State resetting is a fundamental but often overlooked capability of simulators. It supports sample-based planning by allowing resets to previously encountered simulation states, and enables calibration of simulators using real data by resetting to states observed in real-system traces. While often taken for granted, state resetting in complex simulators can be nontrivial: when the simulator comes with latent variables (states), state resetting requires sampling from the posterior over the latent state given the observable history, a.k.a. the belief state (Silver and Veness, 2010). While exact sampling is often infeasible, many approximate belief-state samplers can be constructed, raising the question of how to select among them using only sampling access to the simulator. In this paper, we show that this problem reduces to a general conditional distribution-selection task and develop a new algorithm and analysis under sampling-only access. Building on this reduction, the belief-state selection problem admits two different formulations: latent state-based selection, which directly targets the conditional distribution of the latent state, and observation-based selection, which targets the induced distribution over the observation. Interestingly, these formulations differ in how their guarantees interact with the downstream roll-out methods: perhaps surprisingly, observation-based selection may fail under the most natural roll-out method (which we call Single-Reset) but enjoys guarantees under the less conventional alternative (which we call Repeated-Reset). Together with discussion on issues such as distribution shift and the choice of sampling policies, our paper reveals a rich landscape of algorithmic choices, theoretical nuances, and open questions, in this seemingly simple problem.

Advanced reasoning in LLMs on challenging domains like mathematical reasoning can be tackled using verifiable rewards based reinforced fine-tuning (ReFT). In standard ReFT frameworks, a behavior model generates multiple completions with answers per problem, for the answer to be then scored by a reward function. While such RL post-training methods demonstrate significant performance improvements across challenging reasoning domains, the computational cost of generating completions during training with multiple inference steps makes the training cost non-trivial. To address this, we draw inspiration from off-policy RL, and speculative decoding to introduce a novel ReFT framework, dubbed Nested-ReFT, where a subset of layers of the target model acts as the behavior model to generate off-policy completions during training. The behavior model configured with dynamic layer skipping per batch during training decreases the inference cost compared to the standard ReFT frameworks. Our theoretical analysis shows that Nested-ReFT yields unbiased gradient estimates with controlled variance. Our empirical analysis demonstrates improved computational efficiency measured as tokens/sec across multiple math reasoning benchmarks and model sizes. Additionally, we explore three variants of bias mitigation to minimize the off-policyness in the gradient updates that allows for maintaining performance that matches the baseline ReFT performance.

In this paper, we consider the problem of determining when along a training roll-out feedback from the environment is no longer beneficial, and an intervention such as resetting the agent to the initial state distribution is warranted. We show that such interventions can naturally trade off a small sub-optimality gap for a dramatic decrease in sample complexity. In particular, we focus on the reinforcement learning setting in which the agent has access to a reward signal in addition to either (a) an expert supervisor triggering the e-stop mechanism in real-time or (b) expert state-only demonstrations used to "learn" an automatic e-stop trigger.

In this paper, we propose a simple and efficient method for value model training on long-context reasoning traces. Compared to existing process reward models (PRMs), our method does not require a fine-grained notion of "step," which is difficult to define for long-context reasoning models. By collecting a dataset of 2.5 million reasoning traces, we train a 1.5B token-level value model and apply it to DeepSeek models for improved performance with test-time compute scaling. We find that block-wise value-guided search (VGS) with a final weighted majority vote achieves better test-time scaling than standard methods such as majority voting or best-of-n. Moreover, VGS significantly reduces the inference FLOPs required to achieve the same performance of majority voting. Our dataset, model and codebase are open-sourced.

Data collection in robotics is spread across diverse hardware, and this variation will increase as new hardware is developed. Effective use of this growing body of data requires methods capable of learning from diverse robot embodiments. We consider the setting of training a policy using expert trajectories from a single robot arm (the source), and evaluating on a different robot arm for which no data was collected (the target). We present a data editing scheme termed Shadow, in which the robot during training and evaluation is replaced with a composite segmentation mask of the source and target robots. In this way, the input data distribution at train and test time match closely, enabling robust policy transfer to the new unseen robot while being far more data efficient than approaches that require co-training on large amounts of data from diverse embodiments. We demonstrate that an approach as simple as Shadow is effective both in simulation on varying tasks and robots, and on real robot hardware, where Shadow demonstrates an average of over 2x improvement in success rate compared to the strongest baseline.

