2 papers
cs.LG2026
Exploring Diverse Generation Paths via Inference-time Stiefel Activation Steering
Dongxuan Zhu, Ly Tran Ho Khanh, Andy Yat-Ming Cheung +2
Language models often default to a narrow set of high-probability outputs, leaving their generation paths homogeneous and prone to mode collapse. Sampling-based strategies inject r…
cs.LG2025
Test-time Diverse Reasoning by Riemannian Activation Steering
Ly Tran Ho Khanh, Dongxuan Zhu, Man-Chung Yue +1
Best-of- reasoning improves the accuracy of language models in solving complex tasks by sampling multiple candidate solutions and then selecting the best one based on some crite…