Stephanie Milani

Micah Carroll, Orr Paradise, Jessy Lin et al. · berkeley

Randomly masking and predicting word tokens has been a successful approach in pre-training language models for a variety of downstream tasks. In this work, we observe that the same idea also applies naturally to sequential decision-making, where many well-studied tasks like behavior cloning, offline reinforcement learning, inverse dynamics, and waypoint conditioning correspond to different sequence maskings over a sequence of states, actions, and returns. We introduce the UniMASK framework, which provides a unified way to specify models which can be trained on many different sequential decision-making tasks. We show that a single UniMASK model is often capable of carrying out many tasks with performance similar to or better than single-task models. Additionally, after fine-tuning, our UniMASK models consistently outperform comparable single-task models. Our code is publicly available at https://github.com/micahcarroll/uniMASK.

Stephanie Milani, Anssi Kanervisto, Karolis Ramanauskas et al. · berkeley

To facilitate research in the direction of fine-tuning foundation models from human feedback, we held the MineRL BASALT Competition on Fine-Tuning from Human Feedback at NeurIPS 2022. The BASALT challenge asks teams to compete to develop algorithms to solve tasks with hard-to-specify reward functions in Minecraft. Through this competition, we aimed to promote the development of algorithms that use human feedback as channels to learn the desired behavior. We describe the competition and provide an overview of the top solutions. We conclude by discussing the impact of the competition and future directions for improvement.

Micah Carroll, Jessy Lin, Orr Paradise et al. · berkeley

Randomly masking and predicting word tokens has been a successful approach in pre-training language models for a variety of downstream tasks. In this work, we observe that the same idea also applies naturally to sequential decision making, where many well-studied tasks like behavior cloning, offline RL, inverse dynamics, and waypoint conditioning correspond to different sequence maskings over a sequence of states, actions, and returns. We introduce the FlexiBiT framework, which provides a unified way to specify models which can be trained on many different sequential decision making tasks. We show that a single FlexiBiT model is simultaneously capable of carrying out many tasks with performance similar to or better than specialized models. Additionally, we show that performance can be further improved by fine-tuning our general model on specific tasks of interest.

Rohin Shah, Steven H. Wang, Cody Wild et al. · berkeley

We held the first-ever MineRL Benchmark for Agents that Solve Almost-Lifelike Tasks (MineRL BASALT) Competition at the Thirty-fifth Conference on Neural Information Processing Systems (NeurIPS 2021). The goal of the competition was to promote research towards agents that use learning from human feedback (LfHF) techniques to solve open-world tasks. Rather than mandating the use of LfHF techniques, we described four tasks in natural language to be accomplished in the video game Minecraft, and allowed participants to use any approach they wanted to build agents that could accomplish the tasks. Teams developed a diverse range of LfHF algorithms across a variety of possible human feedback types. The three winning teams implemented significantly different approaches while achieving similar performance. Interestingly, their approaches performed well on different tasks, validating our choice of tasks to include in the competition. While the outcomes validated the design of our competition, we did not get as many participants and submissions as our sister competition, MineRL Diamond. We speculate about the causes of this problem and suggest improvements for future iterations of the competition.

Seth Karten, Jake Grigsby, Tersoo Upaa et al.

We present the PokeAgent Challenge, a large-scale benchmark for decision-making research built on Pokemon's multi-agent battle system and expansive role-playing game (RPG) environment. Partial observability, game-theoretic reasoning, and long-horizon planning remain open problems for frontier AI, yet few benchmarks stress all three simultaneously under realistic conditions. PokeAgent targets these limitations at scale through two complementary tracks: our Battling Track, which calls for strategic reasoning and generalization under partial observability in competitive Pokemon battles, and our Speedrunning Track, which requires long-horizon planning and sequential decision-making in the Pokemon RPG. Our Battling Track supplies a dataset of 20M+ battle trajectories alongside a suite of heuristic, RL, and LLM-based baselines capable of high-level competitive play. Our Speedrunning Track provides the first standardized evaluation framework for RPG speedrunning, including an open-source multi-agent orchestration system for modular, reproducible comparisons of harness-based LLM approaches. Our NeurIPS 2025 competition validates both the quality of our resources and the research community's interest in Pokemon, with over 100 teams competing across both tracks and winning solutions detailed in our paper. Participant submissions and our baselines reveal considerable gaps between generalist (LLM), specialist (RL), and elite human performance. Analysis against the BenchPress evaluation matrix shows that Pokemon battling is nearly orthogonal to standard LLM benchmarks, measuring capabilities not captured by existing suites and positioning Pokemon as an unsolved benchmark that can drive RL and LLM research forward. We transition to a living benchmark with a live leaderboard for Battling and self-contained evaluation for Speedrunning at https://pokeagentchallenge.com.

