While in [61], the authors recover the true supervision signals with peer loss, which punishes over-agreement for avoiding overfitting.

Large Language Models (LLMs) can serve as world models to enhance agent decision-making in digital environments by simulating future states and predicting action outcomes, potentially eliminating costly trial-and-error exploration. However, this capability is fundamentally limited by LLM's tendency to hallucination and their reliance on static training knowledge, which could lead to compounding errors that inhibit long-horizon simulations. To systematically investigate whether LLMs are appropriate for world modeling, we probe two core capabilities of world models - future state prediction and reward estimation - through three tasks: next-state identification, full-procedure planning alignment, and milestone transition recognition. Our analysis shows that while LLMs effectively capture immediate next states and identify meaningful state transitions, their performance rapidly degrades in full-procedure planning. This highlights LLMs' limitations in reliably modeling environment dynamics over long horizons. To address these limitations, we propose the Retrieval-augmented World Model (R-WoM), which grounds LLM simulations by incorporating factual, up-to-date knowledge retrieved from external tutorials. Experiments show that R-WoM achieves substantial improvements of up to 25.3% (OSWorld) and 18.1% (WebArena) compared to baselines, with particular advantage in longer-horizon simulations. World models have evolved from early symbolic planning systems to sophisticated neural architectures that learn latent representations of environment dynamics. Model-based reinforcement learning (MBRL) approaches, such as Dreamer v1-3 (Hafner et al., 2019; 2020; 2023) and MuZero (Schrittwieser et al., 2020), learn latent world models to "imagine" trajectories before selecting actions. More recently, Large Language Model (LLM)-based world models (Hao et al., 2023; Wang et al., 2024; Zhang et al., 2024) have emerged as a new paradigm, leveraging large-scale pre-training to reason about action consequences in realistic digital environments.

Employing large language models (LLMs) to enable embodied agents has become popular, yet it presents several limitations in practice. In this work, rather than using LLMs directly as agents, we explore their use as tools for embodied agent learning. Specifically, to train separate agents via offline reinforcement learning (RL), an LLM is used to provide dense reward feedback on individual actions in training datasets. In doing so, we present a consistency-guided reward ensemble framework (CoREN), designed for tackling difficulties in grounding LLM-generated estimates to the target environment domain. The framework employs an adaptive ensemble of spatio-temporally consistent rewards to derive domain-grounded rewards in the training datasets, thus enabling effective offline learning of embodied agents in different environment domains. Experiments with the VirtualHome benchmark demonstrate that CoREN significantly outperforms other offline RL agents, and it also achieves comparable performance to state-of-the-art LLM-based agents with 8B parameters, despite CoREN having only 117M parameters for the agent policy network and using LLMs only for training.

Online Imitation Learning methods struggle with the gap between extensive online exploration space and limited expert trajectories, which hinder efficient exploration due to inaccurate task-aware reward estimation. Inspired by the findings from cognitive neuroscience that task decomposition could facilitate cognitive processing for efficient learning, we hypothesize that an agent could estimate precise task-aware imitation rewards for efficient online exploration by decomposing the target task into the objectives of "what to do" and the mechanisms of "how to do". In this work, we introduce the hybrid Key-state guided Online Imitation (KOI) learning approach, which leverages the integration of semantic and motion key states as guidance for task-aware reward estimation. Initially, we utilize the visual-language models to segment the expert trajectory into semantic key states, indicating the objectives of "what to do". Within the intervals between semantic key states, optical flow is employed to capture motion key states to understand the process of "how to do". By integrating a thorough grasp of both semantic and motion key states, we refine the trajectory-matching reward computation, encouraging task-aware exploration for efficient online imitation learning. Our experiment results prove that our method is more sample efficient in the Meta-World and LIBERO environments. We also conduct real-world robotic manipulation experiments to validate the efficacy of our method, demonstrating the practical applicability of our KOI method.

Materials discovery driven by statistical property models is an iterative decision process, during which an initial data collection is extended with new data proposed by a model-informed acquisition function--with the goal to maximize a certain "reward" over time, such as the maximum property value discovered so far. While the materials science community achieved much progress in developing property models that predict well on average with respect to the training distribution, this form of in-distribution performance measurement is not directly coupled with the discovery reward. This is because an iterative discovery process has a shifting reward distribution that is over-proportionally determined by the model performance for exceptional materials. We demonstrate this problem using the example of bulk modulus maximization among double perovskite oxides. We find that the in-distribution predictive performance suggests random forests as superior to Gaussian process regression, while the results are inverse in terms of the discovery rewards. We argue that the lack of proper performance estimation methods from pre-computed data collections is a fundamental problem for improving data-driven materials discovery, and we propose a novel such estimator that, in contrast to na\"ive reward estimation, successfully predicts Gaussian processes with the "expected improvement" acquisition function as the best out of four options in our demonstrational study for double perovskites. Importantly, it does so without requiring the over thousand ab initio computations that were needed to confirm this prediction.

