Skip to content

Communicating Activations Between Language Model Agents

Conference: ICML 2025
arXiv: 2501.14082
Code: None
Area: LLM Evaluation
Keywords: Multi-agent communication, activation space, model grafting, LLM reasoning, computational efficiency

TL;DR

A method is proposed to allow LLM agents to communicate via intermediate layer activations (instead of natural language) by injecting the activation vector of Model A into the intermediate layers during the forward pass of Model B. This requires zero additional parameters or training data, while improving performance by up to 27% compared to natural language communication across multiple reasoning benchmarks, using only 1/4 of the computation.

Background & Motivation

Background: Multiple LLM agents can improve reasoning capabilities through natural language dialogue (e.g., debates), but the computational cost grows rapidly with the number of agents and message length.

Limitations of Prior Work: (a) Natural language communication requires a complete generation-and-parsing cycle, which incurs significant computational overhead; (b) The decoding process compresses rich internal representations into a single token, discarding substantial information; (c) Studies show that the intermediate layers of a model contain richer entity representations than the output layer.

Key Challenge: Natural language is a communication medium designed for humans and may not be the optimal communication method for LLMs.

Goal: Can LLMs communicate using a more efficient, higher-density information medium—directly transmitting activation vectors?

Key Insight: Hernandez et al. found that models construct rich entity representations around the halfway point, but compress these into next-token predictions in later layers—implying that intermediate activations carry more information than the final output.

Core Idea: Pause the computation of Model B at layer \(j\), fuse the activation of Model A's \(k\)-th layer using a function \(f\), and then resume the forward pass of Model B.

Method

Overall Architecture

  1. Model A receives prompt \(x_A\) and performs forward propagation to layer \(k\) to obtain activation \(h_{A,k}\).
  2. Model B receives prompt \(x_B\) and performs forward propagation to layer \(j\) to obtain activation \(h_{B,j}\).
  3. Fuses the activation of the last token using function \(f\): \(h_{B,j}^{\text{new}} = f(h_{A,k}, h_{B,j})\).
  4. Resumes the forward propagation of the remaining layers of B to decode the output.

Key Designs

  1. Activation Fusion Function:

    • Function: Fuses the intermediate activations of the two models.
    • Mechanism: Tests four functions: sum (\(a+b\)), mean (\(\frac{a+b}{2}\)), replace (\(a\)), and a learned linear layer.
    • Design Motivation: Simple functions (such as mean) require zero extra parameters while already providing significant improvements.

  2. Layer Selection Strategy:

    • Function: Selects the optimal injection layer \(j\) and extraction layer \(k\).
    • Mechanism: Experiments show that intermediate layers (around 40-60% depth) yield the best results.
    • Design Motivation: Shallow representations are too primitive, while deep layer representations start to degenerate into next-token predictions.

Loss & Training

  • Completely training-free (sum, mean, and replace have no learnable parameters).
  • The learnable linear layer version only requires training on a small amount of task-agnostic data.

Key Experimental Results

Main Results

Task Natural Language Debate Activation Comm. (mean) Gain
GSM8K 52.3% 66.5% +14.2%
MMLU (Avg) 49.8% 53.1% +3.3%
Biographies 68.2% 86.7% +18.5%
Coordination Game 41.0% 68.0% +27.0%

Computational cost: <1/4 of natural language communication

Ablation Study

Configuration Performance Description
sum Good Summing information might be too aggressive
mean Best Balances information between the two models
replace Medium Completely discards information from B
learned linear Slightly better than mean Requires a small amount of training
Shallow Injection (10%) Poor Information is too primitive
Intermediate Injection (50%) Best Equilibrium point

Key Findings

  • Activation communication is also effective on smaller models (1-7B), unlike natural language debate which is only effective on large models.
  • Intermediate layer activations contain richer information than the final output.
  • It requires only a partial forward pass of A + a complete forward pass of B, resulting in extremely high computational efficiency.

Highlights & Insights

  • "Telepathy" between LLMs—skipping the bottleneck of language to directly transmit high-dimensional representations, which is conceptually elegant.
  • The finding that intermediate layers are the "information-richest" layers holds independent research value.
  • The method is extremely lightweight—zero parameters, zero training, and plug-and-play.

Limitations & Future Work

  • Requires the two models to have the same hidden dimension size (otherwise projection layers are needed).
  • Only transmits the activation of the last token, failing to leverage information from all tokens in the sequence.
  • Sensitive to the choice of model pairs—performance varies significantly across different combinations.
  • Safety implications have not been analyzed (whether activation injection might lead to harmful outputs).
  • vs Multi-agent Debate: Natural language communication, 4×+ computational cost, poorer performance.
  • vs CIPHER: Communication via tokenizer embeddings, causing greater information loss than activations.
  • vs CALM: Fusion of activations via learned attention layers, requiring training, whereas ours does not.

Rating

  • Novelty: ⭐⭐⭐⭐⭐ Pioneering use of activations instead of language for LLM-to-LLM communication.
  • Experimental Thoroughness: ⭐⭐⭐⭐ Multiple tasks, multiple fusion functions, and multiple models.
  • Writing Quality: ⭐⭐⭐⭐ Clear motivation and concise methodology.
  • Value: ⭐⭐⭐⭐⭐ Holds significant importance for multi-agent LLM systems.