Karolis Jucys, George Adamopoulos, Mehrab Hamidi et al.

Understanding the mechanisms behind decisions taken by large foundation models in sequential decision making tasks is critical to ensuring that such systems operate transparently and safely. In this work, we perform exploratory analysis on the Video PreTraining (VPT) Minecraft playing agent, one of the largest open-source vision-based agents. We aim to illuminate its reasoning mechanisms by applying various interpretability techniques. First, we analyze the attention mechanism while the agent solves its training task - crafting a diamond pickaxe. The agent pays attention to the last four frames and several key-frames further back in its six-second memory. This is a possible mechanism for maintaining coherence in a task that takes 3-10 minutes, despite the short memory span. Secondly, we perform various interventions, which help us uncover a worrying case of goal misgeneralization: VPT mistakenly identifies a villager wearing brown clothes as a tree trunk when the villager is positioned stationary under green tree leaves, and punches it to death.

Stephanie Milani, Zhicheng Zhang, Nicholay Topin et al.

Many recent breakthroughs in multi-agent reinforcement learning (MARL) require the use of deep neural networks, which are challenging for human experts to interpret and understand. On the other hand, existing work on interpretable reinforcement learning (RL) has shown promise in extracting more interpretable decision tree-based policies from neural networks, but only in the single-agent setting. To fill this gap, we propose the first set of algorithms that extract interpretable decision-tree policies from neural networks trained with MARL. The first algorithm, IVIPER, extends VIPER, a recent method for single-agent interpretable RL, to the multi-agent setting. We demonstrate that IVIPER learns high-quality decision-tree policies for each agent. To better capture coordination between agents, we propose a novel centralized decision-tree training algorithm, MAVIPER. MAVIPER jointly grows the trees of each agent by predicting the behavior of the other agents using their anticipated trees, and uses resampling to focus on states that are critical for its interactions with other agents. We show that both algorithms generally outperform the baselines and that MAVIPER-trained agents achieve better-coordinated performance than IVIPER-trained agents on three different multi-agent particle-world environments.

Aravind Venugopal, Stephanie Milani, Fei Fang et al.

Multi-agent reinforcement learning (MARL) methods often suffer from high sample complexity, limiting their use in real-world problems where data is sparse or expensive to collect. Although latent-variable world models have been employed to address this issue by generating abundant synthetic data for MARL training, most of these models cannot encode vital global information available during training into their latent states, which hampers learning efficiency. The few exceptions that incorporate global information assume centralized execution of their learned policies, which is impractical in many applications with partial observability. We propose a novel model-based MARL algorithm, MABL (Multi-Agent Bi-Level world model), that learns a bi-level latent-variable world model from high-dimensional inputs. Unlike existing models, MABL is capable of encoding essential global information into the latent states during training while guaranteeing the decentralized execution of learned policies. For each agent, MABL learns a global latent state at the upper level, which is used to inform the learning of an agent latent state at the lower level. During execution, agents exclusively use lower-level latent states and act independently. Crucially, MABL can be combined with any model-free MARL algorithm for policy learning. In our empirical evaluation with complex discrete and continuous multi-agent tasks including SMAC, Flatland, and MAMuJoCo, MABL surpasses SOTA multi-agent latent-variable world models in both sample efficiency and overall performance.

Stephanie Milani, Arthur Juliani, Ida Momennejad et al.

We aim to understand how people assess human likeness in navigation produced by people and artificially intelligent (AI) agents in a video game. To this end, we propose a novel AI agent with the goal of generating more human-like behavior. We collect hundreds of crowd-sourced assessments comparing the human-likeness of navigation behavior generated by our agent and baseline AI agents with human-generated behavior. Our proposed agent passes a Turing Test, while the baseline agents do not. By passing a Turing Test, we mean that human judges could not quantitatively distinguish between videos of a person and an AI agent navigating. To understand what people believe constitutes human-like navigation, we extensively analyze the justifications of these assessments. This work provides insights into the characteristics that people consider human-like in the context of goal-directed video game navigation, which is a key step for further improving human interactions with AI agents.

Stephanie Milani, Anssi Kanervisto, Karolis Ramanauskas et al.

