Each year, expert-level performance is attained in increasingly-complex multiagent domains, where notable examples include Go, Poker, and StarCraft II. This rapid progression is accompanied by a commensurate need to better understand how such agents attain this performance, to enable their safe deployment, identify limitations, and reveal potential means of improving them. In this paper we take a step back from performance-focused multiagent learning, and instead turn our attention towards agent behavior analysis. We introduce a model-agnostic method for discovery of behavior clusters in multiagent domains, using variational inference to learn a hierarchy of behaviors at the joint and local agent levels. Our framework makes no assumption about agents' underlying learning algorithms, does not require access to their latent states or policies, and is trained using only offline observational data. We illustrate the effectiveness of our method for enabling the coupled understanding of behaviors at the joint and local agent level, detection of behavior changepoints throughout training, discovery of core behavioral concepts, demonstrate the approach's scalability to a high-dimensional multiagent MuJoCo control domain, and also illustrate that the approach can disentangle previously-trained policies in OpenAI's hide-and-seek domain.

High sample complexity has long been a challenge for RL. On the other hand, humans learn to perform tasks not only from interaction or demonstrations, but also by reading unstructured text documents, e.g., instruction manuals. Instruction manuals and wiki pages are among the most abundant data that could inform agents of valuable features and policies or task-specific environmental dynamics and reward structures. Therefore, we hypothesize that the ability to utilize human-written instruction manuals to assist learning policies for specific tasks should lead to a more efficient and better-performing agent. We propose the Read and Reward framework. Read and Reward speeds up RL algorithms on Atari games by reading manuals released by the Atari game developers. Our framework consists of a QA Extraction module that extracts and summarizes relevant information from the manual and a Reasoning module that evaluates object-agent interactions based on information from the manual. An auxiliary reward is then provided to a standard A2C RL agent, when interaction is detected. Experimentally, various RL algorithms obtain significant improvement in performance and training speed when assisted by our design.

We introduce the forward-backward (FB) representation of the dynamics of a reward-free Markov decision process. It provides explicit near-optimal policies for any reward specified a posteriori. During an unsupervised phase, we use reward-free interactions with the environment to learn two representations via off-the-shelf deep learning methods and temporal difference (TD) learning. In the test phase, a reward representation is estimated either from reward observations or an explicit reward description (e.g., a target state). The optimal policy for thatreward is directly obtained from these representations, with no planning.

This paper introduces a new complexity measure for MDPs called maximum expected hitting cost. Unlike the diameter measure is only a function of the transition dynamics, this new measure takes into account the reward dynamics as well. The authors show theoretically that under the same assumptions as previous authors who introduced diameter, this new measure is a tighter upper bound. Furthermore, they show the usefulness of this measure by showing that it can be used to better understand the informativeness of rewards when using potential based reward shaping and they prove theoretically that in a large class of MDPs potential based reward shaping is bounded by a multiplicative factor of 2 on their maximum expected hitting costs. I enjoyed reading this paper. I appreciated the structure that the authors used in this paper which first introduced all the necessary prior work (related to diameter) cosily but thoroughly enough before introducing their contributions.

The paper introduces a new complexity measure for MDPs, the expected hitting costs. In contrast to former complexity measures, the hitting costs also depend on the reward of the MDP and can provide a tighter bound for UCRL2. The theory also provides an intersting connection between reward shapeing and the complexity of a MDP. All reviewers appreciated the strong theoretical contribution of the paper which improves our theoretical understanding of the complexity of MDPs. The reviewers also liked that the paper is well written and establishes connections to reward shaping, a method that has also a highly practical value. All reviewers recommend acceptance and I agree with their assessment.

Reinforcement learning based fine-tuning of large language models (LLMs) on human preferences has been shown to enhance both their capabilities and safety behavior. However, in cases related to safety, without precise instructions to human annotators, the data collected may cause the model to become overly cautious, or to respond in an undesirable style, such as being judgmental. Additionally, as model capabilities and usage patterns evolve, there may be a costly need to add or relabel data to modify safety behavior. We propose a novel preference modeling approach that utilizes AI feedback and only requires a small amount of human data. Our method, Rule Based Rewards (RBR), uses a collection of rules for desired or undesired behaviors (e.g.

