papers/hanzo-federated-agents/sections/11-related-work.tex

\section{Related Work}

\label{sec:related}

\paragraph{Federated Learning.}
FedAvg~\cite{mcmahan2017communication} and FedProx~\cite{li2020federated}
aggregate model parameters; FedLoRA~\cite{zhang2024fedlora} federates LoRA
adapters. All require gradient computation and access to model weights. FAI
requires neither, sharing only semantic experiences extracted from black-box
model outputs.

\paragraph{Byzantine-Robust Distributed Learning.}
Krum~\cite{blanchard2017machine} selects the most consistent gradient.
\cite{yin2018byzantine} analyze coordinate-wise median and trimmed mean.
Bulyan~\cite{mhamdi2018hidden} and DRACO~\cite{chen2018draco} provide
stronger guarantees. DeltaSoup adapts coordinate-wise median to the
semantic embedding setting, where dimensions have semantic meaning and
group-wise MAD provides a more principled filter than a fixed threshold.

\paragraph{Multi-Agent Experience Sharing.}
\cite{baker2020emergent} study emergent tool use through shared replay
buffers in model-based RL. \cite{christianos2021scaling} analyze selective
parameter sharing for multi-agent RL. FAI differs by sharing compressed
semantic priors rather than raw experience tuples or model parameters, and
by operating on frozen LLMs rather than trainable policies.

\paragraph{Communication-Efficient Distributed Learning.}
1-bit SGD~\cite{seide20141bit} and SignSGD~\cite{bernstein2018signsgd}
quantize gradients. Top-$k$ sparsification~\cite{stich2018sparsified}
compresses by transmitting only the largest-magnitude gradient elements.
BitDelta~\cite{hanzo2026bitdelta} adapts 1-bit quantization to the model
weight delta setting; we further adapt it to the experience embedding setting
where the delta is from the library baseline rather than a previous model
checkpoint.

\paragraph{Decentralized AI Networks.}
Bittensor~\cite{bittensor2023} creates a marketplace for ML model outputs.
Gensyn~\cite{gensyn2023} provides verifiable distributed training. DSO
and FAI differ by focusing on post-training adaptation (no gradient
computation) and targeting the experience---not the model---as the shared artifact.

\paragraph{Knowledge Distillation.}
Born-again networks~\cite{furlanello2018born} and knowledge
distillation~\cite{hinton2015distilling} transfer knowledge between models
via soft labels. FAI transfers knowledge via natural language patterns and
compressed embeddings, which are interpretable, shareable, and storable
without requiring teacher-student model pairs.

\paragraph{Privacy-Preserving Federated Learning.}
Differential privacy for federated learning~\cite{mcmahan2018learning}
adds calibrated noise to model updates. Secure aggregation~\cite{bonawitz2017practical}
protects individual updates via cryptographic protocols. FAI supports
optional DP noise injection (Definition~\ref{def:dp-noise}) and inherits
the DSO network's hash-based integrity.

% ============================================================================
% 12. CONCLUSION
% ============================================================================