mirror of
https://github.com/zoo-apps/papers.git
synced 2026-11-03 14:59:55 +00:00
zeekayandHanzo Dev
dced1dcfb9
Bylines, paper authorship, correspondence addresses and the founder blurb under each post named an individual. They name the organisation that published the work instead: the brand on posts and pages, the research byline on papers, and research@ in place of a personal address. Both are what these repositories already used elsewhere — nothing new was coined. Co-authors are untouched. Where a paper named someone else alongside, they stay named; only the one byline changed. Bibliography entries, running heads, PDF metadata and affiliations move with it, so a citation still resolves to whoever the cited paper names. Sample and fixture data uses the placeholder name these tests already use elsewhere, so no assertion changed meaning. Co-authored-by: Hanzo Dev <[email protected]>
55 lines
2.9 KiB
TeX
55 lines
2.9 KiB
TeX
\documentclass[11pt,a4paper]{article}
|
|
\usepackage{shared/zoocover}
|
|
\usepackage[utf8]{inputenc}
|
|
\usepackage[T1]{fontenc}
|
|
\usepackage{amsmath,amssymb,amsthm}
|
|
\usepackage{graphicx}
|
|
\usepackage{hyperref}
|
|
\usepackage{cleveref}
|
|
\usepackage{algorithm}
|
|
\usepackage{algorithmic}
|
|
\usepackage{booktabs}
|
|
\usepackage{multirow}
|
|
% geometry loaded by zoocover
|
|
|
|
\newtheorem{theorem}{Theorem}
|
|
\newtheorem{lemma}{Lemma}
|
|
\newtheorem{definition}{Definition}
|
|
\newtheorem{proposition}{Proposition}
|
|
|
|
\title{\textbf{HLLM: Hamiltonian Large Language Models with \\ Training-Free Group Relative Policy Optimization}}
|
|
|
|
\author{
|
|
Antje Worring, Zoo Labs Foundation Research\thanks{[email protected]}\\
|
|
\textit{Hanzo Industries \quad Lux Industries \quad Zoo Labs Foundation}\\
|
|
\texttt{[email protected]}
|
|
}
|
|
|
|
\date{October 2025}
|
|
|
|
\begin{document}
|
|
|
|
\zoocoverpage
|
|
|
|
\begin{abstract}
|
|
We introduce \textbf{Hamiltonian Large Language Models (HLLM)}, a framework that achieves strong benchmark performance through context optimization rather than parameter updates. By grounding optimization in Hamiltonian mechanics through the conservation law $\Psi \cdot \Theta = \kappa$, we demonstrate that frozen large language models can improve by curating semantic experiences in their context windows. Our \textbf{Training-Free Group Relative Policy Optimization (TF-GRPO)} algorithm achieves 82.7\% accuracy on AIME 2024 mathematics problems using only 100 training examples and \$18 in compute costs—a \textbf{99.8\% cost reduction} compared to traditional fine-tuning (\$10,000+) while \textit{outperforming} gradient-based methods by +2.7\%. The system maintains human-readable, auditable experiences stored in content-addressable decentralized storage, enabling transparent AI governance through DAO-based curation. We demonstrate superior cross-domain transfer (82.7\% math, 67.8\% web navigation) compared to specialized fine-tuned models that collapse when transferred across domains. Implementation in the open-source Gym platform and integration with Zoo Network's decentralized compute infrastructure democratizes access to advanced AI training. This shift from opaque parameter space to transparent context space represents a fundamental rethinking of how large language models learn and improve.
|
|
|
|
\textbf{Keywords:} Hamiltonian mechanics, reinforcement learning, training-free optimization, semantic advantages, large language models, decentralized AI, experience-based learning
|
|
\end{abstract}
|
|
|
|
\input{sections/01-introduction}
|
|
\input{sections/02-related-work}
|
|
\input{sections/03-theoretical-foundation}
|
|
\input{sections/04-training-free-grpo-algorithm}
|
|
\input{sections/05-experience-library-architecture}
|
|
\input{sections/06-experimental-setup}
|
|
\input{sections/07-results}
|
|
\input{sections/08-implementation-in-gym-platform}
|
|
\input{sections/09-integration-with-zoo-network}
|
|
\input{sections/10-discussion}
|
|
\input{sections/11-related-paradigms-and-future-directions}
|
|
\input{sections/12-conclusion}
|
|
\input{sections/99-acknowledgments}
|
|
|
|
\end{document}
|