An intelligent dialogue system in a multi-turn setting should not only generate the responses which are of good quality, but it should also generate the responses which can lead to long-term success of the dialogue. Although, the current approaches improved the response quality, but they over-look the training signals present in the dialogue data. We can leverage these signals to generate the weakly supervised training data for learning dialog policy and reward estimator, and make the policy take actions (generates responses) which can foresee the future direction for a successful (rewarding) conversation. We simulate the dialogue between an agent and a user (modelled similar to an agent with supervised learning objective) to interact with each other. The agent uses dynamic blocking to generate ranked diverse responses and exploration-exploitation to select among the Top-K responses. Each simulated state-action pair is evaluated (works as a weak annotation) with three quality modules: Semantic Relevant, Semantic Coherence and Consistent Flow. Empirical studies with two benchmarks indicate that our model can significantly out-perform the response quality and lead to a successful conversation on both automatic evaluation and human judgement.

Multi-agent reinforcement learning has drawn increasing attention in practice, e.g., robotics and automatic driving, as it can explore optimal policies using samples generated by interacting with the environment. However, high reward uncertainty still remains a problem when we want to train a satisfactory model, because obtaining high-quality reward feedback is usually expensive and even infeasible. To handle this issue, previous methods mainly focus on passive reward correction. At the same time, recent active reward estimation methods have proven to be a recipe for reducing the effect of reward uncertainty. In this paper, we propose a novel Distributional Reward Estimation framework for effective Multi-Agent Reinforcement Learning (DRE-MARL). Our main idea is to design the multi-action-branch reward estimation and policy-weighted reward aggregation for stabilized training. Specifically, we design the multi-action-branch reward estimation to model reward distributions on all action branches. Then we utilize reward aggregation to obtain stable updating signals during training. Our intuition is that consideration of all possible consequences of actions could be useful for learning policies. The superiority of the DRE-MARL is demonstrated using benchmark multi-agent scenarios, compared with the SOTA baselines in terms of both effectiveness and robustness.

Dialogue policy optimization often obtains feedback until task completion in task-oriented dialogue systems. This is insufficient for training intermediate dialogue turns since supervision signals (or rewards) are only provided at the end of dialogues. To address this issue, reward learning has been introduced to learn from state-action pairs of an optimal policy to provide turn-by-turn rewards. This approach requires complete state-action annotations of human-to-human dialogues (i.e., expert demonstrations), which is labor intensive. To overcome this limitation, we propose a novel reward learning approach for semi-supervised policy learning. The proposed approach learns a dynamics model as the reward function which models dialogue progress (i.e., state-action sequences) based on expert demonstrations, either with or without annotations. The dynamics model computes rewards by predicting whether the dialogue progress is consistent with expert demonstrations. We further propose to learn action embeddings for a better generalization of the reward function. The proposed approach outperforms competitive policy learning baselines on MultiWOZ, a benchmark multi-domain dataset.

Cancer is a complex disease, the understanding and treatment of which are being aided through increases in the volume of collected data and in the scale of deployed computing power. Consequently, there is a growing need for the development of data-driven and, in particular, deep learning methods for various tasks such as cancer diagnosis, detection, prognosis, and prediction. Despite recent successes, however, designing high-performing deep learning models for nonimage and nontext cancer data is a time-consuming, trial-and-error, manual task that requires both cancer domain and deep learning expertise. To that end, we develop a reinforcement-learning-based neural architecture search to automate deep-learning-based predictive model development for a class of representative cancer data. We develop custom building blocks that allow domain experts to incorporate the cancer-data-specific characteristics. We show that our approach discovers deep neural network architectures that have significantly fewer trainable parameters, shorter training time, and accuracy similar to or higher than those of manually designed architectures. We study and demonstrate the scalability of our approach on up to 1,024 Intel Knights Landing nodes of the Theta supercomputer at the Argonne Leadership Computing Facility.