization to different object categories. Website: https://magic-2024.github.io/ .

Yiqing Xu, Jiayuan Mao, Yilun Du et al.

This paper studies the challenge of developing robots capable of understanding under-specified instructions for creating functional object arrangements, such as "set up a dining table for two"; previous arrangement approaches have focused on much more explicit instructions, such as "put object A on the table." We introduce a framework, SetItUp, for learning to interpret under-specified instructions. SetItUp takes a small number of training examples and a human-crafted program sketch to uncover arrangement rules for specific scene types. By leveraging an intermediate graph-like representation of abstract spatial relationships among objects, SetItUp decomposes the arrangement problem into two subproblems: i) learning the arrangement patterns from limited data and ii) grounding these abstract relationships into object poses. SetItUp leverages large language models (LLMs) to propose the abstract spatial relationships among objects in novel scenes as the constraints to be satisfied; then, it composes a library of diffusion models associated with these abstract relationships to find object poses that satisfy the constraints. We validate our framework on a dataset comprising study desks, dining tables, and coffee tables, with the results showing superior performance in generating physically plausible, functional, and aesthetically pleasing object arrangements compared to existing models.

Guangyuan Jiang, Matthias Hofer, Jiayuan Mao et al.

One hallmark of human language is its combinatoriality -- reusing a relatively small inventory of building blocks to create a far larger inventory of increasingly complex structures. In this paper, we explore the idea that combinatoriality in language reflects a human inductive bias toward representational efficiency in symbol systems. We develop a computational framework for discovering structure in a writing system. Built on top of state-of-the-art library learning and program synthesis techniques, our computational framework discovers known linguistic structures in the Chinese writing system and reveals how the system evolves towards simplification under pressures for representational efficiency. We demonstrate how a library learning approach, utilizing learned abstractions and compression, may help reveal the fundamental computational principles that underlie the creation of combinatorial structures in human cognition, and offer broader insights into the evolution of efficient communication systems.

Morris Yau, Ekin Akyürek, Jiayuan Mao et al. · mit

Previous research has explored the computational expressivity of Transformer models in simulating Boolean circuits or Turing machines. However, the learnability of these simulators from observational data has remained an open question. Our study addresses this gap by providing the first polynomial-time learnability results (specifically strong, agnostic PAC learning) for single-layer Transformers with linear attention. We show that linear attention may be viewed as a linear predictor in a suitably defined RKHS. As a consequence, the problem of learning any linear transformer may be converted into the problem of learning an ordinary linear predictor in an expanded feature space, and any such predictor may be converted back into a multiheaded linear transformer. Moving to generalization, we show how to efficiently identify training datasets for which every empirical risk minimizer is equivalent (up to trivial symmetries) to the linear Transformer that generated the data, thereby guaranteeing the learned model will correctly generalize across all inputs. Finally, we provide examples of computations expressible via linear attention and therefore polynomial-time learnable, including associative memories, finite automata, and a class of Universal Turing Machine (UTMs) with polynomially bounded computation histories. We empirically validate our theoretical findings on three tasks: learning random linear attention networks, key--value associations, and learning to execute finite automata. Our findings bridge a critical gap between theoretical expressivity and learnability of Transformers, and show that flexible and general models of computation are efficiently learnable.

Jiayuan Mao, Joshua B. Tenenbaum, Jiajun Wu

This article presents a concept-centric paradigm for building agents that can learn continually and reason flexibly. The concept-centric agent utilizes a vocabulary of neuro-symbolic concepts. These concepts, such as object, relation, and action concepts, are grounded on sensory inputs and actuation outputs. They are also compositional, allowing for the creation of novel concepts through their structural combination. To facilitate learning and reasoning, the concepts are typed and represented using a combination of symbolic programs and neural network representations. Leveraging such neuro-symbolic concepts, the agent can efficiently learn and recombine them to solve various tasks across different domains, ranging from 2D images, videos, 3D scenes, and robotic manipulation tasks. This concept-centric framework offers several advantages, including data efficiency, compositional generalization, continual learning, and zero-shot transfer.

Haowen Liu, Shaoxiong Yao, Haonan Chen et al.

