As large language models (LLMs) continue to improve in reasoning and decision-making, there is a growing need for realistic and interactive environments where their abilities can be rigorously evaluated. We present VirtualEnv, a next-generation simulation platform built on Unreal Engine 5 that enables fine-grained benchmarking of LLMs in embodied and interactive scenarios. VirtualEnv supports rich agent–environment interactions, including object manipulation, navigation, and adaptive multi-agent collaboration, as well as game-inspired mechanics like escape rooms and procedurally generated environments. We provide a user-friendly API built on top of Unreal Engine, allowing researchers to deploy and control LLM-driven agents using natural language instructions. We integrate large-scale LLMs and vision-language models (VLMs), such as GPT-based models, to generate novel environments and structured tasks from multimodal inputs. Our experiments benchmark the performance of several popular LLMs across tasks of increasing complexity, analyzing differences in adaptability, planning, and multi-agent coordination. We also describe our methodology for procedural task generation, task validation, and real-time environment control. VirtualEnv is released as an open-source platform, we aim to advance research at the intersection of AI and gaming, enable standardized evaluation of LLMs in embodied AI settings, and pave the way for future developments in immersive simulations and interactive entertainment.

Large language models (LLMs) have achieved remarkable performance in recent years but are fundamentally limited by the underlying training data. To improve models beyond the training data, recent works have explored how LLMs can be used to generate synthetic data for autonomous self-improvement. However, successive steps of self-improvement can reach a point of diminishing returns. In this work, we propose a complementary approach towards self-improvement where finetuning is applied to a multiagent society of language models. A group of language models, all starting from the same base model, are independently specialized by updating each one using data generated through multiagent interactions among the models. By training each model on independent sets of data, we illustrate how this approach enables specialization across models and diversification over the set of models. As a result, our overall system is able to preserve diverse reasoning chains and autonomously improve over many more rounds of fine-tuning than single-agent self-improvement methods. We quantitatively illustrate the efficacy of the approach across a wide suite of reasoning tasks.

Large language models (LLMs) have demonstrated remarkable capabilities in language generation, understanding, and few-shot learning in recent years. An extensive body of work has explored how their performance may be further improved through the tools of prompting, ranging from verification, self-consistency, or intermediate scratchpads. In this paper, we present a complementary approach to improve language responses where multiple language model instances propose and debate their individual responses and reasoning processes over multiple rounds to arrive at a common final answer. Our findings indicate that this approach significantly enhances mathematical and strategic reasoning across a number of tasks. We also demonstrate that our approach improves the factual validity of generated content, reducing fallacious answers and hallucinations that contemporary models are prone to. Our approach may be directly applied to existing black-box models and uses identical procedure and prompts for all tasks we investigate. Overall, our findings suggest that such "society of minds" approach has the potential to significantly advance the capabilities of LLMs and pave the way for further breakthroughs in language generation and understanding.

Text-to-image generative models have enabled high-resolution image synthesis across different domains, but require users to specify the content they wish to generate. In this paper, we consider the inverse problem – given a collection of different images, can we discover the generative concepts that represent each image? We present an unsupervised approach to discover generative concepts from a collection of images, disentangling different art styles in paintings, objects, and lighting from kitchen scenes, and discovering image classes given ImageNet images. We show how such generative concepts can accurately represent the content of images, be recombined and composed to generate new artistic and hybrid images, and be further used as a representation for downstream classification tasks.

Open-vocabulary image segmentation is attracting increasing attention due to its critical applications in the real world. Traditional closed-vocabulary segmentation methods are not able to characterize novel objects, whereas several recent open-vocabulary attempts obtain unsatisfactory results, i.e., notable performance reduction on the closedvocabulary and massive demand for extra data. To this end, we propose OPSNet, an omnipotent and data-efficient framework for Open-vocabulary Panoptic Segmentation. Specifically, the exquisitely designed Embedding Modulation module, together with several meticulous components, enables adequate embedding enhancement and information exchange between the segmentation model and the visual-linguistic well-aligned CLIP encoder, resulting in superior segmentation performance under both open- and closed-vocabulary settings with much fewer need of additional data. Extensive experimental evaluations are conducted across multiple datasets (e.g., COCO, ADE20K, Cityscapes, and PascalContext) under various circumstances, where the proposed OPSNet achieves state-of-theart results, which demonstrates the effectiveness and generality of the proposed approach. The project page is https://opsnet-page.github.io.

