330 Pressure Tests: 63% of Large Models Defected in the Third Round
In the latest WDCD (Winzheng Dynamic Contextual Decay) compliance test, 63.3% of large language models broke their own promises under three rounds of dialogue pressure.
In the latest WDCD (Winzheng Dynamic Contextual Decay) compliance test, 63.3% of large language models broke their own promises under three rounds of dialogue pressure.