A survey study tracks six years of TrustNLP workshop papers: a shift in reliable natural language processing research from interpretability to proactive control

A recent reference study, co-authored by researcher Raul Gupta and 12 other researchers, namely Abhinav Mohanty, Analia Ovalle, Anil Ramakrishna, Anupama Agarwal, Apurva Verma, Gaurav Dhamaala, Ninareh Mehrabi, Tharindu Kumarajeewa, Yada Brooksatitkul, Yang Trista Cao, Kai-Wei Chang, and Aram Galstyan, reviews the historical and academic trajectory of research on the reliability of language models, documenting the structural shift in the field over six consecutive editions of the TrustNLP workshop, held in conjunction with the main conferences of the Association for Computational Linguistics (ACL) since 2021.

Analytical data in the paper shows that the TrustNLP workshop recorded exponential growth from 8 accepted papers in its first volume to 41 papers in its latest editions.The researchers based their conclusions on the classification and aggregation of scientific insights derived from all 144 published papers in the workshop, revealing that academic interest gradually shifted from post-hoc interpretability of fixed systems to a focus on mechanistic understanding and proactive control of modern generative systems.

To analyze research trends in a systematic manner, all papers were classified based on six main trust dimensions derived from reference frameworks, including TrustLLM and DecodingTrust, and the study found a direct correlation between changes in these dimensions and the emergence of technical capabilities, as the launch of the first high-impact conversational models simultaneously activated all six trust dimensions, and with the emergence of subsequent model generations, research weight clearly shifted towards issues of data truthfulness and safety alignment.

The classification results show that the dimension of truthfulness was the fastest-growing, as TrustNLP workshop papers recorded no presence of this dimension in 2021 and 2022, before rising to comprise 37% of all published papers in the workshop by 2025 and 2026, while the dimension of fairness maintained its position as the most consistent and stable axis across the six years, and explainability took a U-shaped path, as research interest in it declined with the loss of traditional post-hoc methods, only to recover again by 2026, supported by research in mechanistic interpretability.

To verify the extent to which the workshop represents the general trend, a comparison was made between the subject distribution of TrustNLP papers and around 2,000 research papers published in the four main conferences: ACL, NAACL, EACL, and EMNLP during the same period, and the result confirmed that the subject distribution of the TrustNLP workshop closely matches the overall average of the academic field, and the study concludes by providing four structural insights and actionable directions for the research community to guide the development of safer, more interpretable, and controllable models.