A central challenge in building intelligent systems is enabling agents to jointly perceive complex inputs, form hypotheses about hidden patterns, and design informative experiments to test them. To study this problem, we propose ZendoWorld, a controlled interactive environment in which agents must infer a logical rule about visual game observations, acquire information by proposing new scenes, and refine their hypotheses based on feedback from the game environment. We evaluate several agents spanning pure VLM reasoning, Bayesian particle filtering, dynamic concept discovery, and neuro-symbolic methods. Our main findings are: (1) high accuracy in predicting labels for observed examples does not imply recovery of the underlying rule; (2) perception and induction are distinct bottlenecks for different agent classes; and (3) VLM-based agents propose near-uninformative experiments, failing to actively reduce hypothesis uncertainty. To compare these results, we collect human data on the task, which reveals a gap in inductive reasoning, particularly for more complex rules. Overall, ZENDOWORLD takes an important step toward evaluating intelligent agents and identifies concrete avenues for improvement, particularly in domains like scientific discovery.

Long‑horizon embodied planning is challenging because the world does not only change through an agent’s actions: exogenous processes (e.g., water heating, dominoes cascading) unfold concurrently with the agent's actions. We propose a framework for abstract world models that jointly learns (i) symbolic state representations and (ii) causal processes for both endogenous actions and exogenous mechanisms. Each causal process models the time course of a stochastic causal-effect relation. We learn these world models from limited data via variational Bayesian inference combined with LLM proposals. Across five simulated tabletop robotics environments, the learned models enable fast planning that generalizes to held‑out tasks with more objects and more complex goals, outperforming a range of baselines.

Vision-Language Models (VLMs) have made great strides in everyday visual tasks, such as captioning a natural image, or answering commonsense questions about such images. But humans possess the puzzling ability to deploy their visual reasoning abilities in radically new situations, a skill rigorously tested by the classic set of visual reasoning challenges known as the Bongard problems. We present a neurosymbolic approach to solving these problems: given a hypothesized solution rule for a Bongard problem, we leverage LLMs to generate parameterized programmatic representations for the rule and perform parameter fitting using Bayesian optimization. We evaluate our method on classifying Bongard problem images given the ground truth rule, as well as on solving the problems from scratch.

How humans grow and maintain abstract knowledge from the sparse, streaming noisy data of experience is a longstanding challenge in cognitive science. Any computational account must satisfy at least three desiderata: It must be (1) data-efficient and compute-efficient, (2) capture gradations of uncertainty to support intelligent inquiry and information gathering, and (3) be flexible enough to mentally represent the endless range of concepts people can learn and think about. Here we introduce a computational model that captures these three properties, by encoding symbolic knowledge as mental programs that combine natural language with source code, and sequentially inferring mental programs using LLM-guided Bayesian learning algorithms. Across a range of behavioral studies this model successfully reproduces quantitative signatures of human inductive learning and active inquiry, such as anchoring, garden-pathing, and other effects. In contrast, pure LLMs and classic Bayesian models either fail at the underlying task, or do not reproduce human behavior, or succeed only at exorbitant computational cost. These results suggest that one way humans continually grow their knowledge is by mentally representing many hypotheses spanning language-like and program-like representations, then revising those hypotheses to approximate Bayesian updates, while a bottom-up neural mechanism (an LLM) makes inference both tractable and learnable.

World models are central to building AI agents capable of flexible reasoning and planning. Yet current evaluations (i) test only properties measurable from observed interactions within an environment, such as next-frame prediction or task return, and (ii) do not test whether a learned model supports diverse queries about the same environment. In contrast, humans build \textit{general-purpose} models that can answer many different questions about an environment---including questions that require understanding global structure and counterfactual consequences. We propose : a protocol for evaluating agents' ability to learn general-purpose models. A benchmark pairs environments with multiple ---properties of the full environment---rather than objectives defined only on observed trajectories. Individually, these queries can target global and counterfactual properties (e.g., reachability or the effects of interventions) that are not determined by any single rollout distribution. Collectively, they assess model generality across query types. We instantiate WorldTest as , a minimal yet expressive benchmark of 43 interactive grid-world environments and 129 tasks across three query families for both humans and learning agents. AutumnBench supports diverse environments and evaluations, including queries to evaluate prediction, counterfactual reasoning, and long-horizon planning. Experiments with 517 human participants and five frontier models show that humans substantially outperform these models, a gap we attribute to differences in exploration and belief updating. and provide a rigorous framework for evaluating world-model learning and expose critical limitations in current approaches.

