Understanding how linguistic structure emerges in language models is central to interpreting what these systems learn from data and how much supervision they truly require. In particular, semantic role understanding ("who did what to whom") is a core component of meaning representation, yet it remains unclear whether it arises from pre-training alone or depends on task-specific fine-tuning. We study whether semantic role understanding emerges during language model pre-training or requires task-specific fine-tuning. We freeze decoder-only transformers and train linear probes to extract semantic roles, using performance to infer whether role information is already encoded in pre-training or learned during adaptation. Across model scales, we find that frozen representations contain substantial semantic role information, with performance improving but not fully matching fine-tuned models. This indicates partial but incomplete emergence from pre-training alone. We show that semantic role structure emerges from language modeling objectives, but its internal implementation shifts toward more distributed representations as model scale increases.

Chaotic dynamical systems pose a fundamental challenge for Reinforcement Learning (RL): exponential sensitivity to initial conditions induces high-variance bootstrap targets and poorly conditioned gradient updates. Chaotic dynamics arise across scientific and engineering domains, from fluid flows and climate systems to multi-agent systems, where reliable learning is highly desirable. Standard RL methods optimise expected returns through scalar value functions, implicitly averaging over diverging trajectories and entangling trajectory level instability with the learning objective. We show that under mild statistical stability assumptions, the return distribution evolves more regularly than individual trajectories when measured under the 1-Wasserstein metric, yielding a smoother distributional Bellman objective. By aligning optimisation with this measure level structure, distributional RL provides better conditioned learning. We offer a principled explanation for the advantages of distributional methods in chaotic systems and the geometries of RL objectives under chaos.

Large Language Models (LLMs) are increasingly capable of generating text that challenges human performance in domains requiring creativity, yet evaluating creativity in LLM-generated content remains a significant challenge. Here, we investigate whether current automatic evaluation methods can reliably capture human judgments of creativity. We collect human evaluations of human- and AI-generated short stories from the WritingPrompts dataset across 11 dimensions of creativity, and compare these judgments with automated objective metrics and LLM-as-a-Judge evaluations. Our experiments reveal substantial misalignment between automatic evaluations and human assessments. In particular, LLM-based judges exhibit a systematic preference for AI-generated stories, consistently favoring their stylistic characteristics over the unpredictability and other qualities of human-authored texts. Furthermore, correlation analyses show that widely used automatic metrics exhibit near-zero alignment with human judgments across both human- and AI-generated stories, suggesting that they fail to capture important dimensions of creativity. These findings highlight fundamental limitations in current approaches to the automatic evaluation of creative text and underscore the difficulty of reducing the multidimensional and subjective nature of creativity to computational metrics.

Cooperation among independent learning agents is desirable as it enables reaching collectively rewarding states. Recent work has shown that artificial agents can learn to act pro-socially without the need for predefined cooperative preferences or behavioural heuristics, provided that they can observe others' actions or policies and select them as partners accordingly. This paper relaxes this constraint, studying reinforcement learning (RL) agents operating with only minimal information about others' behaviour. We propose a novel `Observer Model', where agents gain insights from direct experience and limited, indirect observations. We show that direct experience alone cannot sustain cooperation, particularly in large societies. However, even minimal observations of third-party interactions, allowing as few as one observer per gameplay, lead to significant improvements, enabling the population to achieve and sustain robust cooperation across varying population sizes. Through numerical analysis, we show the co-evolution of strategy and interaction structure and disentangle how learning happens under various settings. Analysing the partner selection graph, we identify the reasons for cooperation to emerge, and we explore how different learning and exploration rates affect the outcome of social dilemmas played among RL agents.

With increasing awareness of the privacy risks posed by mobile phone location data, researchers need ways to use mobility data while offering stronger privacy guarantees to the individuals included in this data. A promising approach to this challenge is the creation of privacy-preserving mobility insights from decentralized location data using Local Differential Privacy (LDP). However, mobility data generated with LDP, based on the introduction of noise by individual mobile devices, is limited by the volume of noise required to achieve individual privacy. In this paper, we provide a fully reproducible model of the accuracy of mobility networks generated with LDP compared to mobility network data generated with more traditional privacy mechanisms: Central Differential Privacy (CDP) and K-anonymity. Using a simulated mobile phone mobility dataset informed by real-world travel patterns in the USA, we explore the trade-off between privacy and data utility provided by different parameters in a federated system with LDP. We also explore the impact of spatial and temporal aggregation on data accuracy, showing that long-standing considerations regarding the appropriate units of analysis for geographic data play a key role in determining the utility of federated mobility data with LDP. Our paper facilitates an in-depth understanding of the trade-offs between privacy and data utility entailed by the future adoption of a federated approach which uses LDP to generate insights from decentralized mobility data.

