Highly capable AI systems could secretly pursue misaligned goals -- what we call "scheming". Because a scheming AI would deliberately try to hide its misaligned goals and actions, measuring and mitigating scheming requires different strategies than are typically used in ML. We propose that assessing anti-scheming interventions requires at least (1) testing propensity to scheme on far out-of-distribution (OOD) tasks, (2) evaluating whether lack of scheming is driven by situational awareness, and (3) checking for robustness to pre-existing misaligned goals. We use a broad category of "covert actions" -- such as secretly breaking rules or intentionally underperforming in tests -- as a proxy for scheming, and design evaluations for covert actions. We then stress-test deliberative alignment as a case study for anti-scheming. Across 26 OOD evaluations (180+ environments), deliberative alignment reduces covert action rates (OpenAI o3: 13%->0.4%) but does not fully eliminate them. Our mitigation is also able to largely stop agents from pursuing a hidden goal previously trained into the model, but we still find misbehavior after additional red-teaming. We find that models' chain-of-thought (CoT) often demonstrates awareness of being evaluated for alignment, and show causal evidence that this awareness decreases covert behavior, while unawareness increases it. Therefore, we cannot exclude that the observed reductions in covert action rates are at least partially driven by situational awareness. While we rely on human-legible CoT for training, studying situational awareness, and demonstrating clear evidence of misalignment, our ability to rely on this degrades as models continue to depart from reasoning in standard English. We encourage research into alignment mitigations for scheming and their assessment, especially for the adversarial case of deceptive alignment, which this paper does not address.

Large Language Models (LLMs) have recently demonstrated impressive action sequence prediction capabilities but often struggle with dynamic, long-horizon tasks such as real-time strategic games. In a game such as StarCraftII (SC2), agents need to manage resource constraints and adapt to evolving battlefield situations in a partially observable environment. This often overwhelms exisiting LLM-based approaches. To address these challenges, we propose a hierarchical multi-agent framework that employs specialized imitation learning agents under a meta-controller called Strategic Planner (SP). By expert demonstrations, each specialized agent learns a distinctive strategy, such as aerial support or defensive maneuvers, and produces coherent, structured multistep action sequences. The SP then orchestrates these proposals into a single, environmentally adaptive plan that ensures local decisions aligning with long-term strategies. We call this HIMA (Hierarchical Imitation Multi-Agent). We also present TEXTSCII-ALL, a comprehensive SC2 testbed that encompasses all race match combinations in SC2. Our empirical results show that HIMA outperforms state of the arts in strategic clarity, adaptability, and computational efficiency, underscoring the potential of combining specialized imitation modules with meta-level orchestration to develop more robust, general-purpose AI agents.

Reinforcement learning (RL) agents are commonly trained and evaluated in the same environment. In contrast, humans often train in a specialized environment before being evaluated, such as studying a book before taking an exam. The potential of such specialized training environments is still vastly underexplored, despite their capacity to dramatically speed up training. The framework of synthetic environments takes a first step in this direction by meta-learning neural network-based Markov decision processes (MDPs). The initial approach was limited to toy problems and produced environments that did not transfer to unseen RL algorithms. We extend this approach in three ways: Firstly, we modify the meta-learning algorithm to discover environments invariant towards hyperparameter configurations and learning algorithms. Secondly, by leveraging hardware parallelism and introducing a curriculum on an agent's evaluation episode horizon, we can achieve competitive results on several challenging continuous control problems. Thirdly, we surprisingly find that contextual bandits enable training RL agents that transfer well to their evaluation environment, even if it is a complex MDP. Hence, we set up our experiments to train synthetic contextual bandits, which perform on par with synthetic MDPs, yield additional insights into the evaluation environment, and can speed up downstream applications.

Buildings account for 40 % of global energy consumption. A considerable portion of building energy consumption stems from heating, ventilation, and air conditioning (HVAC), and thus implementing smart, energy-efficient HVAC systems has the potential to significantly impact the course of climate change. In recent years, model-free reinforcement learning algorithms have been increasingly assessed for this purpose due to their ability to learn and adapt purely from experience. They have been shown to outperform classical controllers in terms of energy cost and consumption, as well as thermal comfort. However, their weakness lies in their relatively poor data efficiency, requiring long periods of training to reach acceptable policies, making them inapplicable to real-world controllers directly. Hence, common research goals are to improve the learning speed, as well as to improve their ability to generalize, in order to facilitate transfer learning to unseen building environments. In this paper, we take a federated learning approach to training the reinforcement learning controller of an HVAC system. A global control policy is learned by aggregating local policies trained on multiple data centers located in different climate zones. The goal of the policy is to simultaneously minimize energy consumption and maximize thermal comfort. The federated optimization strategy indirectly increases both the rate at which experience data is collected and the variation in the data. We demonstrate through experimental evaluation that these effects lead to a faster learning speed, as well as greater generalization capabilities in the federated policy compared to any individually trained policy.

Retrieval-augmented generation models augment knowledge encoded in a language model by providing additional relevant external knowledge (context) during generation. Although it has been shown that the quantity and quality of context impact the performance of retrieval-augmented generation models during inference, limited research explores how these characteristics affect model training. This paper explores how context quantity and quality during model training affect the performance of Fusion-in-Decoder (FiD), the state-of-the-art retrieval-augmented generation model, in extractive open-domain question answering tasks. Experimental results suggest that FiD models overfit to context quality during training and show suboptimal performance when evaluated on different context quality. Through the experimental results, we also reveal FiD models trained with different context quality have different cross-attention distribution patterns. Specifically, as context quality during training increases, FiD models tend to attend more uniformly to each passage in context. Finally, based on these observations, we propose a method to mitigate overfitting to specific context quality by introducing bias to the cross-attention distribution, which we demonstrate to be effective in improving the performance of FiD models on different context quality.