Real world decision-making requires constructing mental models under uncertainty over evidence, over the underlying causal rules, and over the state of the world itself. Which computational principles underpin human inference under such conditions, and do LLM-based agents exhibit similar behavior given matching constraints? We address these questions using an inductive inference Box Task in which participants, human children and LLM-based agents, infer a latent cause through sequential interaction with an uncertain environment. We formalize this task as program induction with Bayesian particle-based inference, admitting two complementary interpretations: (1) as a constraint satisfaction process over hypotheses, and (2) as a program synthesis problem in which hypotheses are executable programs evaluated against evidence. Using the constraint-based formulation, we show that children's behavior is best explained by a combination of subjective evidence reliability and online hypothesis generation, accounting for both their evidence-seeking patterns and their dissociation between task completion and rule generalization. Using the program synthesis formulation, we treat LLM-based agents as model organisms: controllable systems that allow systematic manipulation of task conditions. Across backends, LLM-based agents replicate children's responses to changes in evidence reliability and observability, including discounting unreliable evidence, seeking to resolve partial information, and dissociating between task completion and causal generalization. At the same time, LLM-based agents tend to over-observe and over-comply with instructions relative to children. These results suggest that while children and LLM-based agents adapt similarly to environmental structure, their information-seeking behavior exhibits distinct underlying costs and inductive biases.

Building agents that can perform new skills by composing existing skills is a long-standing goal of AI agent research. Towards this end, we investigate how to efficiently acquire a sequence of skills, formalized as hierarchical neural options. However, existing model-free hierarchical reinforcement algorithms need a lot of data. We propose a novel method, which we call AgentOWL (Option and World model Learning Agent), that jointly learns – in a sample efficient way – an abstract world model (abstracting across both states and time) and a set of hierarchical neural options. We show, on a subset of Object-Centric Atari games, that our method can learn more skills using much less data than baseline methods.

LLM prompting is widely used for naturally stated tasks, yet it is unreliable it may succeed on a few test cases but fail at deployment time. We study performance prediction: given a program, either symbolic (e.g. Python) or a prompt executed on an LLM, and a few in-domain examples, predict its performance on unseen tasks from the same domain. We use a simple coin-flip model, treating each pass/fail program execution as a Bernoulli random variable, whose success probability is the programs unknown performance. In this model, performance depends entirely on: 1) the observed execution outcomes on test cases, and 2) a prior over performances. We compile empirical performance priors from a corpus of diverse programs and tasks, and find that performance for symbolic programs (e.g., Python) are all or nothing, while prompt programs have a diffuse prior with many nearly-correct programs. This difference explains why a few passing tests can certify symbolic programs but not prompt programs. Building on this insight, we develop RAP (Retrieved Approximate Prior), which retrieves similar tasks and prompt programs from an existing corpus to construct a proxy prior, which is then used to predict performance. We show RAP achieves solid performances.

LLM-powered coding agents are redefining how real-world software is developed. To drive the research towards better coding agents, we require challenging benchmarks that can rigorously evaluate the ability of such agents to perform various software engineering tasks. However, popular coding benchmarks such as HumanEval and SWE-Bench focus on narrowly scoped tasks such as competition programming and patch generation. In reality, software engineers have to handle a broader set of tasks for real-world software development. To address this gap, we propose OmniCode, a novel software engineering benchmark that contains a broader and more diverse set of task categories beyond code or patch generation. Overall, OmniCode contains 1794 tasks spanning three programming languages (Python, Java, and C++) and four key categories: bug fixing, test generation, code review fixing, and style fixing. In contrast to prior software engineering benchmarks, the tasks in OmniCode are (1) manually validated to eliminate ill-defined problems, and (2) synthetically crafted or recently curated to avoid data leakage issues, presenting a new framework for synthetically generating diverse software tasks from limited real-world data. We evaluate OmniCode with popular agent frameworks such as SWE-Agent and show that while they may perform well on bug fixing for Python, they fall short on tasks such as Test Generation and in languages such as C++ and Java. For instance, SWE-Agent achieves a maximum of 20.9

