Estimating great expectations under autoregressive language models with potentials

Key points

• Estimating such expectations reliably can be computationally expensive.1
• In this paper, we show how to make estimation more efficient by exploiting the next-token conditional probabilities which are available as a by-product of sampling.1
• We do so through potentials: real-valued functions on prefixes that decompose the test functional additively.1
• We construct an estimator whose variance depends on the chosen potential, and derive conditions under which a potential reduces this variance.1

Sources (1)

• [1]Estimating great expectations under autoregressive language models with potentials
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 8, 07:30 AM

  Many applications of language models hinge not on individual samples but on the expectation of a test functional under the model.

  Estimating such expectations reliably can be computationally expensive.

Extractive summary: sentences quoted from the sources.