Inference-time ablation
In plain English
An inference-time ablation disables part of an already trained model to measure how the model depends on that part.
Technical definition
Inference-time ablation changes or removes a component during evaluation without retraining the model to adapt to that change.
Engineering details
Suppressing Engram changes downstream features and expert choices in a model trained to use the memory. The resulting loss therefore measures dependence under a training-inference mismatch, not the quality difference between independently trained models with and without Engram.
Why it matters
Different tasks and evaluation procedures can respond differently. The article reports worse token likelihood across evaluated domains while GSM8K accuracy remains within measured run-to-run variation. A single unchanged score cannot establish that the removed component is generally unnecessary.
How to read it in InferenceX
The CRUXEval experiment separately tests natural rerouting and fixed original expert choices under teacher forcing. It also evaluates removing Engram during prefill, decode, or both, keeping those interventions distinct from storage-tier changes that retain the memory.