Sable: A Small-Field Sumcheck Argument for Bit-Exact Inference of Large Language Models

Researchers presented Sable, an argument system and single-GPU prover that verifies a committed, integer-quantized transformer produced a given token sequence from a given prompt bit for bit. Sable fixes an integer function F_O with no floating-point step, in which every rounding and normalization is defined exactly. The work matters because it makes large-language-model inference verifiable at the bit level rather than approximately.

We present Sable, an argument system and single-GPU prover for the statement that a committed, integer-quantized transformer produced a given token sequence from a given prompt, bit for bit. Sable fixes an integer function F O with no floating-point step, in which every rounding, normalization, soft