papers/hanzo-aso/hanzo-aso.tex

% Hanzo ASO (Active 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}

\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}

\DeclareMathOperator*{\argmin}{arg\,min}
\DeclareMathOperator*{\argmax}{arg\,max}
\DeclareMathOperator{\KL}{KL}
\DeclareMathOperator{\Tr}{Tr}
\DeclareMathOperator{\diag}{diag}
\DeclareMathOperator{\sign}{sign}
\DeclareMathOperator{\softmax}{softmax}

\title{Active Semantic Optimization: Training-Free Adaptation\\via Bayesian Product-of-Experts Decoding}
\author{
    Antje Worring, Zach Kelling \\
    Hanzo AI Research\\
    \textit{Hanzo AI Inc (Techstars '17), Los Angeles, CA}\\
    \texttt{[email protected]}
}
\date{September 2025}

\begin{document}

\hanzocoverpage

% ============================================================================
% ABSTRACT
% ============================================================================
\begin{abstract}
We present \textbf{Active Semantic Optimization (ASO)}, a training-free adaptation
framework that achieves performance competitive with fine-tuning at a fraction of
the cost (\$18 vs.\ \$10{,}000+). ASO combines Training-Free Group-Relative Policy
Optimization (TF-GRPO) with decode-time Product-of-Experts (PoE) ensemble to
adapt frozen large language models to new tasks without any gradient computation.
The key insight is that \emph{semantic advantages}---structured descriptions of
what makes one solution better than another---can be extracted from grouped
rollouts and applied as multiplicative token-level prior factors during
autoregressive decoding.

We formalize ASO as a Bayesian Product-of-Experts model where each experiential
prior contributes a factor $\phi_m(y_t \mid x, y_{<t})$ to the decoding
distribution, with closed-form expert weights $\eta_m \propto q_m / (1 - q_m)$
derived from quality attestation reliability. We prove that the PoE posterior
converges to the Bayes-optimal predictor under mild regularity conditions and
that the 1-bit compressed prior representation preserves convergence guarantees
with bounded approximation error.

Empirical evaluation across three code generation benchmarks demonstrates
consistent improvements: \textbf{72.8\%} on HumanEval (+9.1\% over baseline),
\textbf{68.4\%} on MBPP (+7.2\%), and \textbf{18.2\%} on SWE-bench Verified
(+5.7\%). Ablation studies confirm that each component---TF-GRPO rollouts, PoE
decoding, epistemic exploration, and 1-bit compression---contributes
meaningfully to the final result. The total adaptation cost for a new domain is
approximately \$18 in API calls, compared to \$10{,}000--\$50{,}000 for
LoRA/QLoRA fine-tuning and \$100{,}000+ for full RLHF.
\end{abstract}

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

\input{sections/01-introduction}
\input{sections/02-background}
\input{sections/03-training-free-group-relative-policy-opti}
\input{sections/04-product-of-experts-decoding-framework}
\input{sections/05-1-bit-prior-compression-bitdelta}
\input{sections/06-convergence-analysis}
\input{sections/07-cost-analysis}
\input{sections/08-hanzo-dev-agent-architecture}
\input{sections/09-experimental-results}
\input{sections/10-related-work}
\input{sections/11-conclusion}
\input{sections/12-tf-grpo-algorithm-complete-pseudocode}
\input{sections/13-proof-of-theorem-ref-thm-rate}
\input{sections/14-reproducibility-checklist}
\input{sections/99-extended-ablation-group-size-sensitivity}

\end{document}