This paper studies the problem of multiset prediction, where the task in to predict a multiset of labels out of the set of allowed multisets. The proposed method does sequential predictions of labels and is trained to imitate the optimal oracle strategy. The method is evaluated on the two tasks: MultiMNIST and recognition of multiple objects on the COCO dataset. The paper is clearly written, explains the method and some theoretical properties well. The description of the experiments looks good enough.

A central challenge of the clean energy transition is the development of catalysts for low-emissions technologies. Recent advances in Machine Learning for quantum chemistry drastically accelerate the computation of catalytic activity descriptors such as adsorption energies. Here we introduce AdsorbRL, a Deep Reinforcement Learning agent aiming to identify potential catalysts given a multi-objective binding energy target, trained using offline learning on the Open Catalyst 2020 and Materials Project data sets. We experiment with Deep Q-Network agents to traverse the space of all ~160,000 possible unary, binary and ternary compounds of 55 chemical elements, with very sparse rewards based on adsorption energy known for only between 2,000 and 3,000 catalysts per adsorbate. To constrain the actions space, we introduce Random Edge Traversal and train a single-objective DQN agent on the known states subgraph, which we find strengthens target binding energy by an average of 4.1 eV. We extend this approach to multi-objective, goal-conditioned learning, and train a DQN agent to identify materials with the highest (respectively lowest) adsorption energies for multiple simultaneous target adsorbates. We experiment with Objective Sub-Sampling, a novel training scheme aimed at encouraging exploration in the multi-objective setup, and demonstrate simultaneous adsorption energy improvement across all target adsorbates, by an average of 0.8 eV. Overall, our results suggest strong potential for Deep Reinforcement Learning applied to the inverse catalysts design problem.

Monte Carlo Tree Search (MCTS) is a popular tree-based search strategy within the framework of reinforcement learning (RL), which estimates the optimal value of a state and action by building a tree with Monte Carlo simulation. It has been widely used in sequential decision makings, including scheduling problems, inventory, production management, and real-world games, such as Go, Chess, Tic-tac-toe and Chinese Checkers. See Browne et al. (2012), Fu (2018) and Świechowski et al. (2021) for thorough overviews. MCTS uses little or no domain knowledge and self learns by running more simulations. Many variations have been proposed for MCTS to improve its performance. In particular, deep neural networks are combined into MCTS to achieve a remarkable success in the game of Go (Silver et al. 2016, 2017). A basic MCTS is to build a game tree from the root node in an incremental and asymmetric manner, where nodes correspond to states and edges correspond to possible state-action pairs.

Exciting contemporary machine learning problems have recently been phrased in the classic formalism of tree search -- most famously, the game of Go. Interestingly, the state-space underlying these sequential decision-making problems often posses a more general latent structure than can be captured by a tree. In this work, we develop a probabilistic framework to exploit a search space's latent structure and thereby share information across the search tree. The method is based on a combination of approximate inference in jointly Gaussian models for the explored part of the problem, and an abstraction for the unexplored part that imposes a reduction of complexity ad hoc. We empirically find our algorithm to compare favorably to existing non-probabilistic alternatives in Tic-Tac-Toe and a feature selection application.

A common approach to solving physical-reasoning tasks is to train a value learner on example tasks. A limitation of such an approach is it requires learning about object dynamics solely from reward values assigned to the final state of a rollout of the environment. This study aims to address this limitation by augmenting the reward value with additional supervisory signals about object dynamics. Specifically,we define a distance measure between the trajectory of two target objects, and use this distance measure to characterize the similarity of two environment rollouts.We train the model to correctly rank rollouts according to this measure in addition to predicting the correct reward. Empirically, we find that this approach leads to substantial performance improvements on the PHYRE benchmark for physical reasoning: our approach obtains a new state-of-the-art on that benchmark.