Vision-Language Models (VLMs) exhibit remarkable common-sense and semantic reasoning capabilities. However, they lack a grounded understanding of physical dynamics. This limitation arises from training VLMs on static internet-scale visual-language data that contain no causal interactions or action-conditioned changes. Consequently, it remains challenging to leverage VLMs for fine-grained robotic manipulation tasks that require physical understanding, reasoning, and corresponding action planning. To overcome this, we present SIMPACT, a test-time, SIMulation-enabled ACTion Planning framework that equips VLMs with physical reasoning through simulation-in-the-loop world modeling, without requiring any additional training. From a single RGB-D observation, SIMPACT efficiently constructs physics simulations, enabling the VLM to propose informed actions, observe simulated rollouts, and iteratively refine its reasoning. By integrating language reasoning with physics prediction, our simulation-enabled VLM can understand contact dynamics and action outcomes in a physically grounded way. Our method demonstrates state-of-the-art performance on five challenging, real-world rigid-body and deformable manipulation tasks that require fine-grained physical reasoning, outperforming existing general-purpose robotic manipulation models. Our results demonstrate that embedding physics understanding via efficient simulation into VLM reasoning at test time offers a promising path towards generalizable embodied intelligence. Project webpage can be found at https://simpact-bot.github.io

Xinyi Gu, Jiayuan Mao, Zhang-Wei Hong et al.

Pretrained multi-modal large language models (MLLMs) demonstrate strong performance on diverse multimodal tasks, but remain limited in reasoning capabilities for domains where annotations are difficult to collect. In this work, we focus on artificial image domains such as charts, rendered documents, and webpages, which are abundant in practice yet lack large-scale human annotated reasoning datasets. We introduce COGS (COmposition-Grounded instruction Synthesis), a data-efficient framework for equipping MLLMs with advanced reasoning abilities from a small set of seed questions. The key idea is to decompose each seed question into primitive perception and reasoning factors, which can then be systematically recomposed with new images to generate large collections of synthetic question-answer pairs. Each generated question is paired with subquestions and intermediate answers, enabling reinforcement learning with factor-level process rewards. Experiments on chart reasoning show that COGS substantially improves performance on unseen questions, with the largest gains on reasoning-heavy and compositional questions. Moreover, training with a factor-level mixture of different seed data yields better transfer across multiple datasets, suggesting that COGS induces generalizable capabilities rather than dataset-specific overfitting. We further demonstrate that the framework extends beyond charts to other domains such as webpages.

Haonan Chen, Jiaming Xu, Hongyu Chen et al.

Effectively integrating diverse sensory modalities is crucial for robotic manipulation. However, the typical approach of feature concatenation is often suboptimal: dominant modalities such as vision can overwhelm sparse but critical signals like touch in contact-rich tasks, and monolithic architectures cannot flexibly incorporate new or missing modalities without retraining. Our method factorizes the policy into a set of diffusion models, each specialized for a single representation (e.g., vision or touch), and employs a router network that learns consensus weights to adaptively combine their contributions, enabling incremental of new representations. We evaluate our approach on simulated manipulation tasks in {RLBench}, as well as real-world tasks such as occluded object picking, in-hand spoon reorientation, and puzzle insertion, where it significantly outperforms feature-concatenation baselines on scenarios requiring multimodal reasoning. Our policy further demonstrates robustness to physical perturbations and sensor corruption. We further conduct perturbation-based importance analysis, which reveals adaptive shifts between modalities.

Yaoyao Qian, Jindan Huang, Yuanli Wang et al.

