LatentResMAS

Natural language is a lossy channel for agents talking to each other: every exchange gets compressed into words and decompressed again, and something is lost each time. Co-authored work on having agents communicate directly through latent representations instead, with a residual connection so the signal does not degrade as it passes through multiple agents in sequence, the same intuition behind residual connections in deep networks. The gains show up most on complex reasoning and coding tasks. The honest limitation: it only works when every agent shares the same underlying model, so it is more a proof that the architecture matters than a drop-in production system. Currently, I have submitted this work to a conference, so keeping this brief.