Skip to content

diegodiazgarciadev/behavioral-signal-training

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

4 Commits
 
 
 
 
 
 

Repository files navigation

Behavioral Signal Training (BST)

Multi-Agent Coordination Without Inference-Time Communication

Overview

Most modern multi-agent systems coordinate through explicit communication at inference time.

They exchange:

  • natural language messages
  • structured protocols
  • embeddings
  • or latent signals

This works, but it introduces latency, overhead, and additional system complexity.

This repository explores a different hypothesis:

What if agents do not need to communicate at inference at all?
What if coordination can be embedded during training?

Behavioral Signal Training (BST) is an exploratory research direction that investigates whether coordination can be learned into internal representations, rather than transmitted through a runtime channel.

This repository is shared to document the idea and invite discussion and critique.

Working Paper

The conceptual framework is described in detail in the working paper:

➡️ BST Working Paper (PDF)

➡️ BST Academic Format Paper (PDF)

The document formalizes:

  • the core hypothesis
  • the training setup
  • comparisons with existing approaches
  • proposed ablations
  • evaluation criteria

Core Idea

Instead of teaching agents:

  • what another agent says
  • or what another agent does

BST proposes training agents on how other agents internally react to shared situations.

The goal is not representational similarity.
The goal is representational compatibility.

If coordination is embedded during training, there may be nothing to transmit at inference.

Behavior triggers behavior.

What This Is (and Is Not)

BST is NOT:

  • a production-ready algorithm
  • a finished method
  • a claim of state-of-the-art performance

BST IS:

  • a research hypothesis
  • a representational perspective on coordination
  • an attempt to remove inference-time communication altogether

The value of this direction depends entirely on empirical validation.

High-Level Approach

During training:

  • agents are exposed to shared environments
  • and to behavioral signals derived from other agents’ internal activations

The objective is not to align representations, but to make them functionally compatible for joint behavior.

At inference:

  • no messages
  • no embeddings exchanged
  • no coordination channel

Only local observations and internal state.

Status

  • Conceptual framework: defined
  • Working paper: available
  • Experiments: planned or in progress
  • Results: not yet available

Negative results are considered informative.

Positioning

For a detailed comparison with existing approaches, including CTDE, TACO, CADP, LatentMAS, and related work, see the BST Working Paper.

Author

Diego
February 7, 2026

Coordination does not have to be transmitted.
It can be trained.

About

Exploratory research direction on multi-agent coordination without inference-time communication.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors