Assistant Professor, University of Toronto

Research Scientist, Max Planck Institute 

Founder, EuroSafeAI

Title: Multi-Agent LLMs to Assist Causal Reasoning in Science

Abstract:  Causal reasoning sits at the heart of scientific discovery and remains one of the most demanding frontiers for AI. While large language models have demonstrated remarkable breadth across language tasks, their capacity for rigorous causal inference remains poorly understood, largely because the inferential pipeline from raw data to a credible causal claim demands tightly coupled decisions spanning structural assumption encoding, identifiability analysis, estimator selection, and sensitivity analysis. We argue that monolithic LLM prompting systematically fails to compose these steps with the required methodological discipline, and that decomposing causal inference into verifiable, agent-specialized subtasks is the right inductive bias for building AI systems that can reason causally in science. 

This talk presents a unified research program that develops and stress-tests this hypothesis across the full identification-estimation pipeline. We begin by establishing where current agents actually fail, introducing a rigorous benchmark grounded in causal tasks from published scientific literature that exposes systematic breakdowns in identification and estimator selection under realistic confounding regimes (CauSciBench). From this diagnostic foundation, we build toward autonomous causal inference with an end-to-end agent that takes observational data, metadata, and a causal query and performs covariate selection, backdoor and frontdoor identification, and effect estimation with uncertainty quantification (Causal AI Scientist). Two targeted extensions then address the hardest subtasks in this pipeline: for settings requiring exogenous variation, we cast instrumental variable discovery as a multi-agent deliberation problem over domain knowledge and testable exclusion-restriction proxies (IV-Co-Scientist); for verification, we introduce a symbolic layer that grounds LLM-produced causal claims in the do-calculus, catching identifiability violations and algebraic inconsistencies that chain-of-thought reasoning misses entirely (DoVerifier). Together, these systems define both the promise and the precise boundaries of what LLM-based causal agents can and cannot yet do in scientific discovery

Bio: 

Zhijing Jin (she/her) is an Assistant Professor in Computer Science at the University of Toronto, and also a Research Scientist at Max Planck Institute in Germany. She is a faculty member at the Vector Institute, a CIFAR AI Chair, an ELLIS advisor, and faculty affiliate at CHAI at UC Berkeley, Schwartz Reisman Institute in Toronto, and Future of Life Institute. She co-chairs the ACL Ethics Committee, and the ACL Year-Round Mentorship. Her research focuses on Causal Reasoning with LLMs, and AI Safety in Multi-Agent LLMs. She has received the ELLIS PhD Award, three Rising Star awards, two Best Paper awards at NeurIPS 2024 Workshops, two PhD Fellowships, and a postdoc fellowship. She has authored over 100 papers, many of which appear at top AI conferences (e.g., ACL, EMNLP, NAACL, NeurIPS, ICLR, AAAI), and her work have been featured in CHIP Magazine, WIRED, and MIT News. She co-organizes many workshops (e.g., several NLP for Positive Impact Workshops at ACL and EMNLP, and Causal Representation Learning Workshop at NeurIPS 2024), and leads the Tutorial on Causality for LLMs at NeurIPS 2024, and Tutorial on CausalNLP at EMNLP 2022. See more info at zhijing-jin.com

Head of Institute, Vienna University of Economics and Business

Professor, Information Systems and Business Engineering, Department for Information Systems and Operations Management, Vienna, Austria

Title: Towards neuro-symbolic management of causal knowledge

Abstract:  Causality plays a fundamental role in both human reasoning and complex system analysis. For example, as Cyber-Physical Systems (CPS) - such as smart grids or smart factories - become increasingly complex, understanding causal relationships between events happening in these systems is essential for tasks such as anomaly detection and fault diagnosis. Yet, major challenges remain in acquiring and managing causal knowledge. Firstly, in terms of acquisition, while eliciting causal knowledge from domain experts is the most reliable approach, automated methods are needed to increase the efficiency of this acquisition process. In that vein, starting from the context of smart grids, in this talk we will illustrate causal knowledge acquisition methods that augment domain-expert based acquisition with statistical and LLM-based methods. Secondly, when deploying a  variety of causal knowledge acquisition methods, it becomes challenging to integrate and combine their outputs, as these diverge both syntactically and semantically. To that end, we present a neuro-symbolic platform for supporting the management of causal knowledge derived through a variety of acquisition methods.

Bio: 

Prof. Dr. Marta Sabou is a professor for Information Systems and Business Engineering at  the Vienna University of Economics and Business (WU) and the Head of Institute for Data, Process and Knowledge Management (DPKM). She holds  a PhD in Artificial Intelligence from Vrije Universiteit Amsterdam, for which she received the IEEE Intelligent System’s Ten to Watch Award in 2006. During her career, she performed Artificial Intelligence (AI) research at the Open University UK, MODUL University Vienna, Siemens and the Vienna University of Technology.

Prof. Sabou leads the Semantic Systems research group, which performs foundational and applied research on topics ranging from knowledge engineering (knowledge graphs and their evaluation, data integration) to the development of novel intelligent systems that combine both symbolic and sub-symbolic AI techniques, i.e., neuro-symbolic systems. Increasingly, the group addresses topics in the area of Digital Humanism such as the auditing of AI systems and the involvement of human stakeholders in the design of intelligent systems.

Professor of Machine Learning, School of Computer Science at Carnegie Mellon University, Pittsburgh, USA 

Lead a Neurosymbolic AI Group, CMU

Title:

From Opaque Representations to Causal Concepts: A Unified Theory of Concept Extraction

Abstract:

Modern AI systems encode their knowledge in dense, opaque representations, which is a fundamental barrier to the interpretable, causally grounded reasoning that neuro-symbolic and agentic AI demand. Concept extraction methods like sparse autoencoders attempt to recover meaningful symbolic concepts from these representations, but when should we trust the concepts they find? In this talk, we present a unified theoretical framework that recasts concept extraction as a generative model identification problem, yielding a meta-theorem that reduces identifiability to a simple geometric condition on the intersection of two sets. This result clarifies the guarantees, and failure modes, of widely used unsupervised extraction methods. We then show how the framework extends naturally when we move beyond the purely unsupervised setting: lightweight causal side information, such as known intervention targets or environment labels, can dramatically strengthen identifiability, connecting concept extraction to recent progress in causal representation learning. The upshot is a principled path from raw neural representations to the kind of causally meaningful, human-interpretable concepts that trustworthy neuro-symbolic systems require.

Bio: 

Pradeep Ravikumar is a Professor in the Machine Learning Department, School of Computer Science at Carnegie Mellon University. He is a Sloan Fellow, a Siebel Scholar, a recipient of the NSF CAREER Award, and a co-editor-chief of the Journal of Machine Learning Research. His recent research interests are in neuro-symbolic AI, combining statistical machine learning, and symbolic and causal learning.