papers/zip-002-zen-reranker/sections/03-motivation.tex

\section{Motivation}


\subsection{DSO Protocol Overview}

Decentralized Semantic Optimization enables LLMs to improve via shared experiences without parameter synchronization. The protocol operates in three phases:

\textbf{Phase 1: Experience Generation}
\begin{enumerate}
\item LLM generates $G$ rollouts for query $q$: $\{o_1, \ldots, o_G\}$
\item Compute rewards: $\{r_1, \ldots, r_G\}$ via reward model or ground truth
\item Extract semantic advantage via LLM introspection:
\end{enumerate}

\begin{algorithm}[H]
\caption{Semantic Advantage Extraction}
\begin{algorithmic}[1]
\STATE $\text{best} \gets \arg\max_i r_i$
\STATE $\text{worst} \gets \arg\min_i r_i$
\STATE $\text{prompt} \gets \text{``Compare best vs worst, extract strategic insight''}$
\STATE $\text{experience} \gets \text{LLM.generate}(\text{prompt})$
\RETURN $\text{experience}$
\end{algorithmic}
\end{algorithm}

\textbf{Phase 2: Embedding \& Submission}
\begin{enumerate}
\item Encode experience: $\mathbf{e} = \text{ZenReranker.encode}(\text{experience})$
\item Compress: $\mathbf{c} = \text{BitDelta}(\mathbf{e})$ (964 bytes)
\item Submit to network: $\text{Network.submit}(\mathbf{c})$
\end{enumerate}

\textbf{Phase 3: Retrieval \& Application}
\begin{enumerate}
\item Encode query: $\mathbf{q} = \text{ZenReranker.encode}(\text{query})$
\item Retrieve: $\{\mathbf{e}_1, \ldots, \mathbf{e}_k\} = \text{Network.retrieve}(\mathbf{q}, k)$
\item Inject into context: $\text{prompt} = \text{experiences} + \text{query}$
\item Generate: $\text{output} = \text{LLM.generate}(\text{prompt})$
\end{enumerate}

\subsection{The Alignment Bottleneck}

Without native canonical embeddings, each model requires alignment:

\begin{equation}
\text{Latency}_{\text{total}} = \underbrace{t_{\text{encode}}}_{\text{Model-specific}} + \underbrace{t_{\text{align}}}_{\text{Overhead}} + \underbrace{t_{\text{compress}}}_{\text{BitDelta}}
\end{equation}

Measurements on A100 GPU:
\begin{itemize}
\item $t_{\text{encode}}(\text{Qwen3-8B, 4096-dim}) = 18.3$ms
\item $t_{\text{align}}(4096 \to 7680) = 9.7$ms (53\% overhead!)
\item $t_{\text{compress}}(7680 \to 964\text{ bytes}) = 3.2$ms
\item \textbf{Total:} 31.2ms
\end{itemize}

With Zen-Reranker:
\begin{itemize}
\item $t_{\text{encode}}(\text{Zen-Reranker, 7680-dim}) = 18.3$ms
\item $t_{\text{align}} = 0$ms (eliminated!)
\item $t_{\text{compress}}(7680 \to 964\text{ bytes}) = 3.2$ms
\item \textbf{Total:} 21.5ms (\textbf{31\% reduction})
\end{itemize}

\subsection{Compression Inefficiency}

Aligned embeddings exhibit higher entropy, degrading BitDelta compression:

\begin{equation}
\text{Compression Ratio} = \frac{\text{Original Size}}{\text{Compressed Size}} = \frac{7680 \times 4\text{ bytes}}{\text{|BitDelta|}(\mathbf{e})}
\end{equation}

\begin{table}[h]
\centering
\caption{Compression performance comparison}
\begin{tabular}{lrrr}
\toprule
\textbf{Approach} & \textbf{Original (bytes)} & \textbf{Compressed (bytes)} & \textbf{Ratio} \\
\midrule
BGE-Large (1024) & 4,096 & 152 & 26.9$\times$ \\
Qwen3-8B Aligned (7680) & 30,720 & 1,027 & 29.9$\times$ \\
\textbf{Zen-Reranker (7680)} & \textbf{30,720} & \textbf{964} & \textbf{31.87$\times$} \\
\bottomrule
\end{tabular}
\end{table}

\textbf{Why does native 7680-dim compress better?} Alignment introduces noise from projection error. Zen-Reranker learns smooth manifold structure optimized for delta encoding.