Michael Luck

Marc Serramia, William Seymour, Natalia Criado et al.

Smart devices, such as smart speakers, are becoming ubiquitous, and users expect these devices to act in accordance with their preferences. In particular, since these devices gather and manage personal data, users expect them to adhere to their privacy preferences. However, the current approach of gathering these preferences consists in asking the users directly, which usually triggers automatic responses failing to capture their true preferences. In response, in this paper we present a collaborative filtering approach to predict user preferences as norms. These preference predictions can be readily adopted or can serve to assist users in determining their own preferences. Using a dataset of privacy preferences of smart assistant users, we test the accuracy of our predictions.

Marc Serramia, Natalia Criado, Michael Luck

Customising AI technologies to each user's preferences is fundamental to them functioning well. Unfortunately, current methods require too much user involvement and fail to capture their true preferences. In fact, to avoid the nuisance of manually setting preferences, users usually accept the default settings even if these do not conform to their true preferences. Norms can be useful to regulate behaviour and ensure it adheres to user preferences but, while the literature has thoroughly studied norms, most proposals take a formal perspective. Indeed, while there has been some research on constructing norms to capture a user's privacy preferences, these methods rely on domain knowledge which, in the case of AI technologies, is difficult to obtain and maintain. We argue that a new perspective is required when constructing norms, which is to exploit the large amount of preference information readily available from whole systems of users. Inspired by recommender systems, we believe that collaborative filtering can offer a suitable approach to identifying a user's norm preferences without excessive user involvement.

Nathan Gavenski, Juarez Monteiro, Felipe Meneguzzi et al.

Some imitation learning methods combine behavioural cloning with self-supervision to infer actions from state pairs. However, most rely on a large number of expert trajectories to increase generalisation and human intervention to capture key aspects of the problem, such as domain constraints. In this paper, we propose Continuous Imitation Learning from Observation (CILO), a new method augmenting imitation learning with two important features: (i) exploration, allowing for more diverse state transitions, requiring less expert trajectories and resulting in fewer training iterations; and (ii) path signatures, allowing for automatic encoding of constraints, through the creation of non-parametric representations of agents and expert trajectories. We compared CILO with a baseline and two leading imitation learning methods in five environments. It had the best overall performance of all methods in all environments, outperforming the expert in two of them.

Nathan Gavenski, Felipe Meneguzzi, Michael Luck et al.

Imitation learning is an approach in which an agent learns how to execute a task by trying to mimic how one or more teachers perform it. This learning approach offers a compromise between the time it takes to learn a new task and the effort needed to collect teacher samples for the agent. It achieves this by balancing learning from the teacher, who has some information on how to perform the task, and deviating from their examples when necessary, such as states not present in the teacher samples. Consequently, the field of imitation learning has received much attention from researchers in recent years, resulting in many new methods and applications. However, with this increase in published work and past surveys focusing mainly on methodology, a lack of standardisation became more prominent in the field. This non-standardisation is evident in the use of environments, which appear in no more than two works, and evaluation processes, such as qualitative analysis, that have become rare in current literature. In this survey, we systematically review current imitation learning literature and present our findings by (i) classifying imitation learning techniques, environments and metrics by introducing novel taxonomies; (ii) reflecting on main problems from the literature; and (iii) presenting challenges and future directions for researchers.

Gabriele La Malfa, Emanuele La Malfa, Saar Cohen et al.

Self-play red team is an established approach to improving AI safety in which different instances of the same model play attacker and defender roles in a zero-sum game, i.e., where the attacker tries to jailbreak the defender; if self-play converges to a Nash equilibrium, the model is guaranteed to respond safely within the settings of the game. Although the parameter sharing enforced by the use of the same model for the two roles improves stability and performance, it introduces fundamental theoretical and architectural limitations. We show that the set of Nash equilibria that can be reached corresponds to a broad class of behaviours that includes trivial always refuse strategies and oracle-like defenders, thus limiting practical applicability. We then show that when attacker and defender share and update the same base model, the dynamics collapse to self-consistency, so that attacks do not enforce adversarial pressure on the defender. In response, we propose Anchored Bipolicy Self-Play, which trains distinct role-specific LoRA adapters on top of a frozen base model, thereby maintaining stable optimisation while preserving adversarial pressure through explicit role separation. In relation to standard self-play, we show up to 100x greater parameter efficiency than finetuning and consistent improvements in safety compared to self-play fine-tuned models. We evaluate on Qwen2.5-{3B, 7B,14B}-IT models across widely used safety benchmarks, showing improved robustness without loss of reasoning ability. Cross-play experiments further show that our attacker and defender models are superior to self-play in terms of adversarial defence and safety.

Emanuele La Malfa, Gabriele La Malfa, Samuele Marro et al. · oxford

Recent interest in Multi-Agent Systems of Large Language Models (MAS LLMs) has led to an increase in frameworks leveraging multiple LLMs to tackle complex tasks. However, much of this literature appropriates the terminology of MAS without engaging with its foundational principles. In this position paper, we highlight critical discrepancies between MAS theory and current MAS LLMs implementations, focusing on four key areas: the social aspect of agency, environment design, coordination and communication protocols, and measuring emergent behaviours. Our position is that many MAS LLMs lack multi-agent characteristics such as autonomy, social interaction, and structured environments, and often rely on oversimplified, LLM-centric architectures. The field may slow down and lose traction by revisiting problems the MAS literature has already addressed. Therefore, we systematically analyse this issue and outline associated research opportunities; we advocate for better integrating established MAS concepts and more precise terminology to avoid mischaracterisation and missed opportunities.