The task of estimating the world model describing the dynamics of a real world process assumes immense importance for anticipating and preparing for future outcomes. For applications such as video surveillance, robotics applications, autonomous driving, etc. this objective entails synthesizing plausible visual futures, given a few frames of a video to set the visual context. Towards this end, we propose ProgGen, which undertakes the task of video frame prediction by representing the dynamics of the video using a set of neuro-symbolic, human-interpretable set of states (one per frame) by leveraging the inductive biases of Large (Vision) Language Models (LLM/VLM). In particular, ProgGen utilizes LLM/VLM to synthesize programs: (i) to estimate the states of the video, given the visual context (i.e. the frames); (ii) to predict the states corresponding to future time steps by estimating the transition dynamics; (iii) to render the predicted states as visual RGB-frames. Empirical evaluations reveal that our proposed method outperforms competing techniques at the task of video frame prediction in two challenging environments: (i) PhyWorld (ii) Cart Pole. Additionally, ProgGen permits counter-factual reasoning and interpretable video generation attesting to its effectiveness and generalizability for video generation tasks.

Learning how the world works is central to building AI agents that can adapt to complex environments. Traditional world models based on deep-learning demand vast amounts of training data, and do not flexibly update their knowledge from sparse observations. Recent advances in program synthesis using Large Language Models (LLMs) give an alternate approach which learns world models represented as source code, supporting strong generalization from little data. To date, application of program-structured world models remains limited to natural language and grid-world domains. We introduce a novel program synthesis method for effectively modeling complex, non-gridworld domains by representing a world model as an exponentially-weighted product of programmatic experts (PoE-World) synthesized by LLMs. We show that this approach can learn complex, stochastic world models from just a few observations. We evaluate the learned world models by embedding them in a model-based planning agent, demonstrating efficient performance and generalization to unseen levels on Atari's Pong and Montezuma's Revenge.

Making sense of the world and acting in it relies on building simplified mental representations that abstract away aspects of reality. This principle of cognitive mapping is universal to agents with limited resources. Living organisms, people, and algorithms all face the problem of forming functional representations of their world under various computing constraints. In this work, we explore the hypothesis that human resource-efficient planning may arise from representing the world as predictably structured. Building on the metaphor of concepts as programs, we propose that cognitive maps can take the form of generative programs that exploit predictability and redundancy, in contrast to directly encoding spatial layouts. We use a behavioral experiment to show that people who navigate in structured spaces rely on modular planning strategies that align with programmatic map representations. We describe a computational model that predicts human behavior in a variety of structured scenarios. This model infers a small distribution over possible programmatic cognitive maps conditioned on human prior knowledge of the world, and uses this distribution to generate resource-efficient plans. Our models leverages a Large Language Model as an embedding of human priors, implicitly learned through training on a vast corpus of human data. Our model demonstrates improved computational efficiency, requires drastically less memory, and outperforms unstructured planning algorithms with cognitive constraints at predicting human behavior, suggesting that human planning strategies rely on programmatic cognitive maps.

AI for software engineering has made remarkable progress, becoming a notable success within generative AI. Despite this, achieving fully automated software engineering is still a significant challenge, requiring research efforts across both academia and industry. In this position paper, our goal is threefold. First, we provide a taxonomy of measures and tasks to categorize work towards AI software engineering. Second, we outline the key bottlenecks permeating today's approaches. Finally, we highlight promising paths towards making progress on these bottlenecks to guide future research in this rapidly maturing field.

AI for software engineering has made remarkable progress recently, becoming a notable success within generative AI. Despite this, there are still many challenges that need to be addressed before automated software engineering reaches its full potential. It should be possible to reach high levels of automation where humans can focus on the critical decisions of what to build and how to balance difficult tradeoffs while most routine development effort is automated away. Reaching this level of automation will require substantial research and engineering efforts across academia and industry. In this paper, we aim to discuss progress towards this in a threefold manner. First, we provide a structured taxonomy of concrete tasks in AI for software engineering, emphasizing the many other tasks in software engineering beyond code generation and completion. Second, we outline several key bottlenecks that limit current approaches. Finally, we provide an opinionated list of promising research directions toward making progress on these bottlenecks, hoping to inspire future research in this rapidly maturing field.