Ensuring the reliability of voice services in 5G networks requires effective detection of anomalies in IMS signaling. However, this task remains challenging due to the architectural complexity of IMS and the large volume of signaling data. In this paper, we propose SIP-Classifier, an unsupervised methodology that combines Transformer-based representation learning with clustering to identify anomalous SIP sequences. The approach encodes SIP messages through protocol-aware tokenization, learns latent representations via an autoregressive Transformer, and clusters them to distinguish valid from anomalous flows. We evaluate the method on real-world IMS data collected from operational 5G networks. It achieves 98% accuracy, 98% precision, 95% recall, and a 96% F1-score, significantly outperforming state-of-the-art approaches.

Solving the problem of cooperation is fundamentally important for the creation and maintenance of functional societies. Problems of cooperation are omnipresent within human society, with examples ranging from navigating busy road junctions to negotiating treaties. As the use of AI becomes more pervasive throughout society, the need for socially intelligent agents capable of navigating these complex cooperative dilemmas is becoming increasingly evident. Direct punishment is a ubiquitous social mechanism that has been shown to foster the emergence of cooperation in both humans and non-humans. In the natural world, direct punishment is often strongly coupled with partner selection and reputation and used in conjunction with third-party punishment. The interactions between these mechanisms could potentially enhance the emergence of cooperation within populations. However, no previous work has evaluated the learning dynamics and outcomes emerging from multi-agent reinforcement learning populations that combine these mechanisms. This paper addresses this gap. It presents a comprehensive analysis and evaluation of the behaviors and learning dynamics associated with direct punishment, third-party punishment, partner selection, and reputation. Finally, we discuss the implications of using these mechanisms on the design of cooperative AI systems.

Tractography is the process of inferring the trajectories of white-matter pathways in the brain from diffusion magnetic resonance imaging (dMRI). Local tractography methods, which construct streamlines by following local fiber orientation estimates stepwise through an image, are prone to error accumulation and high false positive rates, particularly on noisy or low-resolution data. In contrast, global methods, which attempt to optimize a collection of streamlines to maximize compatibility with underlying fiber orientation estimates, are computationally expensive. To address these challenges, we introduce GenTract, the first generative model for global tractography. We frame tractography as a generative task, learning a direct mapping from dMRI to complete, anatomically plausible streamlines. We compare both diffusion-based and flow matching paradigms and evaluate GenTract's performance against state-of-the-art baselines. Notably, GenTract achieves precision 2.1x higher than the next-best method, TractOracle. This advantage becomes even more pronounced in challenging low-resolution and noisy settings, where it outperforms the closest competitor by an order of magnitude. By producing tractograms with high precision on research-grade data while also maintaining reliability on imperfect, lower-resolution data, GenTract represents a promising solution for global tractography.

Large language models (LLMs), particularly when integrated into agentic systems, have demonstrated human- and even superhuman-level performance across multiple domains. Whether these systems can truly be considered creative, however, remains a matter of debate, as conclusions heavily depend on the definitions, evaluation methods, and specific use cases employed. In this paper, we analyse creativity along two complementary macro-level perspectives. The first is a functionalist perspective, focusing on the observable characteristics of creative outputs. The second is an ontological perspective, emphasising the underlying processes, as well as the social and personal dimensions involved in creativity. We focus on LLM agents and we argue that they exhibit functionalist creativity, albeit not at its most sophisticated levels, while they continue to lack key aspects of ontological creativity. Finally, we discuss whether it is desirable for agentic systems to attain both forms of creativity, evaluating potential benefits and risks, and proposing pathways toward artificial creativity that can enhance human society.

Social scientists have conceptualised different dimensions of human relationships, including knowledge, identity, and social support. At times of crisis, such as during a pandemic, social relationships can be disrupted, affecting the nature of interactions individuals have with others within their community and beyond. Understanding how relationships change, in both structure and content, can reveal how communities adapt and which connections are most resilient. Existing research has examined network structure or communicative content in isolation, but rarely both in combination and over time across multiple communities. In this study, we systematically investigate the temporal evolution of social relationships using readily available social media data. In particular, we make two contributions. First, we develop a computational methodology that leverages ego-networks to study changes to interactions between individuals. Second, we apply this methodology to over a million X/Twitter messages posted by more than two hundred teachers, journalists, and academics in UK from the beginning of 2019 to the end of 2021. In so doing, we longitudinally study and compare changes in social relationships before, during, and after COVID-19 lockdown periods. While results vary across communities, and data sparsity limits generalizability, our findings suggest that out-community ties exhibited more stability than in-community ties over time.