Nathan Gavenski, Michael Luck, Odinaldo Rodrigues

Imitation learning field requires expert data to train agents in a task. Most often, this learning approach suffers from the absence of available data, which results in techniques being tested on its dataset. Creating datasets is a cumbersome process requiring researchers to train expert agents from scratch, record their interactions and test each benchmark method with newly created data. Moreover, creating new datasets for each new technique results in a lack of consistency in the evaluation process since each dataset can drastically vary in state and action distribution. In response, this work aims to address these issues by creating Imitation Learning Datasets, a toolkit that allows for: (i) curated expert policies with multithreaded support for faster dataset creation; (ii) readily available datasets and techniques with precise measurements; and (iii) sharing implementations of common imitation learning techniques. Demonstration link: https://nathangavenski.github.io/#/il-datasets-video

Gabriele La Malfa, Jie M. Zhang, Michael Luck et al.

Fairness in multi-agent systems (MAS) focuses on equitable reward distribution among agents in scenarios involving sensitive attributes such as race, gender, or socioeconomic status. This paper introduces fairness in Proximal Policy Optimization (PPO) with a penalty term derived from a fairness definition such as demographic parity, counterfactual fairness, or conditional statistical parity. The proposed method, which we call Fair-PPO, balances reward maximisation with fairness by integrating two penalty components: a retrospective component that minimises disparities in past outcomes and a prospective component that ensures fairness in future decision-making. We evaluate our approach in two games: the Allelopathic Harvest, a cooperative and competitive MAS focused on resource collection, where some agents possess a sensitive attribute, and HospitalSim, a hospital simulation, in which agents coordinate the operations of hospital patients with different mobility and priority needs. Experiments show that Fair-PPO achieves fairer policies than PPO across the fairness metrics and, through the retrospective and prospective penalty components, reveals a wide spectrum of strategies to improve fairness; at the same time, its performance pairs with that of state-of-the-art fair reinforcement-learning algorithms. Fairness comes at the cost of reduced efficiency, but does not compromise equality among the overall population (Gini index). These findings underscore the potential of Fair-PPO to address fairness challenges in MAS.

Gabriele La Malfa, Jie M. Zhang, Michael Luck et al.

Fairness in Multi-Agent Systems (MAS) has been extensively studied, particularly in reward distribution among agents in scenarios such as goods allocation, resource division, lotteries, and bargaining systems. Fairness in MAS depends on various factors, including the system's governing rules, the behaviour of the agents, and their characteristics. Yet, fairness in human society often involves evaluating disparities between disadvantaged and privileged groups, guided by principles of Equality, Diversity, and Inclusion (EDI). Taking inspiration from the work on algorithmic fairness, which addresses bias in machine learning-based decision-making, we define protected attributes for MAS as characteristics that should not disadvantage an agent in terms of its expected rewards. We adapt fairness metrics from the algorithmic fairness literature -- namely, demographic parity, counterfactual fairness, and conditional statistical parity -- to the multi-agent setting, where self-interested agents interact within an environment. These metrics allow us to evaluate the fairness of MAS, with the ultimate aim of designing MAS that do not disadvantage agents based on protected attributes.

Munkhtulga Battogtokh, Yiwen Xing, Cosmin Davidescu et al.

In natural language processing (NLP), text classification tasks are increasingly fine-grained, as datasets are fragmented into a larger number of classes that are more difficult to differentiate from one another. As a consequence, the semantic structures of datasets have become more complex, and model decisions more difficult to explain. Existing tools, suited for coarse-grained classification, falter under these additional challenges. In response to this gap, we worked closely with NLP domain experts in an iterative design-and-evaluation process to characterize and tackle the growing requirements in their workflow of developing fine-grained text classification models. The result of this collaboration is the development of SemLa, a novel visual analytics system tailored for 1) dissecting complex semantic structures in a dataset when it is spatialized in model embedding space, and 2) visualizing fine-grained nuances in the meaning of text samples to faithfully explain model reasoning. This paper details the iterative design study and the resulting innovations featured in SemLa. The final design allows contrastive analysis at different levels by unearthing lexical and conceptual patterns including biases and artifacts in data. Expert feedback on our final design and case studies confirm that SemLa is a useful tool for supporting model validation and debugging as well as data annotation.

Dong Huang, Jie M. Zhang, Michael Luck et al.

The advancement of natural language processing (NLP) has been significantly boosted by the development of transformer-based large language models (LLMs). These models have revolutionized NLP tasks, particularly in code generation, aiding developers in creating software with enhanced efficiency. Despite their advancements, challenges in balancing code snippet generation with effective test case generation and execution persist. To address these issues, this paper introduces Multi-Agent Assistant Code Generation (AgentCoder), a novel solution comprising a multi-agent framework with specialized agents: the programmer agent, the test designer agent, and the test executor agent. During the coding procedure, the programmer agent will focus on the code generation and refinement based on the test executor agent's feedback. The test designer agent will generate test cases for the generated code, and the test executor agent will run the code with the test cases and write the feedback to the programmer. This collaborative system ensures robust code generation, surpassing the limitations of single-agent models and traditional methodologies. Our extensive experiments on 9 code generation models and 12 enhancement approaches showcase AgentCoder's superior performance over existing code generation models and prompt engineering techniques across various benchmarks. For example, AgentCoder (GPT-4) achieves 96.3\% and 91.8\% pass@1 in HumanEval and MBPP datasets with an overall token overhead of 56.9K and 66.3K, while state-of-the-art obtains only 90.2\% and 78.9\% pass@1 with an overall token overhead of 138.2K and 206.5K.