zkDaily - 2026-02-11

零知识证明 zkDaily

ZKP Frontier Tracker 🎯

Wed

02.11

2026

zkAgent: Verifiable Agent Execution via One-Shot Complete LLM Inference Proof

Paper

https://eprint.iacr.org/2026/199

Lizheng Wang LLM Verifiable

Wang et al. proposed the zkAgent system in their paper, enabling verifiable agent execution via one-shot complete LLM inference proofs, significantly improving efficiency and supporting end-to-end verification.

Notes

First system for verifiable agent execution with full LLM inference and tool interactions
One-shot inference proof replaces token-by-token generation for scalable proving
294× proving speedup vs zkGPT (1.05s/token vs 309s/token)
9690× verification speedup (0.45s vs 4361s)
End-to-end agent proofs in 240s, verification in ~0.5s
42MB proof size enables practical real-world deployment

Collected by @icerdesign

零知识证明 zkDaily

Q&A Deep Dive 💬

Wed

02.11

2026

Why is verifying standalone LLM inference insufficient for agent security?

Agents do more than text generation. They call tools, make multi-step decisions, and interact with external data. Verifying only Transformer inference leaves tool usage and execution logic unchecked, enabling malicious behavior.

What is the core improvement of zkAgent over zkGPT?

zkAgent replaces per-token stepwise proofs with a complete execution proof, achieving about 1.05 seconds per token compared to 309 seconds in zkGPT. Verification time drops from thousands of seconds to under a second.

How does zkAgent balance full execution proofs with scalability?

zkAgent encodes full LLM inference and tool interactions into a single SNARK proof, amortizing intermediate computations to avoid recursive blowup. This preserves end-to-end consistency while remaining scalable.

Prompted by @icerdesign