From interpretability to control: insights from six years of the TrustNLP workshop
From Interpretability to Control: Insights from Six Years of the TrustNLP Workshop
AI summary
An analysis of 144 papers from six editions of the TrustNLP workshop shows that trustworthy NLP research has shifted from post-hoc interpretability of static models to mechanistic understanding and active control of generative systems. Truthfulness grew fastest, rising from absent in 2021-2022 to 37% in 2025-2026; fairness was the most stable, and interpretability rebounded in 2026 through mechanistic interpretability. The workshop's papers increased from 8 to 41.