papers/hanzo-dso/hanzo-dso.tex

% Hanzo DSO (Decentralized Semantic Optimization) Paper
% Fully self-contained -- no \input{} directives
\documentclass[11pt]{article}
\usepackage{../shared/hanzocover}
\usepackage{amsmath, amssymb, amsthm}
\usepackage{mathtools}
\usepackage{bm}
\usepackage{graphicx}
\usepackage{booktabs}
\usepackage{hyperref}
\usepackage{enumitem}
\usepackage{algorithm}
\usepackage{algpseudocode}
\usepackage{xcolor}
\usepackage[numbers]{natbib}
\usepackage{float}
\usepackage{subcaption}
\usepackage{tabularx}
\usepackage{multirow}
\usepackage{tikz}

\hypersetup{colorlinks=true,linkcolor=black,citecolor=blue,urlcolor=blue}

\newtheorem{theorem}{Theorem}
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{proposition}[theorem]{Proposition}
\newtheorem{corollary}[theorem]{Corollary}
\newtheorem{definition}{Definition}
\newtheorem{remark}{Remark}
\newtheorem{assumption}{Assumption}

\DeclareMathOperator*{\argmin}{arg\,min}
\DeclareMathOperator*{\argmax}{arg\,max}
\DeclareMathOperator{\KL}{KL}
\DeclareMathOperator{\Tr}{Tr}
\DeclareMathOperator{\diag}{diag}
\DeclareMathOperator{\sign}{sign}
\DeclareMathOperator{\med}{med}
\DeclareMathOperator{\wmed}{wmed}
\DeclareMathOperator{\rank}{rank}

\title{Decentralized Semantic Optimization:\\Byzantine-Robust Prior Aggregation\\for Collective Model Adaptation}
\author{
    Antje Worring, Hanzo AI Research \\
    Hanzo AI Research\\
    \textit{Hanzo AI Inc (Techstars '17), San Francisco, CA}\\
    \textit{Zoo Labs Foundation (501(c)(3))}\\
    \texttt{[email protected]}
}
\date{December 2023}

\begin{document}

\hanzocoverpage

% ============================================================================
% ABSTRACT
% ============================================================================
\begin{abstract}
We present \textbf{Decentralized Semantic Optimization (DSO)}, a protocol for
sharing and aggregating experiential priors across distributed language model
agents without parameter updates. While Active Semantic Optimization (ASO)
enables individual agents to adapt via training-free Product-of-Experts
decoding, DSO extends this to the multi-agent setting by defining a
Byzantine-fault-tolerant aggregation protocol over compressed semantic priors.

DSO introduces three key mechanisms: (i) a \emph{three-layer semantic storage
architecture} combining on-chain registry, content-addressed off-chain storage
(IPFS/Arweave), and local replica databases; (ii) a \emph{Byzantine-robust
weighted median aggregation} scheme with formal guarantees under $f < n/3$
Byzantine faults; and (iii) a \emph{Sybil-resistant gossip protocol} with
stake-weighted propagation and quality-based filtering.

We prove that DSO aggregation converges to the honest-majority posterior at rate
$O(1/\sqrt{n})$ where $n$ is the number of honest participants, and that the
protocol tolerates up to $f < n/3$ Byzantine nodes while maintaining
$\varepsilon$-optimal prior quality. Empirical evaluation on multi-agent code
generation demonstrates \textbf{15.2\% improvement} when 10 agents share priors
vs.\ isolated operation, with resilience to 20\% Byzantine participants
(degradation $< 0.9\%$).
\end{abstract}

% ============================================================================
% 1. INTRODUCTION
% ============================================================================

\input{sections/01-introduction}
\input{sections/02-background}
\input{sections/03-three-layer-semantic-storage-architectur}
\input{sections/04-dso-protocol-specification}
\input{sections/05-byzantine-robust-aggregation}
\input{sections/06-sybil-resistance}
\input{sections/07-gossip-protocol}
\input{sections/08-convergence-analysis}
\input{sections/09-incentive-mechanisms-and-poai-integratio}
\input{sections/10-experimental-evaluation}
\input{sections/11-related-work}
\input{sections/12-conclusion}
\input{sections/13-security-analysis-extended}
\input{sections/14-gossip-protocol-formal-analysis}
\input{sections/15-full-contract-specification}
\input{sections/99-reproducibility-details}

\end{document}