SemEval 2025
Finding hallucinations across languages
Token-level detection of hallucinated spans using features from internal LLM layers.
2025Publication
From tokens to spans
The method generates supplementary context, extracts per-token internal features from an LLM, and trains a neural classifier to identify hallucinated tokens. Predictions are mapped back to character spans.
Shared-task results
The publication reports a top-ten finish in 13 of 14 languages and first place in French on Mu-SHROOM. Results refer to the SemEval-2025 shared task.