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.

In automated fact-checking (AFC), checkworthiness detection identifies claims requiring verification based on domain-specific criteria.On Wikipedia, this task instantiates as Citation Needed Detection (CND), which flags claims lacking supporting citations.However, existing research has largely overlooked lowerresource languages, and recent AFC pipelines rely on large language models (LLMs), which are inaccessible to low-resource organizations.We introduce MCN, a multilingual CND corpus spanning 18 languages across three resource levels, on which we conduct an extensive study of small decoder-based language models (SLMs).Our experiments show that SLMs fine-tuned with an encoder-style objective substantially outperform prompted LLMs across languages.We further present one of the first studies on cross-lingual CND, demonstrating that SLMs fine-tuned solely on English claims surpass LLMs, even with little to no target-language adaptation.Our findings have important implications for lower-resource Wikipedia communities and suggest that compact, task-specific models are preferable to LLMs for CND.We release all data and code at https://github.com/gerritq/mcnDataset Task Domain Languages (resource-level) Size NLP4IF 2021 (Shaar et al., 2021) CD/CWD Covid-19/Politics 3 (2 high, 1 medium) 1.3K-4K CitationNeeded (Redi et al., 2019) CWD Wikipedia 3 (3 high) 20K Kazemi et al. (2021) CD Covid-19, Politics 5 (2 high, 2 medium, 1 low) 5K Dutta et al. (2023) CD Politics 3 (2 high, 1 medium) 600-1.4KHalitaj and Zubiaga (2024) CWD Wikipedia 3 (2 high, 1 medium) 31K-1.1mCheckThat 2024 (Hasanain et al., 2024) CD/CWD Politics, Covid-19 3 (3 high) 1K-23K Baigutanova et al. (2026) CWD Wikipedia 5 (5 high) 20K MCN (ours) CWD Wikipedia 18 (8 high, 5 medium, 5 low) 2K-1.1m

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.

When bias mitigation methods are applied to make fairer machine learning models in fairness-related classification settings, there is an assumption that the disadvantaged group should be better off than if no mitigation method was applied. However, this is a potentially dangerous assumption because a “fair” model outcome does not automatically imply a positive impact for a disadvantaged individual—they could still be negatively impacted. Modeling and accounting for those impacts is key to ensure that mitigated models are not unintentionally harming individuals; we investigate if mitigated models can still negatively impact disadvantaged individuals and what conditions affect those impacts in a loan repayment example. Our results show that most mitigated models negatively impact disadvantaged group members in comparison to the unmitigated models. The domain-dependent impacts of model outcomes should help drive future bias mitigation method development.

Typically in abstract argumentation, one starts with arguments and a defeat relation, and applies some semantics in order to determine the acceptability status of the arguments. We consider the converse case where we have knowledge of the acceptability status of arguments and want to identify a defeat relation that is consistent with the known acceptability data – the σ-consistency problem. Focusing on complete semantics as underpinning the majority of the major semantic types, we show that the complexity of determining a defeat relation that is consistent with some set of acceptability data is highly dependent on how the data is labelled. The extension-based 2-valued σ-consistency problem for complete semantics is revealed as NP-complete, whereas the labelling-based 3-valued σ-consistency problem is solvable within polynomial time. We then present an informal discussion on application to grounded, stable, and preferred semantics.

Dialogical generalisations of formal logic-based argumentation are typically restricted to a limited set of locutions e.g., assert, why, claim or prefer. However, the use of enthymemes (i.e., arguments with incomplete logical structure) warrant extending this set of locutions. This paper formalises the use of additional novel locutions that account for the use of enthymemes and are typical of real world dialogues. We thus close the gap between formal logic-based models of dialogue and the kinds of dialogue studied by the informal logic community, which focus on more human-oriented models of dialogue. This is important if formal models of dialogues are to provide normative support for human-human debate, as well as for enabling computational and human agents to jointly reason via dialogue.

We consider a one-to-many persuasion setting, where a persuader presents arguments to a multi-party audience, aiming to convince them of some particular goal argument. The individual audience members each have differing personal knowledge, which they use, together with the arguments presented by the persuader, to determine whether they are convinced of the goal. The persuader must, therefore, carefully consider its strategy, i.e., which arguments to assert, in order to maximise the number of convinced audience members. Here, we use evolutionary search to find (near-)optimal strategies for the persuader. We implement our approach using search-based model engineering, which provides a natural and efficient encoding for such problems. We investigate the performance of our approach on a range of settings, considering different structures and sizes of argumentation frameworks (representing the underlying knowledge available to the persuader and audience members), and varying the size of audience and of the audience members' personal knowledge bases. We show that we can find effective strategies for problems with more than 200 arguments and more than 100 audience members. Further, we show that the approach supports multiple persuader objectives, finding persuader strategies that aim to minimise arguments to assert while still maximising the number of convinced audience members.

This paper describes a study which assesses a software prototype designed as part of the POSTURE project. POSTURE (Pressure Offloading Support Technologies for Ulcer REduction) comprises the design and development of a prototype system to help patients suffering from Diabetic Foot Ulcer (DFU) self-manage their treatment. The system is designed to take input from commercial wellness sensors and other sensors specially crafted for tracking aspects of foot ulcer progress. The sensor data is fed into a data-backed decision support tool which interacts with users via an intelligent agent, using a chatbot style interface, underpinned by computational argumentation-based dialogue. A recent Patient and Public Involvement (PPI) study was conducted in which DFU patients were presented with the prototype system design and offered opinions about the utility of the system from their unique perspectives. The preliminary results of this study are presented here.

Argument-based persuasion dialogues provide an effective mechanism for agents to communicate their beliefs, and their reasons for those beliefs, in order to convince another agent of some topic argument. In such dialogues, the persuader has strategic considerations, and must decide which of its known arguments should be asserted, and the order in which they should be asserted. Recent works consider mechanisms for determining an optimal strategy for persuading the responder. However, computing such strategies is expensive, swiftly becoming impractical as the number of arguments increases. In response, we present a strategy that uses heuristic information of the domain arguments and can be computed with high numbers of arguments. Our results show that not only is the heuristic strategy fast to compute, it also performs significantly better than a random strategy.

In most proposals for logic-based models of argumentation dialogues between agents, the arguments exchanged are logical arguments of the form 〈Φ,α〉 where Φ is a set of formulae (called the support) and α is a formula (called the claim) such that Φ is consistent and Φ entails α. However, arguments presented by real-world agents do not normally fit the mould of being logical arguments. They are normally enthymemes, and so they only explicitly represent some of the premises for entailing their claim and/or they do not explicitly state their claim. For example, for a claim that ‘you need an umbrella today’, a husband may give his wife the premise ‘the weather report predicts rain’. Clearly, the premise does not entail the claim, but it is easy for the wife to identify the assumed knowledge used by the husband in order to reconstruct the intended argument correctly (i.e. ‘if the weather report predicts rain, then you need an umbrella’). Whilst humans are constantly handling examples like this, proposals for logic-based formalizations of the process remain underdeveloped. In this article, we present a logic-based framework for handling enthymemes, some design features of which are influenced by aspects of relevance theory (proposed by Sperber and Wilson). In particular, we use the ideas of maximizing cognitive effect and minimizing cognitive effort in order to enable a proponent of an intended logical argument to construct an enthymeme appropriate for the intended recipient, and for the intended recipient to deconstruct the intended logical argument from the enthymeme. We relate our framework back to Sperber andWilson's relevance theory via some formal properties.