← Back to Computation and Language cs.CL
Fixing AI hallucinations by reading hidden patterns across network layers
Tej Sanibh Ranade
May 18, 2026
Language models hallucinate because intermediate layers contain conflicting signals about truth—sometimes accurate information gets suppressed, sometimes multiple candidate answers compete across the network depth. TRACE addresses this by examining the full trajectory of information across layers at inference time, automatically selecting which layer to intervene on and what type of correction to apply (reversal, recovery, or replacement). Tested across 15 models and 3 factuality benchmarks with no training or external data, TRACE improves every configuration by an average of 12 points, reaching 47-point gains on some benchmarks.
Read the original paper →