Dialogue systems often fail when user utterances are semantically complete yet lack the clarity and completeness required for appropriate system action. This mismatch arises because users frequently do not fully understand their own needs, while systems require precise intent definitions. This highlights the critical Intent-Action Alignment Problem: determining when an expression is not just understood, but truly ready for a system to act upon. We present STORM, a framework modeling asymmetric information dynamics through conversations between UserLLM (full internal access) and AgentLLM (observable behavior only). STORM produces annotated corpora capturing trajectories of expression phrasing and latent cognitive transitions, enabling systematic analysis of how collaborative understanding develops. Our contributions include: (1) formalizing asymmetric information processing in dialogue systems; (2) modeling intent formation tracking collaborative understanding evolution; and (3) evaluation metrics measuring internal cognitive improvements alongside task performance. Experiments across four language models reveal that moderate uncertainty (40-60%) can outperform complete transparency in certain scenarios, with model-specific patterns suggesting reconsideration of optimal information completeness in human-AI collaboration. These findings contribute to understanding asymmetric reasoning dynamics and inform uncertainty-calibrated dialogue system design.

Xinyi Gu, Jiayuan Mao

Images not only depict objects but also encapsulate rich interactions between them. However, generating faithful and high-fidelity images involving multiple entities interacting with each other, is a long-standing challenge. While pre-trained text-to-image models are trained on large-scale datasets to follow diverse text instructions, they struggle to generate accurate interactions, likely due to the scarcity of training data for uncommon object interactions. This paper introduces InterActing, an interaction-focused dataset with 1000 fine-grained prompts covering three key scenarios: (1) functional and action-based interactions, (2) compositional spatial relationships, and (3) multi-subject interactions. To address interaction generation challenges, we propose a decomposition-augmented refinement procedure. Our approach, DetailScribe, built on Stable Diffusion 3.5, leverages LLMs to decompose interactions into finer-grained concepts, uses a VLM to critique generated images, and applies targeted interventions within the diffusion process in refinement. Automatic and human evaluations show significantly improved image quality, demonstrating the potential of enhanced inference strategies. Our dataset and code are available at https://concepts-ai.com/p/detailscribe/ to facilitate future exploration of interaction-rich image generation.

Yilun Du, Jiayuan Mao, Joshua B. Tenenbaum

We introduce iterative reasoning through energy diffusion (IRED), a novel framework for learning to reason for a variety of tasks by formulating reasoning and decision-making problems with energy-based optimization. IRED learns energy functions to represent the constraints between input conditions and desired outputs. After training, IRED adapts the number of optimization steps during inference based on problem difficulty, enabling it to solve problems outside its training distribution -- such as more complex Sudoku puzzles, matrix completion with large value magnitudes, and pathfinding in larger graphs. Key to our method's success is two novel techniques: learning a sequence of annealed energy landscapes for easier inference and a combination of score function and energy landscape supervision for faster and more stable training. Our experiments show that IRED outperforms existing methods in continuous-space reasoning, discrete-space reasoning, and planning tasks, particularly in more challenging scenarios. Code and visualizations at https://energy-based-model.github.io/ired/

Weiyu Liu, Jiayuan Mao, Joy Hsu et al.

In this paper, we propose composable part-based manipulation (CPM), a novel approach that leverages object-part decomposition and part-part correspondences to improve learning and generalization of robotic manipulation skills. By considering the functional correspondences between object parts, we conceptualize functional actions, such as pouring and constrained placing, as combinations of different correspondence constraints. CPM comprises a collection of composable diffusion models, where each model captures a different inter-object correspondence. These diffusion models can generate parameters for manipulation skills based on the specific object parts. Leveraging part-based correspondences coupled with the task decomposition into distinct constraints enables strong generalization to novel objects and object categories. We validate our approach in both simulated and real-world scenarios, demonstrating its effectiveness in achieving robust and generalized manipulation capabilities.

Weiyu Liu, Geng Chen, Joy Hsu et al.

This paper presents a framework for learning state and action abstractions in sequential decision-making domains. Our framework, planning abstraction from language (PARL), utilizes language-annotated demonstrations to automatically discover a symbolic and abstract action space and induce a latent state abstraction based on it. PARL consists of three stages: 1) recovering object-level and action concepts, 2) learning state abstractions, abstract action feasibility, and transition models, and 3) applying low-level policies for abstract actions. During inference, given the task description, PARL first makes abstract action plans using the latent transition and feasibility functions, then refines the high-level plan using low-level policies. PARL generalizes across scenarios involving novel object instances and environments, unseen concept compositions, and tasks that require longer planning horizons than settings it is trained on.

