知Yizhi Daily
← Back to this edition

AI

Anthropic and other teams announce FlashProve progress in provable LLM inference

Edition 2026-11-05 · 08:00 Beijing time

Researchers from Anthropic, EPFL, MIT, and other institutions introduced FlashProve, a zkSNARK-based provable inference system used to prove that a response was indeed generated by a specified model. Using a 4-bit quantized 40-billion-parameter Llama 4 Maverick MoE model as an example, a single H100 generates about 110 proof tokens per second. Code and Lean proofs are open-sourced.

The research article reposted on Hacker News was co-authored by personnel from institutions including Anthropic, EPFL, MIT, StarkWare, RESI, and Succinct. The article notes that users currently have difficulty confirming whether an AI response came from the specific model claimed by the developer. Routing services may secretly substitute models to reduce costs, and attackers may tamper with token streams; judging solely by output quality is unreliable.

Cryptographic inference proofs allow model providers to first commit to a model and then prove that each response was generated by that model. The researchers focused on advancing a proof scheme known as zkSNARK and introduced the FlashProve system. The article says Claude has achieved improvements in zkSNARK inference proofs, with speeds significantly faster than previous open-source systems.

In terms of technical results, under a setup using the 400B-parameter mixture-of-experts model Llama 4 Maverick, 4-bit quantized weights, and integer arithmetic during the prefill phase, a single H100 GPU generates proofs at about 110 tokens per second. The proof size for each 1,024-token request is about 19MB, with GPU time roughly 70 to 95 times that of running the model directly with vLLM. Previous open-source systems were orders of magnitude slower.

The system's model commitment and subsequent inference proofs can hide model weights, but hiding the model architecture is not a goal. The team also used the formal verification language Lean to provide top-level auditable claims. The release includes a blog post, technical report, a Claude-written protocol specification, executable code containing a prover, a fast C++ verifier, and a Lean provably correct verifier, as well as Lean top-level mathematical proofs and theorem visualization tools. The code, proofs, and documentation are licensed under Apache 2.0.

Original sources

Hacker News ↗
Back to this editionOpen interactive reader