Files
papers/experience-ledger-dso/sections/09-comparison-with-alternative-approaches.tex

65 lines
2.0 KiB
TeX
Raw Permalink Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
\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}