In case you missed it: ADI’s researchers recap ECAI 2025 

The European Conference on Artificial Intelligence (ECAI) is the EU’s flagship conference was held in Bologna, Italy, at the end of October 2025. ECAI also hosted the Prestigious Applications of Intelligent Systems (PAIS) conference, which focuses on real-world applications of AI.  

Analog Devices, Inc. (ADI) researchers Giorgio Manganini, Antonello Scaldaferri, and Giulia Vilone attended the event, with Giorgio and Giulia presenting their paper  

ADI’s PAIS contribution: Evaluating and FineTuning VisionLanguage Models for Industrial Manufacturing in LowData Regimes (link) 

Our study explored the application of Vision–Language Models (VLMs) to enhance the automation of microchip production processes. VLMs have demonstrated impressive zeroshot and fewshot learning capabilities. Their ability to both reason over images and text makes them a promising solution for interactive industrial systems where robots collaborate with human workers. We evaluated four opensource VLMs for at ADI’s fabrication plants, testing each model in its pretrained state. We then finetuned the models on a domainspecific dataset to determine whether performance could be improved. Our findings show that VLMs still struggle with spatial reasoning, especially in assessing the position of small objects, and cannot provide a viable solution for our problem.  

Here’s a recap of the keynotes and our favorite technical sessions. 

Keynote talks 

Growing concerns over AI’s robustness to adversarial or poisoning attacks motivated University of Oxford’s Marta Kwiatkowska’s talk. The talk, ‘Provably Robust Artificial Intelligence? A Formal Methods Perspective’,argued that formal verification methods from computer science can provide provable guarantees of robustness. Kwiatkowska surveyed emerging verification methodologies aiming to certify AI systems against adversarial manipulation. She wrapped the talk up by discussing the challenges of applying formal methods to today’s AI models and building reliability into their design. 

Marco Dorigo of Université Libre de Bruxelles reviewed the successes and limitations of fully selforganised robot swarms. These swarms rely on local interactions among agents and therefore exhibit scalability and fault tolerance. The lack of direct control, however, makes it difficult to manage swarm behaviour for practical tasks. Dorigo proposed a hybrid architecture that creates a selforganised hierarchy. Using heterogeneous swarms of ground robots and drones, he showed how local asymmetric communication leads to a dynamic control network. Experiments demonstrated swarms can autonomously split and merge subswarms, replace faulty robots, and adjust collective behaviour in real time. The approach preserves the benefits of selforganisation while enabling human operators to steer the swarm. (If you’re interested in robotic swarms, read more about ADI’s contributions to the EU’s OpenSwarm project.) 

Mohit Bansal from the University of North Carolina at Chapel Hill shared work on building planning agents that can collaborate effectively with humans. His talk, Trustworthy Planning Agents for Collaborative Reasoning and Multimodal Generation, began with strategies for teaching agents to be dependable partners through confidence calibration, speaker–listener reasoning, and balancing persuasive behaviors. He also introduced methods for learning reusable abstractions over actions and code, along with data generation techniques to strengthen weak skills. The second half focused on interpretable, controllable multimodal generation, including layout-controlled image creation, consistent video generation guided by language models, and interactive any-to-any multimodal systems.  

Northwestern University’s Edith Elkind’s keynote, Many Facets of Proportionality in Multiwinner Voting, explored multi-winner voting. where voters elect a set of winners (e.g., members of parliament or items for a reception). She summarised recent progress on defining what it means for an outcome to be proportional and explored how to extend proportionality to more complex voting environments. Her talk highlighted the theoretical challenges of balancing fairness and representation when selecting multiple winners and suggested directions for future research. 

The final keynote, delivered by Ben Gurion University of the Negev’s Roni Stern, offered an overview of Multi-Agent Path Finding (MAPF)’s evolution. What started out as  a theoretical combinatorial search problem is today a foundational technology for managing large-scale automated warehouses. Such warehouses house fleets of hundreds (if not thousands) of robots that must transport goods efficiently (computing and following the shortest, quickest path) and safely (avoiding collisions with other robots and with humans). Stern pointed out the gaps between the MAPF-like challenges encountered in warehouses and the classical MAPF definition, covering approaches to bridging these gaps, including lifelong planning, robustness to failures, and techniques to mitigate uncertainty. 

Our favorite sessions 

Planning and Search 2 

Covering new algorithms for classical planning, symbolic search, and domain repair, among others. The talk "Improving Resilient Planning Through Landmarks and Regressed State Formulas” focused on resilient planning enhancement, capable of handling unexpected events, by using landmarks and regressed state formulas. This approach allows the system to respond to failure by working backward to determine the states from which the goal could have been reached. 

Session 2 – Multiagent Systems 2 

This session explored new directions in multi-agent reinforcement learning (MARL). Researchers introduced methods for adding natural language capabilities to agents and factorizing value functions for asynchronous environments. One paper, ADAPT: Auction-Based Dynamic Prioritization for Multi-Agent Coordination, proposed auction-style mechanisms that compute dependency scores to improve coordination. Another study presented a defense strategy for detecting adversarial attacks in MARL with continuous action spaces.  

Session 3 – Machine Learning 10 

This session focused on techniques to reduce the resource consumption of large generative AI models while maintaining or improving accuracy. In “Forward Only Learning for Orthogonal Neural Networks of Any Depth”,the authors presented a new forward-only algorithm for training neural networks without the need for a backward pass at lower time and cost.   

Session 6 – Machine Learning 13 

This session covered both algorithmic progress and medical applications of reinforcement learning (RL). Talks explored theoretical topics like computing Blackwell-optimal policies and secure policy gradient updates. One paper, Improving Actor-Critic Training with Steerable Action-Value Approximation Errors, refined actor–critic algorithms for continuous control by reducing approximation errors through task-specific utility functions. These functions adapt to critic uncertainty, balancing overly cautious strategies that limit learning with overly bold ones that increase risk.  

Session 7 – Machine Learning 16 

This session focused on sample-efficient reinforcement learning (RL), interpretable policy distillation, and domain adaptation in offline RL. One paper, “So, Tell Me About Your Policy…”, introduced an algorithm that extracts interpretable policies from expert behavior using an advantage function to explain why certain actions outperform others. It leverages previously collected data and showed strong results in control tasks and financial trading.  

Session 8 – Machine Learning 20 

This session explored reinforcement learning (RL) alongside domain adaptation and human feedback, introduced new object detection architectures such as domain-adaptive hybrid Mamba–Transformer and semi-supervised detection, and examined risk-averse RL strategies. A theoretical paper, Reasoning Optimization of Multi-Agent Systems via Abstract Domain Semantics, proposed a reasoning optimization approach using abstract domain semantics and a retrieval-and-repair mechanism. The authors demonstrated improved out-of-distribution reasoning on real-world blockchain governance tasks. Overall, the session showed how distributed problem-solving, RL, and computer vision can work together for practical applications.