LAMB: layer-adaptive moment updates for large-batch training

What changed

Combined adaptive moment updates with layerwise trust-ratio scaling, adjusting LARS-style ratio/decay handling to enable large-batch BERT pretraining.

Scientific scope

Combined adaptive moment updates with layerwise trust-ratio scaling, adjusting LARS-style ratio/decay handling to enable large-batch BERT pretraining. Inherits Adam and LARS. Original six-author preprint differs from later expanded theory paper; general nonconvex theorem credit is not imported from the later version. The76minute result uses1024TPUs versus a16TPU baseline, not a hardware-matched64fold algorithmic speedup.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Inherits Adam and LARS. Original six-author preprint differs from later expanded theory paper; general nonconvex theorem credit is not imported from the later version. The76minute result uses1024TPUs versus a16TPU baseline, not a hardware-matched64fold algorithmic speedup.

Contributors

• Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.

• Reducing BERT Pre-Training Time from 3 Days to 76 Minutes

Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction