10/21/2026
Want to know about the latest research? See what our computer scientists are presenting next week at :
In “Mining Legal Arguments in U.S. Corporate Case Law,” Luis Brena, William Jurayj, Andrew Blair-Stanek, Benjamin Van Durme, and collaborators from the University of Maryland Francis King Carey School of Law introduce an expert-annotated dataset of 42 U.S. federal tax opinions on corporate reorganizations.
Miriam Wanner, Benjamin Van Durme, and colleagues from Microsoft and the University of Strathclyde construct a benchmark of 6,733 queries with minimally altered LLM responses designed to omit or falsify key information in “All Claims Are Equal, but Some Claims Are More Equal Than Others: Importance-Sensitive Factuality Evaluation of LLM Generations.”
“Many-Tier Instruction Hierarchy in LLM Agents” by Jingyu Zhang, Tianjian Li, William Jurayj, Hongyuan Zhan, Benjamin Van Durme, and Daniel Khashabi proposes a paradigm for resolving instruction conflicts among instructions with arbitrarily many privilege levels.
In “Seeing Through the MiRAGE: Evaluating Multimodal Retrieval Augmented Generation,” Alexander Martin, William Walden, Reno Kriz, Dengjia Zhang, Kate Sanders, Eugene Yang, Chihsheng Jin, and Benjamin Van Durme introduce an evaluation framework for retrieval-augmented generation from multimodal sources.
Chunsheng Zuo, Liaoyaqi Wang, William Jurayj, William Fleshman, and Benjamin Van Durme study how parameter-side memory interacts with context-side memory stored in the KV cache in document-level question ❓ answering in “Rethinking LoRA Memory Through the Lens of KV Compression.”
“Using Grounded Theory for Agent Behavior Analysis at Scale” by Jie Gao, Ziang Xiao, and researchers from Purdue University, Stevens Institute of Technology, Cornell University, and UTEP - The University of Texas at El Paso proposes the first multi-agent pipeline that automates grounded theory on agent trajectories.
In “What Makes a Good Response? An Empirical Analysis of Quality in Qualitative Interviews,” Jonathan Ivey, Anjalie Field, and Ziang Xiao identify, implement, and evaluate 10 proposed measures of interview response quality to determine which are actually predictive of a response’s contribution to a study’s findings.
Krithika Ramesh, Anjalie Field, and colleagues at IIT Madras and the Indian Institute of Science uncover and investigate a privacy-hallucination tradeoff ⚖️ in differentially private language models in “The Privacy-Hallucination Tradeoff in Differentially Private Language Models.”
“Exploring Sparse Autoencoders in Text-Based Causal Confounding Adjustment” by Mian Zhong, Anjalie Field, and Katherine A. Keith proposes a novel causal adjustment pipeline that iteratively selects a minimal set of SAE features via conditional independence tests.
In “Accurate but Not Humble: Evaluating Epistemic Humility in LLM Agents Under Knowledge Conflict,” Kaiser Sun, Bernal Jimenez Gutierrez, Jingyu Zhang, Jie Gao, Mark Dredze, Daniel Khashabi, and Hongjun Liu propose evaluating agents on their willingness to recognize, act on, and communicate uncertainty during task ex*****on.
and Arda Uzunoglu, Tianjian Li, and Daniel Khashabi study whether aggregate scores faithfully reflect performance across benchmark subdomains in “The Flaw of Averages: Quantifying Uniformity of Performance on Benchmarks.”
