Duration: 0:00

5,735 views

270

Dasha Metropolitansky, Research Data Scientist, Microsoft Research Special Projects, introduces VeriTrail, a new method for closed-domain hallucination detection in multi-step AI workflows. Unlike prior methods, VeriTrail provides traceability: it identifies where hallucinated content was likely introduced, and it establishes the provenance of faithful content by tracing a path to the source text. VeriTrail also outperforms baseline methods in hallucination detection. The combination of traceability and effective hallucination detection makes VeriTrail a powerful tool for auditing the integrity of content generated by language models.

Microsoft Research Special Projects VeriTrail paper: https://www.microsoft.com/en-us/research/publication/veritrail-closed-domain-hallucination-detection-with-traceability/
VeriTrail blog post: https://www.microsoft.com/en-us/research/blog/veritrail-detecting-hallucination-and-tracing-provenance-in-multi-step-ai-workflows/
Claimify video: https://www.microsoft.com/en-us/research/video/claimify-extracting-high-quality-claims-from-language-model-outputs/
Claimify paper: https://www.microsoft.com/en-us/research/publication/towards-effective-extraction-and-evaluation-of-factual-claims/
Claimify blog post: https://www.microsoft.com/en-us/research/blog/claimify-extracting-high-quality-claims-from-language-model-outputs/