hard20 pts · first answer countsllm-evaluationbenchmarks
Passing the needle test
A model scores near 100% on a "needle in a haystack" test (finding one inserted sentence in a long document). Yet it does poorly when asked to compare figures spread across a 100-page report. Why?
Discussion 0
Join free to post in the discussion.
Related questions
- What exact match measuresPrompt Engineer
- Accuracy from four countsPrompt Engineer
- What a false positive isPrompt Engineer
- 99% accuracy that is uselessPrompt Engineer
- Code checks or an LLM judgePrompt Engineer
- Tracking broken JSONPrompt Engineer
- Grading with or without a referencePrompt Engineer
- Offline and online evaluationPrompt Engineer