Zhutian Yang, Jiayuan Mao, Yilun Du et al.

This paper introduces an approach for learning to solve continuous constraint satisfaction problems (CCSP) in robotic reasoning and planning. Previous methods primarily rely on hand-engineering or learning generators for specific constraint types and then rejecting the value assignments when other constraints are violated. By contrast, our model, the compositional diffusion continuous constraint solver (Diffusion-CCSP) derives global solutions to CCSPs by representing them as factor graphs and combining the energies of diffusion models trained to sample for individual constraint types. Diffusion-CCSP exhibits strong generalization to novel combinations of known constraints, and it can be integrated into a task and motion planner to devise long-horizon plans that include actions with both discrete and continuous parameters. Project site: https://diffusion-ccsp.github.io/

Lingjie Mei, Jiayuan Mao, Ziqi Wang et al.

We present a meta-learning framework for learning new visual concepts quickly, from just one or a few examples, guided by multiple naturally occurring data streams: simultaneously looking at images, reading sentences that describe the objects in the scene, and interpreting supplemental sentences that relate the novel concept with other concepts. The learned concepts support downstream applications, such as answering questions by reasoning about unseen images. Our model, namely FALCON, represents individual visual concepts, such as colors and shapes, as axis-aligned boxes in a high-dimensional space (the "box embedding space"). Given an input image and its paired sentence, our model first resolves the referential expression in the sentence and associates the novel concept with particular objects in the scene. Next, our model interprets supplemental sentences to relate the novel concept with other known concepts, such as "X has property Y" or "X is a kind of Y". Finally, it infers an optimal box embedding for the novel concept that jointly 1) maximizes the likelihood of the observed instances in the image, and 2) satisfies the relationships between the novel concepts and the known ones. We demonstrate the effectiveness of our model on both synthetic and real-world datasets.

Jiayuan Mao, Haoyue Shi, Jiajun Wu et al.

We present Grammar-Based Grounded Lexicon Learning (G2L2), a lexicalist approach toward learning a compositional and grounded meaning representation of language from grounded data, such as paired images and texts. At the core of G2L2 is a collection of lexicon entries, which map each word to a tuple of a syntactic type and a neuro-symbolic semantic program. For example, the word shiny has a syntactic type of adjective; its neuro-symbolic semantic program has the symbolic form λx. filter(x, SHINY), where the concept SHINY is associated with a neural network embedding, which will be used to classify shiny objects. Given an input sentence, G2L2 first looks up the lexicon entries associated with each token. It then derives the meaning of the sentence as an executable neuro-symbolic program by composing lexical meanings based on syntax. The recovered meaning programs can be executed on grounded inputs. To facilitate learning in an exponentially-growing compositional space, we introduce a joint parsing and expected execution algorithm, which does local marginalization over derivations to reduce the training time. We evaluate G2L2 on two domains: visual reasoning and language-driven navigation. Results show that G2L2 can generalize from small amounts of data to novel compositions of words.

Jiayuan Mao, Zhezheng Luo, Chuang Gan et al.

We present Temporal and Object Quantification Networks (TOQ-Nets), a new class of neuro-symbolic networks with a structural bias that enables them to learn to recognize complex relational-temporal events. This is done by including reasoning layers that implement finite-domain quantification over objects and time. The structure allows them to generalize directly to input instances with varying numbers of objects in temporal sequences of varying lengths. We evaluate TOQ-Nets on input domains that require recognizing event-types in terms of complex temporal relational patterns. We demonstrate that TOQ-Nets can generalize from small amounts of data to scenarios containing more objects than were present during training and to temporal warpings of input sequences.

Sumith Kulal, Jiayuan Mao, Alex Aiken et al.

