On the Role of Learned Alignment Matrices in LatentMAS

Spursh Deshpande, Wenhao Lu

Keywords: multi-agent systems, latent communication, alignment matrix, ablation study, compositional agents, code generation

Abstract: Latent multi-agent systems propose that agents collaborate more effectively by sharing hidden states via a learned alignment matrix than through natural language. We test this premise through systematic ablations across three reasoning benchmarks. Replacing the trained alignment matrix with an identity transformation causes no statistically significant accuracy drop on any task. Removing cross-agent transfer entirely performs comparably or better than the full system on mathematical and scientific reasoning. On code generation, latent collaboration provides a genuine benefit over single-agent baselines, but text-based multi-agent communication substantially outperforms latent communication. A spectrum-matched random matrix also beats the trained one on code generation. We conclude that the trained alignment matrix may be functionally trivial, that genuine latent collaboration is task-specific, and that text communication remains superior for code despite being substantially slower.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 154