OpenAI o1 and DeepSeek R1 achieve or even surpass human expert-level performance in complex domains like mathematics and science, with reinforcement learning (RL) and reasoning playing a crucial role. In autonomous driving, recent end-to-end models have greatly improved planning performance but still struggle with long-tailed problems due to limited common sense and reasoning abilities. Some studies integrate vision-language models (VLMs) into autonomous driving, but they typically rely on pre-trained models with simple supervised fine-tuning (SFT) on driving data, without further exploration of training strategies or optimizations specifically tailored for planning. In this paper, we propose AlphaDrive, a RL and reasoning framework for VLMs in autonomous driving. AlphaDrive introduces four GRPO-based RL rewards tailored for planning and employs a two-stage planning reasoning training strategy that combines SFT with RL. As a result, AlphaDrive significantly improves both planning performance and training efficiency compared to using only SFT or without reasoning. Moreover, we are also excited to discover that, following RL training, AlphaDrive exhibits some emergent multimodal planning capabilities, which is critical for improving driving safety and efficiency. To the best of our knowledge, AlphaDrive is the first to integrate GRPO-based RL with planning reasoning into autonomous driving. Code will be released to facilitate future research.

TL;DR: Risk-aversion is essential in many RL applications (e.g., driving, robotic surgery and finance). Some modified RL frameworks consider risk (e.g., by optimizing a risk-measure of the return instead of its expectation), but pose new algorithmic challenges. Instead, it is often suggested to stick with the old and good RL framework, and just set the rewards such that negative outcomes are amplified. Unfortunately, as discussed below, modeling risk using expectation over redefined rewards is often unnatural, impractical or even mathematically impossible, hence cannot replace explicit optimization of risk-measures. This is consistent with similar results from decision theory, where risk optimization is not equivalent to expected utility maximization.

We introduce a physiological model-based agent as proof-of-principle that it is possible to define a flexible self-preserving system that does not use a reward signal or reward-maximization as an objective. We achieve this by introducing the Self-Preserving Agent (SPA) with a physiological structure where the system can get trapped in an absorbing state if the agent does not solve and execute goal-directed polices. Our agent is defined using new class of Bellman equations called Operator Bellman Equations (OBEs), for encoding jointly non-stationary non-Markovian tasks formalized as a Temporal Goal Markov Decision Process (TGMDP). OBEs produce optimal goal-conditioned spatiotemporal transition operators that map an initial state-time to the final state-times of a policy used to complete a goal, and can also be used to forecast future states in multiple dynamic physiological state-spaces. SPA is equipped with an intrinsic motivation function called the valence function, which quantifies the changes in empowerment (the channel capacity of a transition operator) after following a policy. Because empowerment is a function of a transition operator, there is a natural synergism between empowerment and OBEs: the OBEs create hierarchical transition operators, and the valence function can evaluate hierarchical empowerment change defined on these operators. The valence function can then be used for goal selection, wherein the agent chooses a policy sequence that realizes goal states which produce maximum empowerment gain. In doing so, the agent will seek freedom and avoid internal death-states that undermine its ability to control both external and internal states in the future, thereby exhibiting the capacity of predictive and anticipatory self-preservation. We also compare SPA to Multi-objective RL, and discuss its capacity for symbolic reasoning and life-long learning.

Humans perceive the world in terms of objects and relations between them. In fact, for any given pair of objects, there is a myriad of relations that apply to them. How does the cognitive system learn which relations are useful to characterize the task at hand? And how can it use these representations to build a relational policy to interact effectively with the environment? In this paper we proposed that this problem can be understood through the lens of a sub-field of symbolic machine learning called relational reinforcement learning (RRL). To demonstrate the potential of our approach, we build a simple model of relational policy learning based on a function approximator developed in RRL. We trained and tested our model in three Atari games that required to consider an increasingly number of potential relations: Breakout, Pong and Demon Attack. In each game, our model was able to select adequate relational representations and build a relational policy incrementally. We discuss the relationship between our model with models of relational and analogical reasoning, as well as its limitations and future directions of research.