Current approaches to video analysis of human motion focus on raw pixels or keypoints as the basic units of reasoning. We posit that adding higher-level motion primitives, which can capture natural coarser units of motion such as backswing or follow-through, can be used to improve downstream analysis tasks. This higher level of abstraction can also capture key features, such as loops of repeated primitives, that are currently inaccessible at lower levels of representation. We therefore introduce Motion Programs, a neuro-symbolic, program-like representation that expresses motions as a composition of high-level primitives. We also present a system for automatically inducing motion programs from videos of human motion and for leveraging motion programs in video synthesis. Experiments show that motion programs can accurately describe a diverse set of human motions and the inferred programs contain semantically meaningful motion primitives, such as arm swings and jumping jacks. Our representation also benefits downstream tasks such as video interpolation and video prediction and outperforms off-the-shelf models. We further demonstrate how these programs can detect diverse kinds of repetitive motion and facilitate interactive video editing.

Zhenfang Chen, Jiayuan Mao, Jiajun Wu et al.

We study the problem of dynamic visual reasoning on raw videos. This is a challenging problem; currently, state-of-the-art models often require dense supervision on physical object properties and events from simulation, which are impractical to obtain in real life. In this paper, we present the Dynamic Concept Learner (DCL), a unified framework that grounds physical objects and events from video and language. DCL first adopts a trajectory extractor to track each object over time and to represent it as a latent, object-centric feature vector. Building upon this object-centric representation, DCL learns to approximate the dynamic interaction among objects using graph networks. DCL further incorporates a semantic parser to parse questions into semantic programs and, finally, a program executor to run the program to answer the question, levering the learned dynamics model. After training, DCL can detect and associate objects across the frames, ground visual properties, and physical events, understand the causal relationship between events, make future and counterfactual predictions, and leverage these extracted presentations for answering queries. DCL achieves state-of-the-art performance on CLEVRER, a challenging causal video reasoning dataset, even without using ground-truth attributes and collision labels from simulations for training. We further test DCL on a newly proposed video-retrieval and event localization dataset derived from CLEVRER, showing its strong generalization capacity.

Ruocheng Wang, Jiayuan Mao, Samuel J. Gershman et al.

We present Language-mediated, Object-centric Representation Learning (LORL), a paradigm for learning disentangled, object-centric scene representations from vision and language. LORL builds upon recent advances in unsupervised object discovery and segmentation, notably MONet and Slot Attention. While these algorithms learn an object-centric representation just by reconstructing the input image, LORL enables them to further learn to associate the learned representations to concepts, i.e., words for object categories, properties, and spatial relationships, from language input. These object-centric concepts derived from language facilitate the learning of object-centric representations. LORL can be integrated with various unsupervised object discovery algorithms that are language-agnostic. Experiments show that the integration of LORL consistently improves the performance of unsupervised object discovery methods on two datasets via the help of language. We also show that concepts learned by LORL, in conjunction with object discovery methods, aid downstream tasks such as referring expression comprehension.

Jianwei Yang, Jiayuan Mao, Jiajun Wu et al.

When answering questions about an image, it not only needs knowing what -- understanding the fine-grained contents (e.g., objects, relationships) in the image, but also telling why -- reasoning over grounding visual cues to derive the answer for a question. Over the last few years, we have seen significant progress on visual question answering. Though impressive as the accuracy grows, it still lags behind to get knowing whether these models are undertaking grounding visual reasoning or just leveraging spurious correlations in the training data. Recently, a number of works have attempted to answer this question from perspectives such as grounding and robustness. However, most of them are either focusing on the language side or coarsely studying the pixel-level attention maps. In this paper, by leveraging the step-wise object grounding annotations provided in the GQA dataset, we first present a systematical object-centric diagnosis of visual reasoning on grounding and robustness, particularly on the vision side. According to the extensive comparisons across different models, we find that even models with high accuracy are not good at grounding objects precisely, nor robust to visual content perturbations. In contrast, symbolic and modular models have a relatively better grounding and robustness, though at the cost of accuracy. To reconcile these different aspects, we further develop a diagnostic model, namely Graph Reasoning Machine. Our model replaces purely symbolic visual representation with probabilistic scene graph and then applies teacher-forcing training for the visual reasoning module. The designed model improves the performance on all three metrics over the vanilla neural-symbolic model while inheriting the transparency. Further ablation studies suggest that this improvement is mainly due to more accurate image understanding and proper intermediate reasoning supervisions.

Yikai Li, Jiayuan Mao, Xiuming Zhang et al.

