Random Seeds in LLM Sampling: Reproducibility Without a Guarantee

Subtitle: A seed can replay a draw. It cannot freeze the whole stack.

• Panel 1 - How one draw picks a token:
• PRNG emits a number from 0 to 1
• Probabilities: [0.60, 0.30, 0.10]
• Cumulative bands: 0-.60 | .60-.90 | .90-1
• Draw 0.72 lands on token two
• Same RNG stream can replay the pick

• Panel 2 - The reproducibility target:
• Same prompt bytes
• Same model and tokenizer versions
• Same temperature, top-p, and seed
• Same provider path and fingerprint
• Target: same inputs -> same output

• Panel 3 - Where the promise breaks:
• Greedy decoding may ignore the seed
• GPU kernels can run in a new order
• Batching can alter floating-point ties
• Hosted backends can change unseen
• A seed controls randomness, not the stack

• Panel 4 - Falsify determinism:
• Fix seed; change temperature once
• Fix seed; change model or provider
• Then freeze all visible settings
• Repeat the request several times
• Report match rate plus fingerprint