TRACE is proposed, a single-pass, decoded-answer-preserving confidence estimator that treats decoding-time uncertainty as a trajectory through three steps: recording token-level surprisal and predictive entropy during decoding, applying local risk operators to preserve uncertainty spikes, and converting localized trace...
Yue-Bin Xu, Xue-Mei Peng, Jun-Lan Chen et al.· 0 citations
This work proposes CompEvo, a trainable multi-agent evolution framework that integrates strategy execution, fitness-based differentiable selection, and competition-induced strategy evolution that enables heterogeneous agents to explore diverse news evidence, converts forecasting feedback into differentiable influence w...
Yuxuan Zhang, Yang-Yang Feng, Yong Guan et al.· 0 citations
Privileged Adaptation from Student Trajectories (PAST), which treats each completed student trajectory as additional privileged information for the OPSD teacher while leaving the student's distillation prefixes unchanged, and improves the Avg@12 macro average over Vanilla OPSD by 5.6 percentage points.