Skip to content

PATTERN Cited by 1 source

Deterministic evaluation cache

Intent

Make LLM evaluation reproducible, efficient, and comparable across runs by caching at the per-sample level, eliminating noise from judge drift and reference regeneration.

Mechanism

Cache on two independent axes:

  1. Reference axis — key: (sample_identifier, reference_generation_config) → cached reference output
  2. Judge axis — key: (sample, model_output, judge_config, metric) → cached score

When inputs are identical to a prior evaluation, cached results are returned directly. Only genuinely different inputs trigger new judge/reference computation.

Properties

  • Deterministic: identical inputs always produce identical evaluation results
  • Durable partial progress: a failed job at example N resumes from cache
  • Incremental: new candidates/metrics reuse existing cached outputs
  • Efficient: >50% of model outputs across candidates are identical strings (Airbnb measurement)

Why this beats probabilistic approaches

Approach Problem
Majority voting Converges to judge's central tendency, not accuracy
Bayesian noise modeling Same infra cost as cache, without reproducibility
Repeated sampling Resource-prohibitive at scale (examples × systems × judges)
Per-sample cache Eliminates noise at source; simpler, cheaper

Prerequisite for

Seen in

Last updated · 571 distilled / 1,747 read