TRACE: Training-Free Algorithm Reduces LLM Hallucinations via Cross-Layer Evidence

Researchers have introduced TRACE (Trajectory Correction from Cross-layer Evidence for Hallucination Reduction), a deterministic, training-free algorithm that corrects hallucinations in large language models at inference time. The method addresses the structural incompleteness of current hallucination reduction approaches, which typically rely on a single intervention form such as contrasting layers, steering along a truthfulness direction, or deferring to external evidence. TRACE derives both the corrective layer and the appropriate correction operation by analyzing cross-layer factual evidence, which does not evolve uniformly across model depth. The algorithm is described in a paper published on arXiv (2605.18163v1).