Several works have investigated the influence of graph topology on cooperation among artificial agents, while the majority of the literature has focused on modelling agents' adaptation through strategy imitation, which relies solely on the cumulative payoffs of others. This paper investigates scenarios in which each agent learns to play the two-player Iterated Prisoner's Dilemma (IPD) using deep reinforcement learning. Each agent is represented as a node in a graph, where its neighbours constitute the pool of opponents with whom it can interact. During each IPD episode, agents are provided with different types of information about their opponent, consisting of action history and opponent identity. Experimental results across different graph topologies show that the number of neighbours per node and the average path length are the main factors affecting the emergence of cooperation. We also show that, while partner selection fosters mutual cooperation by limiting the diversity of the opponent pool, providing agents with the identity of their opponent hinders the proliferation of cooperative strategies.

Large Language Models (LLMs) are increasingly being deployed as autonomous agents in real-world environments. As these deployments scale, multi-agent interactions become inevitable, making it essential to understand strategic behavior in such systems. A central open question is whether LLM agents, like reinforcement learning agents, can shape the learning dynamics and influence the behavior of others through interaction alone. In this paper, we present the first investigation of opponent shaping (OS) with LLM-based agents. Existing OS algorithms cannot be directly applied to LLMs, as they require higher-order derivatives, face scalability constraints, or depend on architectural components that are absent in transformers. To address this gap, we introduce ShapeLLM, an adaptation of model-free OS methods tailored for transformer-based agents. Using ShapeLLM, we examine whether LLM agents can influence co-players' learning dynamics across diverse game-theoretic environments. We demonstrate that LLM agents can successfully guide opponents toward exploitable equilibria in competitive games (Iterated Prisoner's Dilemma, Matching Pennies, and Chicken) and promote coordination and improve collective welfare in cooperative games (Iterated Stag Hunt and a cooperative version of the Prisoner's Dilemma). Our findings show that LLM agents can both shape and be shaped through interaction, establishing opponent shaping as a key dimension of multi-agent LLM research.

The Information Bottleneck (IB) principle offers a compelling theoretical framework to understand how neural networks (NNs) learn. However, its practical utility has been constrained by unresolved theoretical ambiguities and significant challenges in accurate estimation. In this paper, we present a Generalized Information Bottleneck (GIB) framework that reformulates the original IB principle through the lens of synergy, i.e., the information obtainable only through joint processing of features. We provide theoretical and empirical evidence demonstrating that synergistic functions achieve superior generalization compared to their non-synergistic counterparts. Building on these foundations we re-formulate the IB using a computable definition of synergy based on the average interaction information (II) of each feature with those remaining. We demonstrate that the original IB objective is upper bounded by our GIB in the case of perfect estimation, ensuring compatibility with existing IB theory while addressing its limitations. Our experimental results demonstrate that GIB consistently exhibits compression phases across a wide range of architectures (including those with ReLU activations where the standard IB fails), while yielding interpretable dynamics in both CNNs and Transformers and aligning more closely with our understanding of adversarial robustness.

Machine Learning (ML) is increasingly accessible to users with limited knowledge of its theoretical foundations. However, misapplying it can lead to negative consequences. This paper reports on a qualitative study designed to reveal challenges that novices encounter when learning about basic ML concepts and building their first models. Twenty participants were introduced to fundamental ML concepts for classification through an interactive tutorial involving an off-the-shelf GUI application, built their own ML model for a shape gesture dataset, and participated in a semi-structured interview. A thematic analysis revealed insights into these challenges, particularly around problem selection and multi-dimensionality, but also around what constitutes ML, algorithm selection, cross-validation, and interpreting visualizations. Despite these and other misconceptions, participants reflected on good model building practices, discussing that algorithm selection might require knowledge and context and that input features may introduce bias. We discuss the findings’ implications for the design of ML tools for novices.

Interactive Machine Learning (IML) enables users, including non-experts in ML, to iteratively train and improve ML models. However, limited research has been reported on how non-experts interact with these systems. Focusing on thematic analysis as a practical application, we report on a user study where 20 participants interacted with TACA, a functioning IML tool. Thematic analysis involves individual interpretation of ambiguous data, hence it is suited for and can benefit from the iterative customization of models supported by IML. Through a combination of interaction logs and semi-structured interviews, our findings revealed that, by using TACA, participants critically reflected on their analysis, gained new thematic insights, and adapted their interpretative stance. We also document misconceptions of ML concepts, positivist views, and personal blame for poor model performance. We then discuss how applications could be designed to improve the understanding of IML concepts and foster reflexive work practices.

