Communicating Activations Between Language Model Agents

TL;DR

This work introduces activation-based communication between language-model agents, replacing traditional natural-language messages with a one-shot transfer of intermediate activations to a partner model during inference. By pausing one model at layer $j$, fusing its post-layer activation with another model's at layer $k$ via a simple function $f$ (e.g., replace, sum, mean), and continuing the forward pass, the approach achieves compute savings and scales to new tasks without additional parameters or training. The authors provide a formal compute analysis and validate the method on multi-player coordination games and seven reasoning benchmarks, showing up to 27% improvements over natural-language communication with less compute and good generalization across model families. They also explore a task-agnostic linear mapping between activation spaces and discuss limitations around alignment and API access, offering avenues for future work in cross-model collaboration and interpretability of activations.

Abstract

Communication between multiple language model (LM) agents has been shown to scale up the reasoning ability of LMs. While natural language has been the dominant medium for inter-LM communication, it is not obvious this should be the standard: not only does natural language communication incur high inference costs that scale quickly with the number of both agents and messages, but also the decoding process abstracts away too much rich information that could be otherwise accessed from the internal activations. In this work, we propose a simple technique whereby LMs communicate via activations; concretely, we pause an LM $\textit{B}$'s computation at an intermediate layer, combine its current activation with another LM $\textit{A}$'s intermediate activation via some function $\textit{f}$, then pass $\textit{f}$'s output into the next layer of $\textit{B}$ and continue the forward pass till decoding is complete. This approach scales up LMs on new tasks with zero additional parameters and data, and saves a substantial amount of compute over natural language communication. We test our method with various functional forms $\textit{f}$ on two experimental setups--multi-player coordination games and reasoning benchmarks--and find that it achieves up to $27.0\%$ improvement over natural language communication across datasets with $<$$1/4$ the compute, illustrating the superiority and robustness of activations as an alternative "language" for communication between LMs.

Figures (11)

• Figure 1: Overview of activation communication. (Left) Our method involves (1) pausing a Transformer LM RGB]224, 236, 244$B$'s computation at layer $j$ in the residual stream; (2) combining its post-layer $j$ activation with another LM RGB]232,244,220$A$'s post-layer $k$ activation via some function RGB]255,244,204$f$; then (3) passing RGB]255,244,204$f$'s output into the next layer $j+1$ of RGB]224, 236, 244$B$ and continuing the forward pass till decoding is complete. (Right) Any function RGB]255,244,204$f$ can be used to combine $A$ and $B$'s activations; we explore letting RGB]255,244,204$f$ be the sum, mean, and replacement functions, as well as a task-agnostic learned linear layer (details in \ref{['method']}).
• Figure 2: 2D contour plots of accuracy over different values of $k$ and $j$ (the layers at which we access/edit activations for $A$/$B$ respectively).$k=j=26$ is roughly optimal ($\mathop{\raisebox{-0.275ex}{$\begin{tikzpicture}[x=1.2ex,y=1.85ex,line width=.1ex,line join=round, yshift=-0.285ex]\draw [draw=red,fill=red] (0,.5) -- (.5,1) -- (1,.5) -- (.5,0) -- (0,.5) -- cycle;\end{tikzpicture}$}}$) for both (a) Countries and (b) Tip Sheets.
• Figure 3: Accuracy (%) vs. compute (# FLOPs normalized by single $\mathtt{LLaMA}$-$\mathtt{3.2}$-$\mathtt{1B}$ forward pass) for various configurations of AC and NLD on the Biographies dataset. AC () yields the greatest performance gains per additional unit of inference-time compute over each baseline ().
• Figure 4: Example of AC on Biographies dataset.
• Figure 5: Example of AC on GSM8k dataset.