A major challenge for Multi-Agent Systems (MAS) is enabling agents to adapt dynamically to diverse environments in which opponents and teammates may continually change. Agents trained using conventional methods tend to excel only within the confines of their training cohorts; their performance drops significantly when confronting unfamiliar agents. To address this shortcoming, we introduce Inverse Attention Agents that adopt concepts from the Theory of Mind (ToM) implemented algorithmically using an attention mechanism trained in an end-to-end manner. Crucial to determining the final actions of these agents, the weights in their attention model explicitly represent attention to different goals. We furthermore propose an inverse attention network that deduces the ToM of agents based on observations and prior actions. The network infers the attentional states of other agents, thereby refining the attention weights to adjust the agent's final action. We conduct experiments in a continuous environment, tackling demanding tasks encompassing cooperation, competition, and a blend of both. They demonstrate that the inverse attention network successfully infers the attention of other agents, and that this information improves agent performance. Additional human experiments show that, compared to baseline agent models, our inverse attention agents exhibit superior cooperation with humans and better emulate human behaviors.

We propose a novel class of language models, Latent Thought Models (LTMs), which incorporate explicit latent thought vectors that follow an explicit prior model in latent space. These latent thought vectors guide the autoregressive generation of ground tokens through a Transformer decoder. Training employs a dual-rate optimization process within the classical variational Bayes framework: fast learning of local variational parameters for the posterior distribution of latent vectors (inference-time computation), and slow learning of global decoder parameters. Empirical studies reveal that LTMs possess additional scaling dimensions beyond traditional Large Language Models (LLMs), such as the number of iterations in inference-time computation and number of latent thought vectors. Higher sample efficiency can be achieved by increasing training compute per token, with further gains possible by trading model size for more inference steps. Designed based on these scaling properties, LTMs demonstrate superior sample and parameter efficiency compared to autoregressive models and discrete diffusion models. They significantly outperform these counterparts in validation perplexity and zero-shot language modeling tasks. Additionally, LTMs exhibit emergent few-shot in-context reasoning capabilities that scale with model size, and achieve competitive performance in conditional and unconditional text generation. The project page is available at https://deqiankong.github.io/blogs/ltm.

The hippocampus supports spatial navigation by encoding cognitive maps through collective place cell activity. We model the place cell population as non-negative spatial embeddings derived from the spectral decomposition of multi-step random walk transition kernels. In this framework, inner product or equivalently Euclidean distance between embeddings encode similarity between locations in terms of their transition probability across multiple scales, forming a cognitive map of adjacency. The combination of non-negativity and inner-product structure naturally induces sparsity, providing a principled explanation for the localized firing fields of place cells without imposing explicit constraints. The temporal parameter that defines the diffusion scale also determines field size, aligning with the hippocampal dorsoventral hierarchy. Our approach constructs global representations efficiently through recursive composition of local transitions, enabling smooth, trap-free navigation and preplay-like trajectory generation. Moreover, theta phase arises intrinsically as the angular relation between embeddings, linking spatial and temporal coding within a single representational geometry.

Driving behavior modeling is crucial in autonomous driving systems for preventing traffic accidents. Inverse reinforcement learning (IRL) allows autonomous agents to learn complicated behaviors from expert demonstrations. Similar to how humans learn by trial and error, failed demonstrations can help an agent avoid failures. However, expert and failed demonstrations generally have some common behaviors, which could cause instability in an IRL model. To improve the stability, this work proposes a novel method that introduces time-series labeling for the optimization of IRL to help distinguish the behaviors in demonstrations. Experimental results in a simulated driving environment show that the proposed method converged faster than and outperformed other baseline methods. The results also show consistency for various data balances of the number of expert and failed demonstrations.

In tasks aiming for long-term returns, planning becomes essential. We study generative modeling for planning with datasets repurposed from offline reinforcement learning. Specifically, we identify temporal consistency in the absence of step-wise rewards as one key technical challenge. We introduce the Latent Plan Transformer (LPT), a novel model that leverages a latent variable to connect a Transformer- based trajectory generator and the final return. LPT can be learned with maximum likelihood estimation on trajectory-return pairs. In learning, posterior sampling of the latent variable naturally integrates sub-trajectories to form a consistent abstrac- tion despite the finite context. At test time, the latent variable is inferred from an expected return before policy execution, realizing the idea of planning as inference. Our experiments demonstrate that LPT can discover improved decisions from sub- optimal trajectories, achieving competitive performance across several benchmarks, including Gym-Mujoco, Franka Kitchen, Maze2D, and Connect Four. It exhibits capabilities in nuanced credit assignments, trajectory stitching, and adaptation to environmental contingencies. These results validate that latent variable inference can be a strong alternative to step-wise reward prompting.

Recent conditional and unconditional video generation tasks have been accomplished mainly based on generative adversarial network (GAN), diffusion, and autoregressive models. However, in some circumstances, using only one modality cannot provide enough semantic information. Therefore, in this paper, we propose text-audio to video (TA2V) generation, a new task for generating realistic videos from two different guided modalities, text and audio, which has not been explored much thus far. Compared to image generation, video generation is a harder task because of the complexity of processing higher-dimensional data and scarcer suitable datasets, especially for multimodal video generation. To overcome these limitations, (i) we propose the Text&Audio-guided-Video-Maker (TAgVM) model, which consists of two modules: a text-guided video generator and a text&audio-guided video modifier. (ii) This model uses a 3D VQ-GAN to compress high-dimension video data to a low-dimension discrete sequence, followed by an autoregressive model to guide text-conditional generation in the latent space. Then, we apply a text&audio-guided diffusion model to the generated video scenes, providing additional semantic details corresponding to the audio and text. (iii) We introduce a newly produced music performance video dataset, the University of Rochester Multimodal Music Performance with Video-Audio-Text (URMP-VAT), and a landscape dataset, Landscape with Video-Audio-Text (Landscape-VAT), both of which include three modalities (text, audio, and video) that are aligned with each other. The results demonstrate that our model can create videos with satisfactory quality and semantic information. The source code and datasets are available athttps://github.com/Minglu58/TA2V.

Predicting hard drive failures in datacenters can help avoid wasting resources and waiting time for recovery. Anomaly detection from sensing data is commonly used for predicting failures. Usually, conventional threshold-based anomaly detection methods consider each sensor independently. However, deciding an optimal threshold for each type of sensors is not trivial, especially for large-scale systems in datacenters. To detect failures that cannot conventionally be detected, multimodal anomaly detection becomes crucial integrating sensing data from different types of sensors. This work proposes a correlation-based multimodal anomaly detection approach. This approach is applied to a Network-Attached Storage (NAS) system with multiple hard disk drives (HDDs) and three sensors, which are a thermal camera, a microphone, and system performance logs. The unimodal results show that the auditory and system performance model can detect temporal anomalies, and the thermal model can detect spatial anomalies. The multimodal results show that even with a simple filter and detection algorithms, the multimodal approach was able to detect failure signs before the real failure and also earlier than the auditory unimodal approach.