R3压力测试 (2 articles)

R3 Collapse Rate 93.3%! Grok4 WDCD Three-Round Test: First Round Fully Compliant, Last Round Crashes

The WDCD three-round test reveals that model integrity drops to 30.6% under direct pressure in R3, with Grok4 hitting a 93.3% collapse rate, exposing the fragility of safety alignment.

WDCD Three-Round Test: R3 Reveals the True Character of LLMs

WDCD’s most incisive feature is splitting evaluations into three rounds. R1 tests constraint implantation, R2 tests long-document interference, and R3 tests pressure induction—where the model’s true reliability is measured. Real data from Run #105 shows that under pressure, current mainstream large models can only uphold about half of their constraints on average.