TRACE: Trajectory Correction from Cross-layer Evidence for Hallucination Reduction
The paper introduces TRACE, a new algorithm designed to reduce hallucinations in language models by utilizing cross-layer evidence. This approach corrects hallucinations during inference without requiring training or external data. Evaluation across multiple models and benchmarks shows significant improvements in factual accuracy.
- ▪TRACE is a deterministic, training-free algorithm that corrects hallucinations at inference time.
- ▪The method derives corrective layers and operators from each input's cross-layer candidate trajectory.
- ▪Evaluation of TRACE across 15 models and 3 factuality benchmarks yielded mean gains of +12.26 MC1 points and +8.65 MC2-style points.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.18163 (cs) [Submitted on 18 May 2026] Title:TRACE: Trajectory Correction from Cross-layer Evidence for Hallucination Reduction Authors:Tej Sanibh Ranade View a PDF of the paper titled TRACE: Trajectory Correction from Cross-layer Evidence for Hallucination Reduction, by Tej Sanibh Ranade View PDF HTML (experimental) Abstract:Hallucination correction is not a one-direction problem. We show that intermediate layers are neither uniformly more 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.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
