AI RESEARCH
TRACE: Trajectory Correction from Cross-layer Evidence for Hallucination Reduction
arXiv CS.CL
•
ArXi:2605.18163v1 Announce Type: cross Hallucination correction is not a one-direction problem. We show that intermediate layers are neither uniformly truthful than final layers nor uniformly less trustworthy. Yet hallucination reduction is usually instantiated through one fixed intervention form: contrast one layer against another, steer along a truthfulness direction, or defer to external evidence. This framing is structurally incomplete.