We consider two important aspects in understanding and editing images: modeling regular, program-like texture or patterns in 2D planes, and 3D posing of these planes in the scene. Unlike prior work on image-based program synthesis, which assumes the image contains a single visible 2D plane, we present Box Program Induction (BPI), which infers a program-like scene representation that simultaneously models repeated structure on multiple 2D planes, the 3D position and orientation of the planes, and camera parameters, all from a single image. Our model assumes a box prior, i.e., that the image captures either an inner view or an outer view of a box in 3D. It uses neural networks to infer visual cues such as vanishing points, wireframe lines to guide a search-based algorithm to find the program that best explains the image. Such a holistic, structured scene representation enables 3D-aware interactive image editing operations such as inpainting missing pixels, changing camera parameters, and extrapolate the image contents.

Yikai Li, Jiayuan Mao, Xiuming Zhang et al.

We study the inverse graphics problem of inferring a holistic representation for natural images. Given an input image, our goal is to induce a neuro-symbolic, program-like representation that jointly models camera poses, object locations, and global scene structures. Such high-level, holistic scene representations further facilitate low-level image manipulation tasks such as inpainting. We formulate this problem as jointly finding the camera pose and scene structure that best describe the input image. The benefits of such joint inference are two-fold: scene regularity serves as a new cue for perspective correction, and in turn, correct perspective correction leads to a simplified scene structure, similar to how the correct shape leads to the most regular texture in shape from texture. Our proposed framework, Perspective Plane Program Induction (P3I), combines search-based and gradient-based algorithms to efficiently solve the problem. P3I outperforms a set of baselines on a collection of Internet images, across tasks including camera pose estimation, global structure inference, and down-stream image manipulation tasks.

Chi Han, Jiayuan Mao, Chuang Gan et al.

Humans reason with concepts and metaconcepts: we recognize red and green from visual input; we also understand that they describe the same property of objects (i.e., the color). In this paper, we propose the visual concept-metaconcept learner (VCML) for joint learning of concepts and metaconcepts from images and associated question-answer pairs. The key is to exploit the bidirectional connection between visual concepts and metaconcepts. Visual representations provide grounding cues for predicting relations between unseen pairs of concepts. Knowing that red and green describe the same property of objects, we generalize to the fact that cube and sphere also describe the same property of objects, since they both categorize the shape of objects. Meanwhile, knowledge about metaconcepts empowers visual concept learning from limited, noisy, and even biased data. From just a few examples of purple cubes we can understand a new color purple, which resembles the hue of the cubes instead of the shape of them. Evaluation on both synthetic and real-world datasets validates our claims.

Jiayuan Mao, Xiuming Zhang, Yikai Li et al.

Humans are capable of building holistic representations for images at various levels, from local objects, to pairwise relations, to global structures. The interpretation of structures involves reasoning over repetition and symmetry of the objects in the image. In this paper, we present the Program-Guided Image Manipulator (PG-IM), inducing neuro-symbolic program-like representations to represent and manipulate images. Given an image, PG-IM detects repeated patterns, induces symbolic programs, and manipulates the image using a neural network that is guided by the program. PG-IM learns from a single image, exploiting its internal statistics. Despite trained only on image inpainting, PG-IM is directly capable of extrapolation and regularity editing in a unified framework. Extensive experiments show that PG-IM achieves superior performance on all the tasks.

Sidi Lu, Jiayuan Mao, Joshua B. Tenenbaum et al.

Most structure inference methods either rely on exhaustive search or are purely data-driven. Exhaustive search robustly infers the structure of arbitrarily complex data, but it is slow. Data-driven methods allow efficient inference, but do not generalize when test data have more complex structures than training data. In this paper, we propose a hybrid inference algorithm, the Neurally-Guided Structure Inference (NG-SI), keeping the advantages of both search-based and data-driven methods. The key idea of NG-SI is to use a neural network to guide the hierarchical, layer-wise search over the compositional space of structures. We evaluate our algorithm on two representative structure inference tasks: probabilistic matrix decomposition and symbolic program parsing. It outperforms data-driven and search-based alternatives on both tasks.