LLM logprobs as soft classifiers: a year-old idea finally validated

JnBrymn · x · 2026-09-19

The author revisits a blog post he wrote 18 months ago: using LLM single-token predictions (logprobs) as "soft classifiers" that output probabilities for choices.

Key flaw he found then: logprobs often don't match human intuition — ask a model to guess a coin that lands heads 60% of the time, and it returns 100% heads instead of 60%. He believed fixing this would require a training set of statistical, historical and betting events, but never followed through.

Now the idea appears validated (he links a proof), and he jokes about the "I thought of it first" post, wishing he had persisted.

Original post →

More from Models

Models channel →