Why can two same-seed training runs differ?
A seed controls some sources of randomness, but training also depends on hardware, software, parallelism and the order of numerical operations. Floating-point addition is not perfectly associative, so different reduction orders can produce slightly different values that later updates may amplify.
Answered in
Why Two Identical AI Training Runs Can Still DisagreeBitwise reproducibility requires matching the full training setup and operation order, not merely reusing the same random seed.
Read the full analysisOther questions this article answers
More tech questions
- What does bitwise deterministic training mean?
- Can rounded loss values prove two runs are identical?
- Why is correct OCR not enough to understand a document?
- What does grounding add to OCR output?
- How should teams validate extracted document fields?
- What is a decision model interface?
- Can a model's confidence score be treated as a reliable probability?
- Does a routing decision authorize the next action?
Every answer on Crashtech is written by the editor of the article it comes from — never auto-summarised. Browse all answers or the Tech beat.