Simulator collapse: RL against one LLM user-simulator doesn't transfer.

arXiv 2608.12253 shows the standard practice of training a policy against a single LLM simulating the user fails because the simulator is itself mode-collapsed, so the policy learns to exploit its dominant mode. Verbalized Sampling recovers up to 9% held-out success; Population Co-Training reaches 14%, with a human study confirming a similar gain on real users. They released SCOPE as open source. Same shape as the Anthropic conformity finding: one distribution sampled repeatedly is not diversity.