the 28th European Conference on Artificial Intelligence (ECAI-2025) It is currently taking place in Bologna, Italy, from 25 to 30 October 2025. During the opening ceremony, the winners of the prestigious ECAI-2025 and Applications of Intelligent Systems (PAIS-2025) awards were announced. And the winners are…
ECAI-2025 Outstanding Papers
FAIRGAME: A framework for identifying bias of artificial intelligence agents using game theory
Alessio Buscemi, Daniele Proverbio, Alessandro Di Stefano, Anh Hanh, German Castignani, Pietro Liu
a summary: Allowing AI agents to interact in multi-agent applications adds a layer of complexity to the interpretability and prediction of AI outcomes, with profound implications for its trustworthy adoption in research and society. Game theory offers powerful models to capture and explain strategic interaction between agents, but requires the support of replicable, standardized, and easy-to-use IT frameworks to enable comparison and interpretation of results. To this end, we present FAIRGAME, a framework for identifying the bias of AI agents using game theory. We describe its implementation and use, and use it to detect biased outcomes in popular games among AI agents, depending on the large language model (LLM) used and the language used, as well as on the agents’ personal traits or strategic knowledge. Overall, FAIRGAME allows users to reliably and easily simulate desired games and scenarios and compare results across simulation campaigns and game theoretical predictions, enabling systematic discovery of biases, prediction of behavior emerging from strategic interactions, and enabling further research into strategic decision making using LLM agents.
Read the full paper here.
Conditional dominance analysis for classical planning
Anna Wilhelm, Alvaro Torralba
a summary: Dominance analysis methods compare pairs of situations in a planning task to demonstrate that one is at least as close to the goal as the other. Existing methods calculate truth-dominance relationships, which determine which truths are no less good than other truths in any given situation. However, this is only possible when one truth is at least as good as the other in every possible context. We introduce a new idea of conditional dominance, which can specify that one fact dominates another under certain conditions. We extend previous methods of calculating dominance by taking into account a range of “contexts” in order to find maximal dominance relationships. We propose several strategies to automatically find relevant contexts and show that even with a single condition, one can achieve significant pruning in certain domains.
Read the full paper here.
Analysis of temporal inference in description logic using grammar rules
Camille Borgo, Anton Gnatenko, Michael Tommaso
a summary: We establish correspondence between (parts of) the TEL◯a time extension of the EL description logic with the LTL operator ◯yourand some specific types of formal grammars, in particular, conjunction grammars (context-free grammars equipped with the intersection operation). This connection means that TEL◯ It does not have the final periodicity property of the models, and also results in the inability to determine the answer to the query in the TEL◯to close a question that has remained open since the introduction of TEL◯. Moreover, it also allows identifying the possibility of specifying the answer to the query for some interesting new parts of TEL◯And reusing existing tools and algorithms for this purpose in connected grammars.
Read the full paper here.
PAIS-2025 Premium Paper
Improving parcel sorting through reinforcement learning for internal logistics
Loris Rovida, Marco Maccarini, Filippo Borra, Fabio Risso, Bellerina Spahio
a summary: Parcel sorting is a critical process in domestic logistics for proper processing and dispatch of parcels. In general, such a process is performed manually by operators along the plant, without any added value, and may lead to musculoskeletal injuries due to uncomfortable working conditions. Automation solutions are also present in the market and scientific literature. However, available solutions are typically implemented using simplified, pre-defined sorting rules/finite state machines capable of managing only a limited number of parcel types/sorting scenarios. To fully generalize and automate the sorting process in inbound logistics, we propose to use reinforcement learning (RL) to derive sorting policies in combination with machine vision for online parcel tracking, which is used as an RL case. In more detail, a proximity policy optimization (PPO) algorithm is used on the policy for RL, and Yolo is chosen as the machine vision algorithm for parcel recognition and tracking. Based on SAIET Engineering’s AMS sorting unit, a modular kinematic model (with parcel collision modeling) of the sorting system was created ( n by M AMS – any,Dual-motion-matrix operators are derived and used as the ,environment for PPO. Training of the sorting policy is performed offline by randomizing the parcel number, size, and input positions. The trained policy is then deployed to the sorting unit, which is equipped with cameras to implement machine vision and evaluate performance. inside distribution and outside distribution (any(With package types not taken into account in off-line training) the tests achieved the target performance of 96.5% and 94% sorting accuracy, respectively.
Read the full paper here.
You can view the conference proceedings here.
Tags: Ikay, ICAI2025, Quick read
Lucy Smith is Senior Editorial Director at AIhub.
