Thesis chapter argues LLM logprobs can power probabilistic programming again

xuanalogue · x · 2026-09-28

The author shares a thesis chapter on probabilistic programming with LLMs, using model logprobs directly as priors and likelihoods.

Key point: around the GPT-3.5 era, language models could serve as reasonable priors and likelihoods, but finetuning progressively ruined their calibration, making the approach less useful. The author argues newer models like Jev make logprobs useful again, reviving the combination of probabilistic programming and language models.

Related event: New LLMs May Make Probabilistic Programming Viable Again(2 posts)→

Original post →

More from Models

Models channel →