LLMs Silently Accumulate Approximation Errors Without External Verification, Argues Engineer

gerardsans · x · 2026-09-19

In a discussion thread, Gerard Sans argues that since LLMs are distribution-based systems, they pass through phase transitions conditioned on inputs—from in-distribution, to sparse sampling, to fully out-of-distribution regions of the input space. Without strong external verification, users silently absorb approximation errors, which is compounded by the enormous output space and input-conditioned error behavior. His conclusion: external verification is mandatory, not optional.

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

Original post →

More from Models

Models channel →