Trustworthy AI II
Sat 17 Oct · 14:00–15:00 · Room B (First floor)
Paper session · Trustworthy and explainable AI · Chair: Anna Kowalska
Papers
Sign in to ask or upvote questions. Sign in
- Full paperPresented
Do counterfactual explanations survive model updates?
Felix Brandt (Alpine AI Institute) (presenting), Sara Romano (Northbridge University)
Abstract
Counterfactual explanations tell a person what to change to get a different decision. When the model is retrained, those promises can silently break. We measure how often counterfactuals remain valid across monthly retraining of a credit model and find that 38% become invalid within three updates. We propose a robustness regulariser that keeps 91% valid at a cost of 0.4 points of accuracy.
counterfactual explanations · model updates · robustness · credit scoring
Questions from the audience 0
No questions yet.
- Full paperPresented
Measuring hallucination in retrieval-augmented assistants for public administration
Lina Weber (Lakeside Technical University) (presenting), Arjun Mehta (Meridian Research Centre), Clara Fischer (Thalassa University)
Abstract
Public bodies are piloting language-model assistants that answer questions from official documents. We build a benchmark of 1,200 questions from Greek and English administrative guidance and evaluate five retrieval-augmented systems. Unsupported statements appear in 7 to 19 percent of answers, and citations do not reliably indicate support. We release the benchmark and an annotation guide.
large language models · hallucination · retrieval · public sector
Questions from the audience 0
No questions yet.