Turns out AI hallucination fixes have a shelf life, and it is measured in layers, not time.
Researchers tested a training-free technique called visual intervention, which edits a multimodal AI model's internal representations to reduce hallucinated details in image-based answers. They applied the same edit at every decoder layer across six checkpoints from three different model families, holding the edit's strength constant, then measured how much each layer's edit changed the model's output. The effect dropped sharply past a specific depth the researchers call the Evidence Anchoring Boundary: interventions applied before that point were 13.78 to 26.10 times more effective than ones applied after it. The team also found that a common technique, subtracting a "counterfactual" signal, could just as easily erase a correct answer as fix a wrong one.
Most hallucination-reduction tools get applied the same way at every layer, on the assumption that more intervention is always better. This paper shows that assumption often fails. Once a model has locked in its answer, further edits mostly add noise or break things that were already working. The researchers' fix, a calibrated selective routing method that only intervenes when calibration data supports it for a given question type, lifted attribute-question accuracy by 1.56 to 1.81 percentage points across three checkpoints, while applying the same edit everywhere was neutral or actively harmful.
It is a useful reminder that in AI safety tooling, as in medicine, dosage and timing matter more than the drug itself.