Evidence can conflict
Statistical tests, causal assumptions and expert knowledge may support incompatible conclusions.
Causal AI, open to challenge
CArLA
Transparent, explainable and contestable causal discovery.
European Research Council Proof of Concept Grantbuilding upon the Advanced Grant ADIX
The problem
Causal discovery should be open to scrutiny.
Causal discovery helps build the assumptions needed for causal analysis, uncover scientific knowledge and support better decisions when randomised controlled trials are impossible or too costly. Because those assumptions shape what follows, they should remain open to scrutiny.
Experts need a voice
A domain expert should be able to inspect, challenge and qualify a causal claim—not simply override a graph.
Revision should be traceable
CArLA keeps the origin, confidence and consequences of each judgement visible through revision.
The foundational workflow
Turn causal discovery into a debate about evidence.
First introduced in Argumentative Causal Discovery, the method represents causal evidence as arguments whose support and conflicts can be inspected.
Bring evidence together
Combine statistical evidence from data with causal knowledge supplied by experts. Neither source is assumed to be infallible.
Make disagreement explicit
Translate evidence into arguments and preserve the conflicts between incompatible causal claims.
Evaluate supported alternatives
Use argumentative reasoning to identify the causal structures supported by the available evidence—without hiding uncertainty.
What CArLA is building
A learning assistant for accountable causal analysis.
CArLA aims to move from a one-shot graph towards an interactive process in which evidence, explanations and revisions remain available for review.
Transparency
Expose the evidence and assumptions behind each supported causal structure.
Explanation
Support why, why-not and what-if questions about causal claims and alternatives.
Contestability
Let experts challenge inputs and make the consequences of revisions visible.
Scalability
Combine exact reasoning with efficient solvers and learned approximations.
Planned application domains: healthcare and finance provide high-stakes settings in which transparent, expert-led revision matters. The public demo currently uses the Asia research benchmark.
Research
Peer-reviewed publications, tools and dissemination.
The work moves from the foundational method to finer-grained semantics, scalable approximation, language-model knowledge, and methods for contestability and redress.
- 01
Argumentative Causal Discovery
Combines statistical evidence and expert knowledge through assumption-based argumentation, making causal conflicts explicit and returning supported alternative structures.
- 02
On Gradual Semantics for Assumption-Based Argumentation
Assigns dialectical strength to assumptions, enabling more nuanced comparisons than binary acceptance alone.
- 03
Heterogeneous Graph Neural Networks for Assumption-Based Argumentation
Uses heterogeneous graph neural networks to approximate credulous acceptance and reconstruct stable extensions for large ABA frameworks.
- 04
Leveraging Large Language Models for Causal Discovery: a Constraint-based, Argumentation-driven Approach
Uses language models as imperfect experts that propose causal constraints, then evaluates those constraints alongside statistical evidence through Causal ABA.
- 05
Optimal Correction Sets for Argumentative Causal Discovery
Recasts the release of inconsistent causal constraints as a cost-aware repair problem using optimal minimal correction sets.
People
Come argue with us.
CArLA brings together causal discovery, computational argumentation, explainable AI and practical experience in high-stakes modelling.
Prof. Francesca Toni
Principal Investigator and strategic lead
Argumentation, interactive explainability, contestability and project leadership.
Dr. Fabrizio Russo
Technical lead
Causal discovery, statistical learning, explainable AI, contestability and software engineering.
Dr. Anna Rapberger
Argumentation and explainability
Computational argumentation theory, scalability and explainable AI.
Student contributors and supervised research
Dhruv Himatsingka · Zihao Li · Yeva Hunanyan · Preesha Gehlot
Meet the people behind the work