The Needle In a Haystack Test

Source: https://towardsdatascience.com/the-needle-in-a-haystack-test-a94974c1ad38 Published: 15 February 2024 · Author: Aparna Dhinakaran

Summary

An account of the needle-in-a-haystack evaluation — originated by Greg Kamradt, extended by Arize — which embeds “a specific, targeted piece of information (the needle) within a larger body of text (the haystack)” and measures whether the model can use it.

Method

An out-of-place statement is inserted into essay text at varying depths (0–100%) and context lengths (1K to the model’s limit). The model answers using only the provided context; accuracy is charted across both axes.

Results

ModelFinding
ChatGPT-4Decline below 64K tokens; sharp decline past 100K. Overall leader.
Claude 2.127% retrieval accuracy initially → 98% with prompt adjustments
Mixtral-8x7B-v0.1 / 7B InstructOutperformed expectations for their size

Arize’s follow-up cut Claude 2.1’s misses from 165 to 74 through revised prompting alone.

Reading It Critically

A 71-point swing on prompt phrasing means the benchmark partly measures Prompt Sensitivity rather than context capability — worth holding in mind before citing any single number from it.