mirror of
https://github.com/zoo-apps/papers.git
synced 2026-10-30 14:56:54 +00:00
65 lines
2.0 KiB
TeX
65 lines
2.0 KiB
TeX
\section{Comparison with Alternative Approaches}
|
||
|
||
|
||
\subsection{Fine-Tuning}
|
||
|
||
\textbf{Advantages of Fine-Tuning}:
|
||
\begin{itemize}
|
||
\item Optimal performance on single domain (if sufficient data)
|
||
\item No inference-time overhead (parameters encode all knowledge)
|
||
\end{itemize}
|
||
|
||
\textbf{Disadvantages}:
|
||
\begin{itemize}
|
||
\item Cost: 667× more expensive (\$12K vs. \$18)
|
||
\item Catastrophic forgetting across domains (Table~\ref{tab:transfer})
|
||
\item Opacity: No interpretability of parameter changes
|
||
\item Non-composable: Cannot combine multiple fine-tunes
|
||
\end{itemize}
|
||
|
||
\subsection{RLHF and GRPO}
|
||
|
||
\textbf{Vanilla GRPO} improves upon RLHF by eliminating value networks, reducing training instability. However:
|
||
|
||
\begin{itemize}
|
||
\item Still updates parameters ($\theta \neq \theta_0$)
|
||
\item Advantages are numerical, not semantic
|
||
\item No audit trail for why performance changed
|
||
\end{itemize}
|
||
|
||
\textbf{\DSO{}} extends GRPO into the semantic realm: advantages become human-readable experiences, enabling decentralized curation.
|
||
|
||
\subsection{Prompt Engineering}
|
||
|
||
Manual prompt crafting is effective but:
|
||
|
||
\begin{itemize}
|
||
\item Requires expert knowledge (not scalable)
|
||
\item Static (does not improve with model usage)
|
||
\item No mechanism for community contributions
|
||
\end{itemize}
|
||
|
||
\textbf{\DSO{}} automates prompt refinement via introspection and scales via decentralized governance.
|
||
|
||
\subsection{RAG}
|
||
|
||
RAG retrieves factual documents, while \DSO{} retrieves \textit{strategic insights}. Key differences:
|
||
|
||
\begin{table}[t]
|
||
\centering
|
||
\caption{RAG vs. \DSO{}}
|
||
\label{tab:rag-vs-dso}
|
||
\begin{tabular}{lcc}
|
||
\toprule
|
||
\textbf{Aspect} & \textbf{RAG} & \textbf{\DSO{}} \\
|
||
\midrule
|
||
Retrieval target & Facts & Strategies \\
|
||
Source & External corpus & Model introspection \\
|
||
Update mechanism & Manual corpus edit & Automated extraction \\
|
||
Governance & Centralized & Decentralized \DAO{} \\
|
||
Performance gain & +2--5\% & +10--15\% \\
|
||
\bottomrule
|
||
\end{tabular}
|
||
\end{table}
|
||
|