New Method Restores Diversity in LLMs, Blocking ‘Artificial Hivemind’
The authors introduce meta‑persona anchoring plus a two‑stage temperature sieve that forces models to adopt distinct personas and then sample at extreme temperatures. On the 20‑billion‑parameter INFINITY‑CHAT benchmark, average pairwise similarity fell from ~0.85 to ~0.65, reviving response diversity.