Conceptcompute confound
Computational Confound
In AI evaluation, a phenomenon in which the method under assessment achieves improved performance because it consumes more computational resources, such as tokens or interaction rounds, making it impossible to attribute the performance difference to the method itself.
Timeline (last 90 days)
Oct 16
Rigorous agent research requires matching computational budgets, or at least reporting call counts and latency, to avoid computational confounding.
Unverified50%