To make effective decisions in novel environments with long-horizon goals, it is crucial to engage in hierarchical reasoning across spatial and temporal scales. This entails planning abstract subgoal sequences, visually reasoning about the underlying plans, and executing actions in accordance with the devised plan through visual-motor control. We propose Compositional Foundation Models for Hierarchical Planning (HiP), a foundation model which leverages multiple expert foundation model trained on language, vision and action data individually jointly together to solve long-horizon tasks. We use a large language model to construct symbolic plans that are grounded in the environment through a large video diffusion model. Generated video plans are then grounded to visual-motor control, through an inverse dynamics model that infers actions from generated videos. To enable effective reasoning within this hierarchy, we enforce consistency between the models via iterative refinement. We illustrate the efficacy and adaptability of our approach in three different long-horizon table-top manipulation tasks.

Labeling neural network submodules with human-legible descriptions is useful for many downstream tasks: such descriptions can surface failures, guide interventions, and perhaps even explain important model behaviors. To date, most mechanistic descriptions of trained networks have involved small models, narrowly delimited phenomena, and large amounts of human labor. Labeling all human-interpretable sub-computations in models of increasing size and complexity will almost certainly require tools that can generate and validate descriptions automatically. Recently, techniques that use learned models in-the-loop for labeling have begun to gain traction, but methods for evaluating their efficacy are limited and ad-hoc. How should we validate and compare open-ended labeling tools? This paper introduces FIND (Function INterpretation and Description), a benchmark suite for evaluating the building blocks of automated interpretability methods. FIND contains functions that resemble components of trained neural networks, and accompanying descriptions of the kind we seek to generate. The functions are procedurally constructed across textual and numeric domains, and involve a range of real-world complexities, including noise, composition, approximation, and bias. We evaluate methods that use pretrained language models (LMs) to produce code-based and natural language descriptions of function behavior. Additionally, we introduce a new interactive method in which an Automated Interpretability Agent (AIA) generates function descriptions. We find that an AIA, built with an off-the-shelf LM augmented with black-box access to functions, can sometimes infer function structure—acting as a scientist by forming hypotheses, proposing experiments, and updating descriptions in light of new data. However, FIND also reveals that LM-based descriptions capture global function behavior while missing local details. These results suggest that FIND will be useful for characterizing the performance of more sophisticated interpretability methods before they are applied to real-world models.

Deep learning has excelled on complex pattern recognition tasks such as image classification and object recognition. However, it struggles with tasks requiring nontrivial reasoning, such as algorithmic computation. Humans are able to solve such tasks through iterative reasoning – spending more time to think about harder tasks. Most existing neural networks, however, exhibit a fixed computational budget controlled by the neural network architecture, preventing additional computational processing on harder tasks. In this work, we present a new framework for iterative reasoning with neural networks. We train a neural network to parameterize an energy landscape over all outputs, and implement each step of the iterative reasoning as an energy minimization step to find a minimal energy solution. By formulating reasoning as an energy minimization problem, for harder problems that lead to more complex energy landscapes, we may then adjust our underlying computational budget by running a more complex optimization procedure. We empirically illustrate that our iterative reasoning approach can solve more accurate and generalizable algorithmic reasoning tasks in both graph and continuous domains. Finally, we illustrate that our approach can recursively solve algorithmic problems requiring nested reasoning.

Language model (LM) pre-training is useful in many language processing tasks. But can pre-trained LMs be further leveraged for more general machine learning problems? We propose an approach for using LMs to scaffold learning and generalization in general sequential decision-making problems. In this approach, goals and observations are represented as a sequence of embeddings, and a policy network initialized with a pre-trained LM predicts the next action. We demonstrate that this framework enables effective combinatorial generalization across different environments and supervisory modalities. We begin by assuming access to a set of expert demonstrations, and show that initializing policies with LMs and fine-tuning them via behavior cloning improves task completion rates by 43.6% in the VirtualHome environment. Next, we integrate an active data gathering procedure in which agents iteratively interact with the environment, relabel past "failed" experiences with new goals, and update their policies in a self-supervised loop. Active data gathering further improves combinatorial generalization, outperforming the best baseline by 25.1%. Finally, we explain these results by investigating three possible factors underlying the effectiveness of the LM-based policy. We find that sequential input representations (vs. fixed-dimensional feature vectors) and LM-based weight initialization are both important for generalization. Surprisingly, however, the format of the policy inputs encoding (e.g. as a natural language string vs. an arbitrary sequential encoding) has little influence. Together, these results suggest that language modeling induces representations that are useful for modeling not just language, but also goals and plans; these representations can aid learning and generalization even outside of language processing.

