Researchers introduced a method to detect hallucinations in LLMs by analyzing attention graph topologies and information flow patterns. Their approach provides consistent improvements over existing methods and identifies impaired context sharing as a key characteristic of hallucinations. This affects developers building and operating LLMs, as it may inform improvements to LLM architectures and training procedures.

Researchers designed and tested privacy-preserving methods for algorithmic shortlisting in participatory budgeting, using large language models (LLMs) to predict project funding. The findings suggest that user preferences are stable enough for algorithmic shortlisting to effectively approximate an initial selection of projects.

Researchers propose Re2A, a framework for situated conversational recommendation. It uses rubric-based preference reasoning to guide response generation and optimize for user preference satisfaction and situation consistency.

Researchers found hidden instability in Vision Language Models (VLMs) by measuring internal representation drift, spectral sensitivity, and structural smoothness. This instability can lead to models producing unchanged outputs while undergoing significant internal changes. The study also found that larger VLMs are not necessarily more robust and that perturbations affect tasks differently, highlighting the need for more robust evaluation frameworks for VLMs.