Climate policy development faces significant challenges due to deep uncertainty, complex system dynamics, and competing stakeholder interests. Climate simulation methods, such as Earth System Models, have become valuable tools for policy exploration. However, their typical use is for evaluating potential polices, rather than directly synthesizing them. The problem can be inverted to optimize for policy pathways, but the traditional optimization approaches often struggle with non-linear dynamics, heterogeneous agents, and comprehensive uncertainty quantification. We propose a framework for augmenting climate simulations with Multi-Agent Reinforcement Learning (MARL) to address these limitations. We identify key challenges at the interface between climate simulations and the application of MARL in the context of policy synthesis, including reward definition, scalability with increasing agents and state spaces, uncertainty propagation across linked systems, and solution validation. Additionally, we discuss challenges in making MARL-derived solutions interpretable and useful for policy-makers. Our framework provides a foundation for more sophisticated climate policy exploration while acknowledging important limitations and areas for future research.

Reinforcement learning from human feedback (RLHF) is a widely used method for aligning large language models with human preferences. However, RLHF often suffers from reward model overoptimisation, in which models overfit to the reward function, resulting in non-generalisable policies that exploit the idiosyncrasies and peculiarities of the reward function. A common mitigation is iterated RLHF, in which reward models are repeatedly retrained with updated human feedback and policies are re-optimised. Despite its increasing adoption, the dynamics of overoptimisation in this setting remain poorly understood. In this work, we present the first comprehensive study of overoptimisation in iterated RLHF. We systematically analyse key design choices - how reward model training data is transferred across iterations, which reward function is used for optimisation, and how policies are initialised. Using the controlled AlpacaFarm benchmark, we observe that overoptimisation tends to decrease over successive iterations, as reward models increasingly approximate ground-truth preferences. However, performance gains diminish over time, and while reinitialising from the base policy is robust, it limits optimisation flexibility. Other initialisation strategies often fail to recover from early overoptimisation. These findings offer actionable insights for building more stable and generalisable RLHF pipelines.

The concept of rationality is central to the field of artificial intelligence (AI). Whether we are seeking to simulate human reasoning, or trying to achieve bounded optimality, our goal is generally to make artificial agents as rational as possible. Despite the centrality of the concept within AI, there is no unified definition of what constitutes a rational agent. This article provides a survey of rationality and irrationality in AI, and sets out the open questions in this area. We consider how the understanding of rationality in other fields has influenced its conception within AI, in particular work in economics, philosophy and psychology. Focusing on the behaviour of artificial agents, we examine irrational behaviours that can prove to be optimal in certain scenarios. Some methods have been developed to deal with irrational agents, both in terms of identification and interaction, however work in this area remains limited. Methods that have up to now been developed for other purposes, namely adversarial scenarios, may be adapted to suit interactions with artificial agents. We further discuss the interplay between human and artificial agents, and the role that rationality plays within this interaction; many questions remain in this area, relating to potentially irrational behaviour of both humans and artificial agents.

The coevolution of signalling is a complex problem within animal behaviour, and is also central to communication between artificial agents. The Sir Philip Sidney game was designed to model this dyadic interaction from an evolutionary biology perspective, and was formulated to demonstrate the emergence of honest signalling. We use Multi-Agent Reinforcement Learning (MARL) to show that in the majority of cases, the resulting behaviour adopted by agents is not that shown in the original derivation of the model. This paper demonstrates that MARL can be a powerful tool to study evolutionary dynamics and understand the underlying mechanisms of learning over generations; particularly advantageous is the interpretability of this type of approach, as well as the fact that it allows us to study emergent behaviour without the need to constrain the strategy space from the outset. Although it originally set out to exemplify honest signalling, we show that the game provides no incentive for such behaviour. In the majority of cases, the optimal outcome is one that does not require a signal for the resource to be given. This type of interaction is observed within animal behaviour and is sometimes referred to as proactive prosociality. High learning and low discount rates of the reinforcement learning model are shown to be optimal in order to achieve the outcome that maximises both agents' reward, and proximity to the given threshold leads to suboptimal learning.

Policy gradient methods usually rely on entropy regularization to prevent premature convergence. However, maximizing entropy indiscriminately pushes the policy towards a uniform distribution, often overriding the reward signal if not optimally tuned. We propose replacing the standard entropy term with a self-regulating complexity term, defined as the product of Shannon entropy and disequilibrium, where the latter quantifies the distance from the uniform distribution. Unlike pure entropy, which favors maximal disorder, this complexity measure is zero for both fully deterministic and perfectly uniform distributions, i.e., it is strictly positive for systems that exhibit a meaningful interplay between order and randomness. These properties ensure the policy maintains beneficial stochasticity while reducing regularization pressure when the policy is highly uncertain, allowing learning to focus on reward optimization. We introduce Complexity-Regularized Proximal Policy Optimization (CR-PPO), a modification of PPO that leverages this dynamic. We empirically demonstrate that CR-PPO is significantly more robust to hyperparameter selection than entropy-regularized PPO, achieving consistent performance across orders of magnitude of regularization coefficients and remaining harmless when regularization is unnecessary, thereby reducing the need for expensive hyperparameter tuning.