Partially Observable Markov Decision Processes (POMDPs) model decision making under uncertainty. While there are many approaches to approximately solving POMDPs, we aim to address the problem of learning such models. In particular, we are interested in a subclass of POMDPs wherein the components of the model, including the observation function, reward function, transition function, and initial state distribution function, can be modeled as low-complexity probabilistic graphical models in the form of a short probabilistic program. Our strategy to learn these programs uses an LLM as a prior, generating candidate probabilistic programs that are then tested against the empirical distribution and adjusted through feedback. We experiment on a number of classical toy POMDP problems, simulated MiniGrid domains, and two real mobile-base robotics search domains involving partial observability. Our results show that using an LLM to guide in the construction of a low-complexity POMDP model can be more effective than tabular POMDP learning, behavior cloning, or direct LLM planning.

Maintainable and general software allows developers to build robust applications efficiently, yet achieving these qualities often requires refactoring specialized solutions into reusable components. This challenge becomes particularly relevant as code agents become increasingly accurate at solving isolated programming problems. We investigate code agents' capacity to refactor code in ways supporting growth and reusability. We present both a method and a benchmark for refactoring: Librarian, a sample-and-rerank method for generating reusable libraries, and Minicode, a benchmark where code agents must minimize and refactor multiple independent solutions into a joint library. Compared to state-of-the-art code agents, Librarian achieves strong results on both compression and correctness on Minicode, obtaining compression rates 1.6-2x better than coding agents while also improving correctness. We open-source our code and benchmark at https://code-refactor.github.io/.

Broadly intelligent agents should form task-specific abstractions that selectively expose the essential elements of a task, while abstracting away the complexity of the raw sensorimotor space. In this work, we present Neuro-Symbolic Predicates, a first-order abstraction language that combines the strengths of symbolic and neural knowledge representations. We outline an online algorithm for inventing such predicates and learning abstract world models. We compare our approach to hierarchical reinforcement learning, vision-language model planning, and symbolic predicate invention approaches, on both in- and out-of-distribution tasks across five simulated robotic domains. Results show that our approach offers better sample complexity, stronger out-of-distribution generalization, and improved interpretability.

Systematic compositional generalization - constructing and understanding novel combinations of known building blocks - remains a core challenge for AI systems. Human cognition achieves this flexibility via the interplay of the hippocampus (HPC) and prefrontal cortex (PFC): the hippocampus rapidly encodes episodes, and the prefrontal cortex consolidates them into reusable schemas for reasoning. Drawing on these insights, we present MIRAGE (Meta-Inference with Rules and Abstractions from Generalized Experience), a framework that achieves systematic generalization on compositional tasks. MIRAGE has two interacting modules mirroring the brain's deliberative HPC-PFC loop and intuitive neocortical pattern recognition. (1) The meta-trained Transformer Neural Decomposer, paralleling neocortical "System 1" computation, is trained on a task-agnostic stream of randomly sampled compositional grammars and applies one decomposition step per pass, with successive passes iteratively refining the sequence representation. (2) The Schema Engine, analogous to the HPC-PFC "System 2" loop, dynamically extracts, ranks, and applies reusable schemas, storing variable bindings in episodic memory and expanding them when needed. By explicitly equipping the Transformer component of MIRAGE with actively managed schematic structures, our model performs systematic compositional operations through explicit schema application and transformation, relying solely on frozen weights when solving entirely novel tasks. This approach demonstrates systematic compositional generalization on the SCAN benchmark, achieving > 99

When learning an input-output mapping from very few examples, is it better to first infer a latent function that explains the examples, or is it better to directly predict new test outputs, e.g. using a neural network? We study this question on ARC by training neural models for \emph{induction} (inferring latent functions) and \emph{transduction} (directly predicting the test output for a given test input). We train on synthetically generated variations of Python programs that solve ARC training tasks. We find inductive and transductive models solve different kinds of test problems, despite having the same training problems and sharing the same neural architecture: Inductive program synthesis excels at precise computations, and at composing multiple concepts, while transduction succeeds on fuzzier perceptual concepts. Ensembling them approaches human-level performance on ARC.