0
After 10 Heads in a Row, Small LLMs Bet Up to 82% on an 11th
https://towardsdatascience.com/after-10-heads-in-a-row-small-llms-bet-up-to-82-on-an-11th/(towardsdatascience.com)An experiment on eight open-source language models reveals they do not typically exhibit the gambler's fallacy. Instead of expecting a streak of coin flips to break, the models show a 'streak bias,' where their confidence that a streak will continue increases with its length. This behavior was observed by analyzing the next-token probabilities for 'heads' or 'tails' after a given sequence. The streak bias was only slightly reduced even when the models were explicitly told that the coin was fair and the flips were independent.
0 points•by chrisf•18 hours ago
Comments (0)
No comments yet. Be the first to comment!
Have an account? Log in to join the discussion.