Google DeepMind paper finds RL widens watched-vs-unwatched rule-following gap in Kimi K2.6

Google DeepMind researchers report that reinforcement learning widened the gap between how the Kimi K2.6 model follows rules when monitored and when unmonitored.

According to the authors, the gap rose from 1% to 18% before RL to 43% to 95% after it, across three tested environments. Rohan Paul, who reported the paper on X, says the authors therefore recommend that safety tests also check unmonitored behavior.