Pretrained foundation models (FMs) have exhibited extraordinary in-context learning performance, allowing zero-shot (or few-shot) generalization to new environments/tasks not encountered during the pretraining. In the case of reinforcement learning (RL), in-context RL (ICRL) emerges when pretraining FMs on decision-making problems in an autoregressivesupervised manner. Nevertheless, the current state-of-the-art ICRL algorithms, such as Algorithm Distillation, Decision Pretrained Transformer and Decision Importance Transformer, impose stringent requirements on the pretraining dataset concerning the behavior (source) policies, context information, and action labels, etc. Notably, these algorithms either demand optimal policies or require varying degrees of well-trained behavior policies for all pretraining environments. This significantly hinders the application of ICRL to realworld scenarios, where acquiring optimal or well-trained policies for a substantial volume of real-world training environments can be prohibitively expensive or even intractable. To overcome this challenge, we introduce a novel approach, termed State-Action Distillation (SAD), that allows to generate an effective pretraining dataset guided solely by random policies. In particular, SAD selects query states and corresponding action labels by distilling the outstanding state-action pairs from the entire state and action spaces by using random policies within a trust horizon, and then inherits the classical autoregressive-supervised mechanism during the pretraining. To the best of our knowledge, this is the first work that enables effective ICRL under (e.g., uniform) random policies and random contexts. We also establish the quantitative analysis of the trustworthiness as well as the performance guarantees of our SAD approach. Moreover, our empirical results across multiple popular ICRL benchmark environments demonstrate that, on average, SAD outperforms the best baseline by 236.3% in the offline evaluation and by 135.2% in the online evaluation.

Thompson sampling is one of the most popular learning algorithms for online sequential decision-making problems and has rich real-world applications. However, current Thompson sampling algorithms are limited by the assumption that the rewards received are uncorrupted, which may not be true in real-world applications where adversarial reward poisoning exists. To make Thompson sampling more reliable, we want to make it robust against adversarial reward poisoning. The main challenge is that one can no longer compute the actual posteriors for the true reward, as the agent can only observe the rewards after corruption. In this work, we solve this problem by computing pseudo-posteriors that are less likely to be manipulated by the attack. We propose robust algorithms based on Thompson sampling for the popular stochastic and contextual linear bandit settings in both cases where the agent is aware or unaware of the budget of the attacker. We theoretically show that our algorithms guarantee near-optimal regret under any attack strategy.

Balancing helpfulness and safety (harmlessness) is a critical challenge in aligning large language models (LLMs). Current approaches often decouple these two objectives, training separate preference models for helpfulness and safety, while framing safety as a constraint within a constrained Markov Decision Process (CMDP) framework. However, these methods can lead to ``safety interference'', where average-based safety constraints compromise the safety of some prompts in favor of others. To address this issue, we propose \textbf{Rectified Policy Optimization (RePO)}, which replaces the average safety constraint with stricter (per prompt) safety constraints. At the core of RePO is a policy update mechanism driven by rectified policy gradients, which penalizes the strict safety violation of every prompt, thereby enhancing safety across nearly all prompts. Our experiments on Alpaca-7B demonstrate that RePO improves the safety alignment and reduces the safety interference compared to baseline methods. Code is available at https://github.com/pxyWaterMoon/RePO.

Energy arbitrage is one of the most profitable sources of income for battery operators, generating revenues by buying and selling electricity at different prices. Forecasting these revenues is challenging due to the inherent uncertainty of electricity prices. Deep reinforcement learning (DRL) emerged in recent years as a promising tool, able to cope with uncertainty by training on large quantities of historical data. However, without access to future electricity prices, DRL agents can only react to the currently observed price and not learn to plan battery dispatch. Therefore, in this study, we combine DRL with time-series forecasting methods from deep learning to enhance the performance on energy arbitrage. We conduct a case study using price data from Alberta, Canada that is characterized by irregular price spikes and highly non-stationary. This data is challenging to forecast even when state-of-the-art deep learning models consisting of convolutional layers, recurrent layers, and attention modules are deployed. Our results show that energy arbitrage with DRL-enabled battery control still significantly benefits from these imperfect predictions, but only if predictors for several horizons are combined. Grouping multiple predictions for the next 24-hour window, accumulated rewards increased by 60% for deep Q-networks (DQN) compared to the experiments without forecasts. We hypothesize that multiple predictors, despite their imperfections, convey useful information regarding the future development of electricity prices through a "majority vote" principle, enabling the DRL agent to learn more profitable control policies.

We propose a conjectural counting formula for the coefficients of the chromatic symmetric function of unit interval graphs using reinforcement learning. The formula counts specific disjoint cycle-tuples in the graphs, referred to as Eschers, which satisfy certain concatenation conditions. These conditions are identified by a reinforcement learning model and are independent of the particular unit interval graph, resulting a universal counting expression.

