Skip to content
← Research
Experiment2 min read

One neuron against hallucination

A recognizer can emit familiar phrases when the input is a pause. We treated this as an intervention problem. Can we change how often the decoder emits words on no-speech input while preserving its ability to transcribe speech?

In the forced-development experiment, an edit to unit 1619 at layer 27 reduced the fraction of silence clips that produced text from 99.8% to 1.6%. Measured WER on the accompanying speech development set moved by 0.000 percentage points. The measurement covers that set and decoding configuration.

A separate locked evaluation tested a combined intervention that also changed loop and head behavior. Its starting hallucination rate was about 98%. Attributing that entire locked result to a single unit would mix experiments. The demo identifies the intervention behind each readout.

Sources and experiments

Continue exploring

Hear the difference on your own audio.

Try an English recording without an account, inspect the transcript and vocal-expression annotations, and use the quickstart to bring the result into your application.