A new hallucination detector called HalluTracer catches AI mistakes by reading every layer of a model's thinking, not just one.
Researchers built HalluTracer, a white-box framework that pulls truthfulness signals from every layer of a language model's forward pass before it generates a single answer token. Earlier detectors picked one layer or isolated component to inspect, discarding evidence scattered across the rest of the network. The team found that per-layer signals are only weakly correlated with each other, so simply averaging them across depth cancels out layer-specific noise while keeping nearly all the useful signal intact. Tested across six open-source language models and five hallucination benchmarks, HalluTracer beat matched white-box baselines by one to fourteen points.
Hallucination is still one of the more embarrassing failure modes in language models, especially in fields like law, medicine, and finance where a confidently wrong answer can cause real damage. This work reframes the problem: instead of hunting for the single best layer to monitor, which has occupied much of the prior detection research, it argues the fix is aggregating evidence from all of them at once.
Better benchmark scores are not the same as fewer wrong answers in the wild, and reading a model's internal layers only works if you can access them - which rules out proprietary systems offered solely through an API.