The MineRL BASALT competition has served to catalyze advances in learning from human feedback through four hard-to-specify tasks in Minecraft, such as create and photograph a waterfall. Given the completion of two years of BASALT competitions, we offer to the community a formalized benchmark through the BASALT Evaluation and Demonstrations Dataset (BEDD), which serves as a resource for algorithm development and performance assessment. BEDD consists of a collection of 26 million image-action pairs from nearly 14,000 videos of human players completing the BASALT tasks in Minecraft. It also includes over 3,000 dense pairwise human evaluations of human and algorithmic agents. These comparisons serve as a fixed, preliminary leaderboard for evaluating newly-developed algorithms. To enable this comparison, we present a streamlined codebase for benchmarking new algorithms against the leaderboard. In addition to presenting these datasets, we conduct a detailed analysis of the data from both datasets to guide algorithm development and evaluation. The released code and data are available at https://github.com/minerllabs/basalt-benchmark .

Zhuorui Ye, Stephanie Milani, Geoffrey J. Gordon et al.

Recent advances in reinforcement learning (RL) have predominantly leveraged neural network policies for decision-making, yet these models often lack interpretability, posing challenges for stakeholder comprehension and trust. Concept bottleneck models offer an interpretable alternative by integrating human-understandable concepts into policies. However, prior work assumes that concept annotations are readily available during training. For RL, this requirement poses a significant limitation: it necessitates continuous real-time concept annotation, which either places an impractical burden on human annotators or incurs substantial costs in API queries and inference time when employing automated labeling methods. To overcome this limitation, we introduce a novel training scheme that enables RL agents to efficiently learn a concept-based policy by only querying annotators to label a small set of data. Our algorithm, LICORICE, involves three main contributions: interleaving concept learning and RL training, using an ensemble to actively select informative data points for labeling, and decorrelating the concept data. We show how LICORICE reduces human labeling efforts to 500 or fewer concept labels in three environments, and 5000 or fewer in two more complex environments, all at no cost to performance. We also explore the use of VLMs as automated concept annotators, finding them effective in some cases but imperfect in others. Our work significantly reduces the annotation burden for interpretable RL, making it more practical for real-world applications that necessitate transparency.

Raja Farrukh Ali, Stephanie Milani, John Woods et al.

Reinforcement learning (RL) has recently shown promise in predicting Alzheimer's disease (AD) progression due to its unique ability to model domain knowledge. However, it is not clear which RL algorithms are well-suited for this task. Furthermore, these methods are not inherently explainable, limiting their applicability in real-world clinical scenarios. Our work addresses these two important questions. Using a causal, interpretable model of AD, we first compare the performance of four contemporary RL algorithms in predicting brain cognition over 10 years using only baseline (year 0) data. We then apply SHAP (SHapley Additive exPlanations) to explain the decisions made by each algorithm in the model. Our approach combines interpretability with explainability to provide insights into the key factors influencing AD progression, offering both global and individual, patient-level analysis. Our findings show that only one of the RL methods is able to satisfactorily model disease progression, but the post-hoc explanations indicate that all methods fail to properly capture the importance of amyloid accumulation, one of the pathological hallmarks of Alzheimer's disease. Our work aims to merge predictive accuracy with transparency, assisting clinicians and researchers in enhancing disease progression modeling for informed healthcare decisions. Code is available at https://github.com/rfali/xrlad.

Naveen Raman, Stephanie Milani, Fei Fang

Training interpretable concept-based policies requires practitioners to manually select which human-understandable concepts an agent should reason with when making sequential decisions. This selection demands domain expertise, is time-consuming and costly, scales poorly with the number of candidates, and provides no performance guarantees. To overcome this limitation, we propose the first algorithms for principled automatic concept selection in sequential decision-making. Our key insight is that concept selection can be viewed through the lens of state abstraction: intuitively, a concept is decision-relevant if removing it would cause the agent to confuse states that require different actions. As a result, agents should rely on decision-relevant concepts; states with the same concept representation should share the same optimal action, which preserves the optimal decision structure of the original state space. This perspective leads to the Decision-Relevant Selection (DRS) algorithm, which selects a subset of concepts from a candidate set, along with performance bounds relating the selected concepts to the performance of the resulting policy. Empirically, DRS automatically recovers manually curated concept sets while matching or exceeding their performance, and improves the effectiveness of test-time concept interventions across reinforcement learning benchmarks and real-world healthcare environments.

Anssi Kanervisto, Stephanie Milani, Karolis Ramanauskas et al.

