Self-Supervised Inference of Agents in Trustless Environments
Vladyslav Larin, Ivan Nikitin, Alexander Firsov
TL;DR
This work tackles trustless decentralized AI inference by introducing a swarm of self-supervised agents that jointly infer and rank responses, using LLMs as both generators and classifiers. The framework deploys a three-phase swarm consensus—response generation, selective ranking, and final selection—with encrypted communications and a weighted aggregation to deter manipulation. It also models adversarial behaviors, including Sybil and prompt-engineering attacks, and proposes economic and architectural defenses, achieving ultra-low latency (under 125 ms) on large language models such as Llama 3 405B. Empirical evaluation contrasts latency against PoQ, ZK-based methods, and TEEs, demonstrating significant speedups while maintaining robustness through agent-rating mechanisms and anti-manipulation incentives. This approach has practical implications for scalable, real-time, trustless AI services in decentralized infrastructures.
Abstract
In this paper, we propose a novel approach where agents can form swarms to produce high-quality responses effectively. This is accomplished by utilizing agents capable of data inference and ranking, which can be effectively implemented using LLMs as response classifiers. We assess existing approaches for trustless agent inference, define our methodology, estimate practical parameters, and model various types of malicious agent attacks. Our method leverages the collective intelligence of swarms, ensuring robust and efficient decentralized AI inference with better accuracy, security, and reliability. We show that our approach is an order of magnitude faster than other trustless inference strategies reaching less than 125 ms validation latency.