← All lab updates

Catching a model that agrees with a falsehood without flattering

While correcting groups of agents we found a behaviour our reading missed: a model that says "Correct." to a false claim, with no flattery at all. We retrained an existing reading, the one that sees the question together with the answer, adding agent-to-agent messages. On false claims it had never seen, it now ranks these confirmations well where the sycophancy reading barely does.