From Trajectories to Grounded Preferences: Process Preference Synthesis via Interaction Element Graphs for Web PRMs
SURFPRM is proposed, a graph-guided process preference synthesis framework for comparative Web PRMs that outperforms baseline-trained models on average on WEBPRMBENCH and rival leading proprietary LLMs, and elevates the GMCP proportion from 24.19% to 74.60%, producing the curated SURFPRM-DATA dataset.
Yang-Zhe Peng, Xiao-Yang Wang, Yi-Yang Zhao et al.
· 0 citations