AI inference glossary
Model architecture

Teacher forcing

In plain English

Teacher forcing supplies the reference token history so different model variants are scored on the same continuation.

Technical definition

Teacher forcing conditions a model on reference tokens rather than its own sampled outputs, allowing token-level losses to be compared along a fixed sequence.

Engineering details

The article uses a teacher-forced CRUXEval experiment to hold tokens constant while changing Engram and expert routing. This separates changes in reference-answer likelihood from differences caused by models generating different continuations.

Why it matters

Teacher-forced loss and free-generation task accuracy are different measurements. Fixed token histories help isolate an intervention, but they do not measure the entire behavior of a model choosing and extending its own answers.

How to read it in InferenceX

Removing Engram raises answer loss from 0.2848 to 0.3093 bits/token; forcing the ablated model to retain the original expert choices raises it further to 0.3375. These results support partial compensation through rerouting in that experiment, not a universal division between memory and reasoning.