zkDaily - 2026-03-27

零知识证明 zkDaily

ZKP Frontier Tracker 🎯

Fri

03.27

2026

zk-autoresearch implementation on Plonky3

OSS

https://x.com/realbarnakiss/status/2037120329131147430

@realbarnakiss AI Rust NTT

@realbarnakiss implemented zk-autoresearch on the Plonky3 ZK prover, using AI agents to optimize NTT implementation, achieving ~3% speedup.

Notes

First application of autoresearch pattern to a production ZK prover, Plonky3
AI agents analyzed Rust and Montgomery arithmetic to find redundancies in expert-optimized code
Optimizations targeted NTT inner loop: precomputing products, hoisting broadcasts, skipping multiplications by 1
74 iterations yielded 6 improvements and 57 regressions, all changes verified as mathematically equivalent
Safety ensured via end-to-end ZK proof generation and verification, soundness review pending
Round 2 planned to fix issues, full code and findings to be open-sourced

Collected by @icerdesign

零知识证明 zkDaily

Q&A Deep Dive 💬

Fri

03.27

2026

What is zk-autoresearch and what did it do in this experiment?

zk-autoresearch is an automated iterative optimization loop. It lets an AI repeatedly modify code, test performance, and keep improvements. In this experiment, it optimized the NTT in Plonky3 and achieved about a 3% speedup while preserving correctness.

How does the approach avoid breaking proof soundness?

The agent was restricted to implementation-level code like NTT and could not modify soundness-critical parts like FRI. Each iteration ran correctness tests and full end-to-end proof verification to ensure equivalence.

Why do NTT optimizations have system-wide impact in ZK provers?

NTT underpins polynomial operations used in constraint encoding, sum-check-related computations, and FRI preprocessing. Since it dominates runtime, even small optimizations propagate to system-wide speedups.

Prompted by @icerdesign