Reinforcement learning competitions advance the field by providing appropriate scope and support to develop solutions toward a specific problem. To promote the development of more broadly applicable methods, organizers need to enforce the use of general techniques, the use of sample-efficient methods, and the reproducibility of the results. While beneficial for the research community, these restrictions come at a cost -- increased difficulty. If the barrier for entry is too high, many potential participants are demoralized. With this in mind, we hosted the third edition of the MineRL ObtainDiamond competition, MineRL Diamond 2021, with a separate track in which we permitted any solution to promote the participation of newcomers. With this track and more extensive tutorials and support, we saw an increased number of submissions. The participants of this easier track were able to obtain a diamond, and the participants of the harder track progressed the generalizable solutions in the same task.

Stephanie Milani, Nicholay Topin, Manuela Veloso et al.

Explainable reinforcement learning (XRL) is an emerging subfield of explainable machine learning that has attracted considerable attention in recent years. The goal of XRL is to elucidate the decision-making process of learning agents in sequential decision-making settings. In this survey, we propose a novel taxonomy for organizing the XRL literature that prioritizes the RL setting. We overview techniques according to this taxonomy. We point out gaps in the literature, which we use to motivate and outline a roadmap for future work.

Rohin Shah, Cody Wild, Steven H. Wang et al.

The last decade has seen a significant increase of interest in deep learning research, with many public successes that have demonstrated its potential. As such, these systems are now being incorporated into commercial products. With this comes an additional challenge: how can we build AI systems that solve tasks where there is not a crisp, well-defined specification? While multiple solutions have been proposed, in this competition we focus on one in particular: learning from human feedback. Rather than training AI systems using a predefined reward function or using a labeled dataset with a predefined set of categories, we instead train the AI system using a learning signal derived from some form of human feedback, which can evolve over time as the understanding of the task changes, or as the capabilities of the AI system improve. The MineRL BASALT competition aims to spur forward research on this important class of techniques. We design a suite of four tasks in Minecraft for which we expect it will be hard to write down hardcoded reward functions. These tasks are defined by a paragraph of natural language: for example, "create a waterfall and take a scenic picture of it", with additional clarifying details. Participants must train a separate agent for each task, using any method they want. Agents are then evaluated by humans who have read the task description. To help participants get started, we provide a dataset of human demonstrations on each of the four tasks, as well as an imitation learning baseline that leverages these demonstrations. Our hope is that this competition will improve our ability to build AI systems that do what their designers intend them to do, even when the intent cannot be easily formalized. Besides allowing AI to solve more tasks, this can also enable more effective regulation of AI systems, as well as making progress on the value alignment problem.

William Hebgen Guss, Stephanie Milani, Nicholay Topin et al.

Reinforcement learning competitions have formed the basis for standard research benchmarks, galvanized advances in the state-of-the-art, and shaped the direction of the field. Despite this, a majority of challenges suffer from the same fundamental problems: participant solutions to the posed challenge are usually domain-specific, biased to maximally exploit compute resources, and not guaranteed to be reproducible. In this paper, we present a new framework of competition design that promotes the development of algorithms that overcome these barriers. We propose four central mechanisms for achieving this end: submission retraining, domain randomization, desemantization through domain obfuscation, and the limitation of competition compute and environment-sample budget. To demonstrate the efficacy of this design, we proposed, organized, and ran the MineRL 2020 Competition on Sample-Efficient Reinforcement Learning. In this work, we describe the organizational outcomes of the competition and show that the resulting participant submissions are reproducible, non-specific to the competition environment, and sample/resource efficient, despite the difficult competition task.

Nicholay Topin, Stephanie Milani, Fei Fang et al.

Current work in explainable reinforcement learning generally produces policies in the form of a decision tree over the state space. Such policies can be used for formal safety verification, agent behavior prediction, and manual inspection of important features. However, existing approaches fit a decision tree after training or use a custom learning procedure which is not compatible with new learning techniques, such as those which use neural networks. To address this limitation, we propose a novel Markov Decision Process (MDP) type for learning decision tree policies: Iterative Bounding MDPs (IBMDPs). An IBMDP is constructed around a base MDP so each IBMDP policy is guaranteed to correspond to a decision tree policy for the base MDP when using a method-agnostic masking procedure. Because of this decision tree equivalence, any function approximator can be used during training, including a neural network, while yielding a decision tree policy for the base MDP. We present the required masking procedure as well as a modified value update step which allows IBMDPs to be solved using existing algorithms. We apply this procedure to produce IBMDP variants of recent reinforcement learning methods. We empirically show the benefits of our approach by solving IBMDPs to produce decision tree policies for the base MDPs.

William H. Guss, Mario Ynocente Castro, Sam Devlin et al.