The landscape of software development has witnessed a paradigm shift with the advent of AI-powered assistants, exemplified by GitHub Copilot. However, existing solutions are not leveraging all the potential capabilities available in an IDE such as building, testing, executing code, git operations, etc. Therefore, they are constrained by their limited capabilities, primarily focusing on suggesting code snippets and file manipulation within a chat-based interface. To fill this gap, we present AutoDev, a fully automated AI-driven software development framework, designed for autonomous planning and execution of intricate software engineering tasks. AutoDev enables users to define complex software engineering objectives, which are assigned to AutoDev's autonomous AI Agents to achieve. These AI agents can perform diverse operations on a codebase, including file editing, retrieval, build processes, execution, testing, and git operations. They also have access to files, compiler output, build and testing logs, static analysis tools, and more. This enables the AI Agents to execute tasks in a fully automated manner with a comprehensive understanding of the contextual information required. Furthermore, AutoDev establishes a secure development environment by confining all operations within Docker containers. This framework incorporates guardrails to ensure user privacy and file security, allowing users to define specific permitted or restricted commands and operations within AutoDev. In our evaluation, we tested AutoDev on the HumanEval dataset, obtaining promising results with 91.5% and 87.8% of Pass@1 for code generation and test generation respectively, demonstrating its effectiveness in automating software engineering tasks while maintaining a secure and user-controlled development environment.

In order for autonomous mobile robots to navigate in human spaces, they must abide by our social norms. Reinforcement learning (RL) has emerged as an effective method to train sequential decision-making policies that are able to respect these norms. However, a large portion of existing work in the field conducts both RL training and testing in simplistic environments. This limits the generalization potential of these models to unseen environments, and the meaningfulness of their reported results. We propose a method to improve the generalization performance of RL social navigation methods using curriculum learning. By employing multiple environment types and by modeling pedestrians using multiple dynamics models, we are able to progressively diversify and escalate difficulty in training. Our results show that the use of curriculum learning in training can be used to achieve better generalization performance than previous training methods. We also show that results presented in many existing state-of-the-art RL social navigation works do not evaluate their methods outside of their training environments, and thus do not reflect their policies' failure to adequately generalize to out-of-distribution scenarios. In response, we validate our training approach on larger and more crowded testing environments than those used in training, allowing for more meaningful measurements of model performance.

Deep reinforcement learning (RL) approaches have been broadly applied to a large number of robotics tasks, such as robot manipulation and autonomous driving. However, an open problem in deep RL is learning policies that are robust to variations in the environment, which is an important condition for such systems to be deployed into real-world, unstructured settings. Curriculum learning is one approach that has been applied to improve generalization performance in both supervised and reinforcement learning domains, but selecting the appropriate curriculum to achieve robustness can be a user-intensive process. In our work, we show that performing probabilistic inference of the underlying curriculum-reward function using Bayesian Optimization can be a promising technique for finding a robust curriculum. We demonstrate that a curriculum found with Bayesian optimization can outperform a vanilla deep RL agent and a hand-engineered curriculum in the domain of autonomous racing with obstacle avoidance. Our code is available at https://github.com/PRISHIta123/Curriculum_RL_for_Driving.

Unsupervised Environment Design (UED) is a paradigm for training generally capable agents to achieve good zero-shot transfer performance. This paradigm hinges on automatically generating a curriculum of training environments. Leading approaches for UED predominantly use randomly generated environment instances to train the agent. While these methods exhibit good zero-shot transfer performance, they often encounter challenges in effectively exploring large design spaces or leveraging previously discovered underlying structures, To address these challenges, we introduce a novel framework based on Hierarchical MDP (Markov Decision Processes). Our approach includes an upper-level teacher's MDP responsible for training a lower-level MDP student agent, guided by the student's performance. To expedite the learning of the upper leavel MDP, we leverage recent advancements in generative modeling to generate synthetic experience dataset for training the teacher agent. Our algorithm, called Synthetically-enhanced Hierarchical Environment Design (SHED), significantly reduces the resource-intensive interactions between the agent and the environment. To validate the effectiveness of SHED, we conduct empirical experiments across various domains, with the goal of developing an efficient and robust agent under limited training resources. Our results show the manifold advantages of SHED and highlight its effectiveness as a potent instrument for curriculum-based learning within the UED framework. This work contributes to exploring the next generation of RL agents capable of adeptly handling an ever-expanding range of complex tasks.

Advancements in reinforcement learning (RL) have demonstrated superhuman performance in complex tasks such as Starcraft, Go, Chess etc. However, knowledge transfer from Artificial "Experts" to humans remain a significant challenge. A promising avenue for such transfer would be the use of curricula. Recent methods in curricula generation focuses on training RL agents efficiently, yet such methods rely on surrogate measures to track student progress, and are not suited for training robots in the real world (or more ambitiously humans). In this paper, we introduce a method named Parameterized Environment Response Model (PERM) that shows promising results in training RL agents in parameterized environments. Inspired by Item Response Theory, PERM seeks to model difficulty of environments and ability of RL agents directly. Given that RL agents and humans are trained more efficiently under the "zone of proximal development", our method generates a curriculum by matching the difficulty of an environment to the current ability of the student. In addition, PERM can be trained offline and does not employ non-stationary measures of student ability, making it suitable for transfer between students. We demonstrate PERM's ability to represent the environment parameter space, and training with RL agents with PERM produces a strong performance in deterministic environments. Lastly, we show that our method is transferable between students, without any sacrifice in training quality.