Debate: without strong verification, LLM approximation errors go silently unnoticed

gerardsans · x · 2026-09-19

A technical argument over LLM sampling and verification: the poster claims that without strong verification, models absorb the last tokens across the input space, so many silent approximation errors go unnoticed — and because outputs are conditioned on input over a huge output space, external verification is mandatory. The replier counters that these are fundamentals of next-token sampling, urging a review before misunderstandings snowball into catastrophic outcomes in high-stakes scenarios.

Related event: Researchers Debate Silent Error Accumulation in LLMs Without Verification(3 posts)→

Original post →

More from Models

Models channel →