Tanaka links multi-agent consensus to sampling noise

Hot take: this is a strong caution against treating multi-agent agreement as evidence of truth or deliberation; in some regimes, consensus is just randomized path dependence with better branding.

• –The core claim is that group alignment can emerge from mutual in-context learning even when no agent had a real prior preference.
• –Quantized Simplex Gossip gives the paper a clean mechanistic story: sampled outputs feed back as evidence, so early randomness can lock in.
• –The scaling-law angle is the practical value here, since it suggests when larger groups or more bandwidth make drift worse or when weak signal starts to dominate.
• –The paper’s framing is useful for evaluating agent swarms, debate systems, and self-consistency setups where “agreement” may be an artifact.
• –Because it is a research paper, the impact is mainly conceptual and methodological rather than a shipping product feature.