We introduce the task of weakly supervised learning for detecting human and object interactions in videos. Our task poses unique challenges as a system does not know what types of human-object interactions are present in a video or the actual spatiotemporal location of the human and the object. To address these challenges, we introduce a contrastive weakly supervised training loss that aims to jointly associate spatiotemporal regions in a video with an action and object vocabulary and encourage temporal continuity of the visual appearance of moving objects as a form of self-supervision. To train our model, we introduce a dataset comprising over 6.5k videos with human-object interaction annotations that have been semi-automatically curated from sentence captions associated with the videos. We demonstrate improved performance over weakly supervised baselines adapted to our task on our video dataset.

Contrastive divergence is a popular method of training energy-based models, but is known to have difficulties with training stability. We propose an adaptation to improve contrastive divergence training by scrutinizing a gradient term that is difficult to calculate and is often left out for convenience. We show that this gradient term is numerically significant and in practice is important to avoid training instabilities, while being tractable to estimate. We further highlight how data augmentation and multi-scale processing can be used to improve model robustness and generation quality. Finally, we empirically evaluate stability of model architectures and show improved performance on a host of benchmarks and use cases, such as image generation, OOD detection, and compositional generation.

Humans are able to rapidly understand scenes by utilizing concepts extracted from prior experience. Such concepts are diverse, and include global scene descriptors, such as the weather or lighting, as well as local scene descriptors, such as the color or size of a particular object. So far, unsupervised discovery of concepts has focused on either modeling the global scene-level or the local object-level factors of variation, but not both. In this work, we propose COMET, which discovers and represents concepts as separate energy functions, enabling us to represent both global concepts as well as objects under a unified framework. COMET discovers energy functions through recomposing the input image, which we find captures independent factors without additional supervision. Sample generation in COMET is formulated as an optimization process on underlying energy functions, enabling us to generate images with permuted and composed concepts. Finally, discovered visual concepts in COMET generalize well, enabling us to compose concepts between separate modalities of images as well as with other concepts discovered by a separate instance of COMET trained on a different dataset. Code and data available at https://energy-based-model.github.io/comet/.

The visual world around us can be described as a structured set of objects and their associated relations. An image of a room may be conjured given only the description of the underlying objects and their associated relations. While there has been significant work on designing deep neural networks which may compose individual objects together, less work has been done on composing the individual relations between objects. A principal difficulty is that while the placement of objects is mutually independent, their relations are entangled and dependent on each other. To circumvent this issue, existing works primarily compose relations by utilizing a holistic encoder, in the form of text or graphs. In this work, we instead propose to represent each relation as an unnormalized density (an energy-based model), enabling us to compose separate relations in a factorized manner. We show that such a factorized decomposition allows the model to both generate and edit scenes that have multiple sets of relations more faithfully. We further show that decomposition enables our model to effectively understand the underlying relational scene structure.

A vital aspect of human intelligence is the ability to compose increasingly complex concepts out of simpler ideas, enabling both rapid learning and adaptation of knowledge. In this paper we show that energy-based models can exhibit this ability by directly combining probability distributions. Samples from the combined distribution correspond to compositions of concepts. For example, given a distribution for smiling faces, and another for male faces, we can combine them to generate smiling male faces. This allows us to generate natural images that simultaneously satisfy conjunctions, disjunctions, and negations of concepts. We evaluate compositional generation abilities of our model on the CelebA dataset of natural faces and synthetic 3D scene images. We also demonstrate other unique advantages of our model, such as the ability to continually learn and incorporate new concepts, or infer compositions of concept properties underlying an image.

