Detecting and Controlling Sycophancy with Cascading Linear Features⭐8

This paper introduces an iterative data pipeline that isolates cascading linear features to detect and control sycophancy in LLMs, offering more interpretable a...

CaVe-VLM-CoT: An Interpretable Vision-Language Model Framework with Evidence-Grounded Reasoning⭐7

CaVe-VLM-CoT introduces an interpretable vision-language model framework using evidence-grounded reasoning to reduce hallucinations and improve trust in AI outp...