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.

As embodied agents learn to interact, it is crucial for them to understand when, what, and to whom they should respond. While advances in natrual language processing and speech technologies have enabled conversational agents to focus on what to respond, they still struggle to determine when and to whom they should respond. In this paper, we address the addressee detection (Talking-To-Me, TTM) problem under the egocentric view. Instead of relying solely on short-term audio and video data, we propose a simple architecture SICNet with self/cross-modality attention that leverages long-term social interaction context. By leveraging long-term information, our approach has achieved a mean Average Precision (mAP) of 68.98% on the Ego4D TTM task, surpassing the previous state-of-the-art single-task model by 10.07%. We also conducted a detailed ablation study to demonstrate the effectiveness of each component in the long-term social interaction context.

This work explores the challenging problem of molecule design by framing it as a conditional generative modeling task, where target biological properties or desired chemical constraints serve as conditioning variables. We propose the Latent Prompt Transformer (LPT), a novel generative model comprising three components: (1) a latent vector with a learnable prior distribution modeled by a neural transformation of Gaussian white noise; (2) a molecule generation model based on a causal Transformer, which uses the latent vector as a prompt; and (3) a property prediction model that predicts a molecule's target properties and/or constraint values using the latent prompt. LPT can be learned by maximum likelihood estimation on molecule-property pairs. During property optimization, the latent prompt is inferred from target properties and constraints through posterior sampling and then used to guide the autoregressive molecule generation. After initial training on existing molecules and their properties, we adopt an online learning algorithm to progressively shift the model distribution towards regions that support desired target properties. Experiments demonstrate that LPT not only effectively discovers useful molecules across single-objective, multi-objective, and structure-constrained optimization tasks, but also exhibits strong sample efficiency.

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.

The capability to generate responses with diversity and faithfulness using factual knowledge is paramount for creating a human-like, trustworthy dialogue system. Common strategies either adopt a two-step paradigm, which optimizes knowledge selection and response generation separately, and may overlook the inherent correlation between these two tasks, or leverage conditional variational method to jointly optimize knowledge selection and response generation by employing an inference network. In this paper, we present an end-to-end learning framework, termed Sequential Posterior Inference (SPI), capable of selecting knowledge and generating dialogues by approximately sampling from the posterior distribution. Unlike other methods, SPI does not require the inference network or assume a simple geometry of the posterior distribution. This straightforward and intuitive inference procedure of SPI directly queries the response generation model, allowing for accurate knowledge selection and generation of faithful responses. In addition to modeling contributions, our experimental results on two common dialogue datasets (Wizard of Wikipedia and Holl-E) demonstrate that SPI outperforms previous strong baselines according to both automatic and human evaluation metrics.

Generation of molecules with desired chemical and biological properties such as high drug-likeness, high binding affinity to target proteins, is critical for drug discovery. In this paper, we propose a probabilistic generative model to capture the joint distribution of molecules and their properties. Our model assumes an energy-based model (EBM) in the latent space. Conditional on the latent vector, the molecule and its properties are modeled by a molecule generation model and a property regression model respectively. To search for molecules with desired properties, we propose a sampling with gradual distribution shifting (SGDS) algorithm, so that after learning the model initially on the training data of existing molecules and their properties, the proposed algorithm gradually shifts the model distribution towards the region supported by molecules with desired values of properties. Our experiments show that our method achieves very strong performances on various molecule design tasks.

We study the machine’s understanding of embodied reference: One agent uses both language and gesture to refer to an object to another agent in a shared physical environment. Of note, this new visual task requires understanding multimodal cues with perspective-taking to identify which object is being referred to. To tackle this problem, we introduce YouRefIt, a new crowd-sourced dataset of embodied reference collected in various physical scenes; the dataset contains 4,195 unique reference clips in 432 indoor scenes. To the best of our knowledge, this is the first embodied reference dataset that allows us to study referring expressions in daily physical scenes to understand referential behavior, human communication, and human-robot interaction. We further devise two benchmarks for image-based and video-based embodied reference understanding. Comprehensive baselines and extensive experiments provide the very first result of machine perception on how the referring expressions and gestures affect the embodied reference understanding. Our results provide essential evidence that gestural cues are as critical as language cues in understanding the embodied reference.

Brain-inspired Hyperdimensional (HD) computing emulates cognition tasks by computing with hypervectors rather than traditional numerical values. In HD, an encoder maps inputs to high dimensional vectors (hypervectors) and combines them to generate a model for each existing class. During inference, HD performs the task of reasoning by looking for similarities of the input hypervector and each pre-stored class hypervector However, there is not a unique encoding in HD which can perfectly map inputs to hypervectors. This results in low HD classification accuracy over complex tasks such as speech recognition. In this paper we propose MHD, a multi-encoder hierarchical classifier, which enables HD to take full advantages of multiple encoders without increasing the cost of classification. MHD consists of two HD stages: a main stage and a decider stage. The main stage makes use of multiple classifiers with different encoders to classify a wide range of input data. Each classifier in the main stage can trade between efficiency and accuracy by dynamically varying the hypervectors' dimensions. The decider stage, located before the main stage, learns the difficulty of the input data and selects an encoder within the main stage that will provide the maximum accuracy, while also maximizing the efficiency of the classification task. We test the accuracy/efficiency of the proposed MHD on speech recognition application. Our evaluation shows that MHD can provide a 6.6× improvement in energy efficiency and a 6.3× speedup, as compared to baseline single level HD.

Brain-inspired hyperdimensional (HD) computing emulates cognition tasks by computing with hypervectors as an alternative to computing with numbers. At its very core, HD computing is about manipulating and comparing large patterns, stored in memory as hypervectors: the input symbols are mapped to a hypervector and an associative search is performed for reasoning and classification. For every classification event, an associative memory is in charge of finding the closest match between a set of learned hypervectors and a query hypervector by using a distance metric. Hypervectors with the i.i.d. components qualify a memory-centric architecture to tolerate massive number of errors, hence it eases cooperation of various methodological design approaches for boosting energy efficiency and scalability. This paper proposes architectural designs for hyperdimensional associative memory (HAM) to facilitate energy-efficient, fast, and scalable search operation using three widely-used design approaches. These HAM designs search for the nearest Hamming distance, and linearly scale with the number of dimensions in the hypervectors while exploring a large design space with orders of magnitude higher efficiency. First, we propose a digital CMOS-based HAM (D-HAM) that modularly scales to any dimension. Second, we propose a resistive HAM (R-HAM) that exploits timing discharge characteristic of nonvolatile resistive elements to approximately compute Hamming distances at a lower cost. Finally, we combine such resistive characteristic with a currentbased search method to design an analog HAM (A-HAM) that results in faster and denser alternative. Our experimental results show that R-HAM and A-HAM improve the energy-delay product by 9.6× and 1347× compared to D-HAM while maintaining a moderate accuracy of 94% in language recognition.