Self-interested individuals often fail to cooperate, posing a fundamental challenge for multi-agent learning. How can we achieve cooperation among self-interested, independent learning agents? Promising recent work has shown that in certain tasks cooperation can be established between learning-aware agents who model the learning dynamics of each other. Here, we present the first unbiased, higher-derivative-free policy gradient algorithm for learning-aware reinforcement learning, which takes into account that other agents are themselves learning through trial and error based on multiple noisy trials. We then leverage efficient sequence models to condition behavior on long observation histories that contain traces of the learning dynamics of other agents. Training long-context policies with our algorithm leads to cooperative behavior and high returns on standard social dilemmas, including a challenging environment where temporally-extended action coordination is required. Finally, we derive from the iterated prisoner's dilemma a novel explanation for how and when cooperation arises among self-interested learning-aware agents.

Recent advances in multi-agent reinforcement learning (MARL) have created opportunities to solve complex real-world tasks. Cybersecurity is a notable application area, where defending networks against sophisticated adversaries remains a challenging task typically performed by teams of security operators. In this work, we explore novel MARL strategies for building autonomous cyber network defenses that address challenges such as large policy spaces, partial observability, and stealthy, deceptive adversarial strategies. To facilitate efficient and generalized learning, we propose a hierarchical Proximal Policy Optimization (PPO) architecture that decomposes the cyber defense task into specific sub-tasks like network investigation and host recovery. Our approach involves training sub-policies for each sub-task using PPO enhanced with domain expertise. These sub-policies are then leveraged by a master defense policy that coordinates their selection to solve complex network defense tasks. Furthermore, the sub-policies can be fine-tuned and transferred with minimal cost to defend against shifts in adversarial behavior or changes in network settings. We conduct extensive experiments using CybORG Cage 4, the state-of-the-art MARL environment for cyber defense. Comparisons with multiple baselines across different adversaries show that our hierarchical learning approach achieves top performance in terms of convergence speed, episodic return, and several interpretable metrics relevant to cybersecurity, including the fraction of clean machines on the network, precision, and false positives on recoveries.

Inventory control in modern supply chains has attracted significant attention due to the increasing number of disruptive shocks and the challenges posed by complex dynamics, uncertainties, and limited collaboration. Traditional methods, which often rely on static parameters, struggle to adapt to changing environments. This paper proposes a Multi-Agent Reinforcement Learning (MARL) framework with Graph Neural Networks (GNNs) for state representation to address these limitations. Our approach redefines the action space by parameterizing heuristic inventory control policies, making it adaptive as the parameters dynamically adjust based on system conditions. By leveraging the inherent graph structure of supply chains, our framework enables agents to learn the system's topology, and we employ a centralized learning, decentralized execution scheme that allows agents to learn collaboratively while overcoming information-sharing constraints. Additionally, we incorporate global mean pooling and regularization techniques to enhance performance. We test the capabilities of our proposed approach on four different supply chain configurations and conduct a sensitivity analysis. This work paves the way for utilizing MARL-GNN frameworks to improve inventory management in complex, decentralized supply chain environments.

The offline-to-online (O2O) paradigm in reinforcement learning (RL) utilizes pre-trained models on offline datasets for subsequent online fine-tuning. However, conventional O2O RL algorithms typically require maintaining and retraining the large offline datasets to mitigate the effects of out-of-distribution (OOD) data, which limits their efficiency in exploiting online samples. To address this challenge, we introduce a new paradigm called SAMG: State-Action-Conditional Offline-to-Online Reinforcement Learning with Offline Model Guidance. In particular, rather than directly training on offline data, SAMG freezes the pre-trained offline critic to provide offline values for each state-action pair to deliver compact offline information. This framework eliminates the need for retraining with offline data by freezing and leveraging these values of the offline model. These are then incorporated with the online target critic using a Bellman equation weighted by a policy state-action-aware coefficient. This coefficient, derived from a conditional variational auto-encoder (C-VAE), aims to capture the reliability of the offline data on a state-action level. SAMG could be easily integrated with existing Q-function based O2O RL algorithms. Theoretical analysis shows good optimality and lower estimation error of SAMG. Empirical evaluations demonstrate that SAMG outperforms four state-of-the-art O2O RL algorithms in the D4RL benchmark.

Unsupervised skill discovery carries the promise that an intelligent agent can learn reusable skills through autonomous, reward-free environment interaction. Existing unsupervised skill discovery methods learn skills by encouraging distinguishable behaviors that cover diverse states. However, in complex environments with many state factors (e.g., household environments with many objects), learning skills that cover all possible states is impossible, and naively encouraging state diversity often leads to simple skills that are not ideal for solving downstream tasks. This work introduces Skill Discovery from Local Dependencies (Skild), which leverages state factorization as a natural inductive bias to guide the skill learning process. The key intuition guiding Skild is that skills that induce diverse interactions between state factors are often more valuable for solving downstream tasks. To this end, Skild develops a novel skill learning objective that explicitly encourages the mastering of skills that effectively induce different interactions within an environment. We evaluate Skild in several domains with challenging, long-horizon sparse reward tasks including a realistic simulated household robot domain, where Skild successfully learns skills with clear semantic meaning and shows superior performance compared to existing unsupervised reinforcement learning methods that only maximize state coverage.