Interpreting Latent Agent Communication

As multi-agent systems powered by Vision-Language Models (VLMs) become more complex, the internal "language" or latent representations used for inter-agent communication remain largely opaque. This research proposes a post-hoc sparse coding approach to decompose these latent signals. By applying sparse autoencoders or similar dictionary learning techniques to the hidden states of communicating agents, the authors aim to map high-dimensional, dense latent vectors into human-interpretable, sparse features.

Methodology and Implications

The core argument is that latent communication is not merely noise but contains structured, task-relevant information that can be reconstructed through sparse coding. By treating the latent space as a superposition of features, the method allows researchers to:

• Identify specific communication primitives: Isolate which latent features correspond to visual grounding, spatial reasoning, or strategic planning during agent interaction.
• Post-hoc analysis: Unlike training-time interpretability methods, this approach does not require modifying the underlying VLM architecture, making it applicable to existing, pre-trained models.
• Improve transparency: By visualizing the activation of sparse features during multi-step tasks, developers can debug agent failures where communication breakdown is suspected.

This work provides a technical foundation for auditing agent-to-agent protocols, moving toward more reliable and explainable multi-agent AI systems.