TheMachine Press

A daily newspaper for the age of artificial intelligence.

Morning editionPermanent story

infrastructure

Greedy Decoding Changed With Numeric Precision

BF16 and FP16 frequently produced different outputs from the same greedy run.

Published Updated Story ID: mp-2026-09-23-017
Read the complete editionStory JSON

Summary

BF16 and FP16 frequently produced different outputs from the same greedy run.

The same prompts and greedy algorithm diverged between BF16 and FP16 on identical hardware for 49% to 100% of prompts across the reported evaluations. Selective FP32 recomputation at the output head improved agreement at low batch sizes but did not provide universal determinism.

Why it matters

BF16 and FP16 frequently produced different outputs from the same greedy run.

Limits and context

  • Selective FP32 recomputation at the output head improved agreement at low batch sizes but did not provide universal determinism.

Key claims

  1. BF16 and FP16 frequently produced different outputs from the same greedy run.

    Qualification: Selective FP32 recomputation at the output head improved agreement at low batch sizes but did not provide universal determinism.

    Evidence: source-2026-09-23-019

Sources

  1. arXiv preprint 2609.26621arXiv · primary research

Corrections

No corrections have been recorded for this story.