Video-based person re-identification matches video clips of people across non-overlapping cameras. Most existing methods tackle this problem by encoding each video frame in its entirety and computing an aggregate representation across all frames. In practice, people are often partially occluded, which can corrupt the extracted features. Instead, we propose a new spatiotemporal attention model that automatically discovers a diverse set of distinctive body parts. This allows useful information to be extracted from all frames without succumbing to occlusions and misalignments. The network learns multiple spatial attention models and employs a diversity regularization term to ensure multiple models do not discover the same body part. Features extracted from local image regions are organized by spatial attention model and are combined using temporal attention. As a result, the network learns latent representations of the face, torso and other body parts using the best available image patches from the entire video sequence. Extensive evaluations on three datasets show that our framework outperforms the state-of-the-art approaches by large margins on multiple metrics.

Searching persons in large-scale image databases with the query of natural language description has important applications in video surveillance. Existing methods mainly focused on searching persons with image-based or attribute-based queries, which have major limitations for a practical usage. In this paper, we study the problem of person search with natural language description. Given the textual description of a person, the algorithm of the person search is required to rank all the samples in the person database then retrieve the most relevant sample corresponding to the queried description. Since there is no person dataset or benchmark with textual description available, we collect a large-scale person description dataset with detailed natural language annotations and person samples from various sources, termed as CUHK Person Description Dataset (CUHK-PEDES). A wide range of possible models and baselines have been evaluated and compared on the person search benchmark. An Recurrent Neural Network with Gated Neural Attention mechanism (GNA-RNN) is proposed to establish the state-of-the art performance on person search.

Existing person re-identification benchmarks and methods mainly focus on matching cropped pedestrian images between queries and candidates. However, it is different from real-world scenarios where the annotations of pedestrian bounding boxes are unavailable and the target person needs to be searched from a gallery of whole scene images. To close the gap, we propose a new deep learning framework for person search. Instead of breaking it down into two separate tasks-pedestrian detection and person re-identification, we jointly handle both aspects in a single convolutional neural network. An Online Instance Matching (OIM) loss function is proposed to train the network effectively, which is scalable to datasets with numerous identities. To validate our approach, we collect and annotate a large-scale benchmark dataset for person search. It contains 18,184 images, 8,432 identities, and 96,143 pedestrian bounding boxes. Experiments show that our framework outperforms other separate approaches, and the proposed OIM loss function converges much faster and better than the conventional Softmax loss.

Textual-visual matching aims at measuring similarities between sentence descriptions and images. Most existing methods tackle this problem without effectively utilizing identity-level annotations. In this paper, we propose an identity-aware two-stage framework for the textual-visual matching problem. Our stage-1 CNN-LSTM network learns to embed cross-modal features with a novel Cross-Modal Cross-Entropy (CMCE) loss. The stage-1 network is able to efficiently screen easy incorrect matchings and also provide initial training point for the stage-2 training. The stage-2 CNN-LSTM network refines the matching results with a latent co-attention mechanism. The spatial attention relates each word with corresponding image regions while the latent semantic attention aligns different sentence structures to make the matching results more robust to sentence structure variations. Extensive experiments on three datasets with identity-level annotations show that our framework outperforms state-of-the-art approaches by large margins.

Articulated human pose estimation is a fundamental yet challenging task in computer vision. The difficulty is particularly pronounced in scale variations of human body parts when camera view changes or severe foreshortening happens. Although pyramid methods are widely used to handle scale changes at inference time, learning feature pyramids in deep convolutional neural networks (DCNNs) is still not well explored. In this work, we design a Pyramid Residual Module (PRMs) to enhance the invariance in scales of DCNNs. Given input features, the PRMs learn convolutional filters on various scales of input features, which are obtained with different subsampling ratios in a multibranch network. Moreover, we observe that it is inappropriate to adopt existing methods to initialize the weights of multi-branch networks, which achieve superior performance than plain networks in many tasks recently. Therefore, we provide theoretic derivation to extend the current weight initialization scheme to multi-branch network structures. We investigate our method on two standard benchmarks for human pose estimation. Our approach obtains state-of-the-art results on both benchmarks. Code is available at https://github.com/bearpaw/PyraNet.