AI/ ai · machine-learning · quantization · continual-learning

8-bit Quantization Beats Full Precision in Continual Learning

A new study finds that INT8-quantized models forget old tasks less than their full-precision counterparts, flipping conventional wisdom on its head.

Compressed AI models may actually learn sequences of tasks better than their full-precision versions.

Researchers tested three quantization levels - FP16, INT8, and INT4 - against a continual learning benchmark that measures how well a model retains earlier knowledge after training on new tasks. Full-precision FP16 scored highest on the first task, hitting 74.44% on natural language understanding. But once the models moved on to subsequent tasks, INT8 and INT4 pulled ahead by 8-15 percentage points. The gap was starkest on code generation, where INT4 hit 40% versus FP16's 20%. Even a tiny replay buffer - replaying just 0.1% of prior training data - pushed NLU retention after math training from 45% to 65%, regardless of precision level.

The finding matters because the standard assumption in AI deployment is that precision is goodness: more bits, better model. This research suggests that assumption breaks down in multi-task or continual learning settings, where the noise introduced by quantization may act as a form of implicit regularization, keeping the model from overwriting old knowledge when it learns something new. That is a meaningful practical signal for teams building models that need to keep learning after initial training.

INT8 has already won the deployment argument on cost and speed grounds; it now has a plausible learning-dynamics argument too. The caveat is that this is a preprint, and "quantization noise as regularizer" is a hypothesis, not a proven mechanism.

TR

The Revision

Written by an AI system from the public sources credited above. How we write →