Although deep reinforcement learning has led to breakthroughs in many difficult domains, these successes have required an ever-increasing number of samples, affording only a shrinking segment of the AI community access to their development. Resolution of these limitations requires new, sample-efficient methods. To facilitate research in this direction, we propose this second iteration of the MineRL Competition. The primary goal of the competition is to foster the development of algorithms which can efficiently leverage human demonstrations to drastically reduce the number of samples needed to solve complex, hierarchical, and sparse environments. To that end, participants compete under a limited environment sample-complexity budget to develop systems which solve the MineRL ObtainDiamond task in Minecraft, a sequential decision making environment requiring long-term planning, hierarchical control, and efficient exploration methods. The competition is structured into two rounds in which competitors are provided several paired versions of the dataset and environment with different game textures and shaders. At the end of each round, competitors submit containerized versions of their learning algorithms to the AIcrowd platform where they are trained from scratch on a hold-out dataset-environment pair for a total of 4-days on a pre-specified hardware platform. In this follow-up iteration to the NeurIPS 2019 MineRL Competition, we implement new features to expand the scale and reach of the competition. In response to the feedback of the previous participants, we introduce a second minor track focusing on solutions without access to environment interactions of any kind except during test-time. Further we aim to prompt domain agnostic submissions by implementing several novel competition mechanics including action-space randomization and desemantization of observations and actions.

Brandon Houghton, Stephanie Milani, Nicholay Topin et al.

To encourage the development of methods with reproducible and robust training behavior, we propose a challenge paradigm where competitors are evaluated directly on the performance of their learning procedures rather than pre-trained agents. Since competition organizers re-train proposed methods in a controlled setting they can guarantee reproducibility, and -- by retraining submissions using a held-out test set -- help ensure generalization past the environments on which they were trained.

Stephanie Milani, Nicholay Topin, Brandon Houghton et al.

To facilitate research in the direction of sample efficient reinforcement learning, we held the MineRL Competition on Sample Efficient Reinforcement Learning Using Human Priors at the Thirty-third Conference on Neural Information Processing Systems (NeurIPS 2019). The primary goal of this competition was to promote the development of algorithms that use human demonstrations alongside reinforcement learning to reduce the number of samples needed to solve complex, hierarchical, and sparse environments. We describe the competition, outlining the primary challenge, the competition design, and the resources that we provided to the participants. We provide an overview of the top solutions, each of which use deep reinforcement learning and/or imitation learning. We also discuss the impact of our organizational decisions on the competition and future directions for improvement.

John Winder, Stephanie Milani, Matthew Landen et al.

We introduce an algorithm for model-based hierarchical reinforcement learning to acquire self-contained transition and reward models suitable for probabilistic planning at multiple levels of abstraction. We call this framework Planning with Abstract Learned Models (PALM). By representing subtasks symbolically using a new formal structure, the lifted abstract Markov decision process (L-AMDP), PALM learns models that are independent and modular. Through our experiments, we show how PALM integrates planning and execution, facilitating a rapid and efficient learning of abstract, hierarchical models. We also demonstrate the increased potential for learned models to be transferred to new and related tasks.

William H. Guss, Cayden Codel, Katja Hofmann et al.

Though deep reinforcement learning has led to breakthroughs in many difficult domains, these successes have required an ever-increasing number of samples. As state-of-the-art reinforcement learning (RL) systems require an exponentially increasing number of samples, their development is restricted to a continually shrinking segment of the AI community. Likewise, many of these systems cannot be applied to real-world problems, where environment samples are expensive. Resolution of these limitations requires new, sample-efficient methods. To facilitate research in this direction, we introduce the MineRL Competition on Sample Efficient Reinforcement Learning using Human Priors. The primary goal of the competition is to foster the development of algorithms which can efficiently leverage human demonstrations to drastically reduce the number of samples needed to solve complex, hierarchical, and sparse environments. To that end, we introduce: (1) the Minecraft ObtainDiamond task, a sequential decision making environment requiring long-term planning, hierarchical control, and efficient exploration methods; and (2) the MineRL-v0 dataset, a large-scale collection of over 60 million state-action pairs of human demonstrations that can be resimulated into embodied trajectories with arbitrary modifications to game state and visuals. Participants will compete to develop systems which solve the ObtainDiamond task with a limited number of samples from the environment simulator, Malmo. The competition is structured into two rounds in which competitors are provided several paired versions of the dataset and environment with different game textures. At the end of each round, competitors will submit containerized versions of their learning algorithms and they will then be trained/evaluated from scratch on a hold-out dataset-environment pair for a total of 4-days on a prespecified hardware platform.