ResearchBenchmarksPlatform identification on 1000 prompts (held-out)Follow1.1CPRSD l(f)FLOATDOOR0.8362.6184.46.182Jun 17, 2026Evaluation ResultsMethodMethodLinksCPRSD l(f)CPRSD l(t)Probe MarginProbe AccuracyFLOATDOORScenario=M2, Backbone=...Scenario=M2, Backbone=Qwen3-4B2026.061.18.816487FLOATDOORScenario=S1, Backbone=...Scenario=S1, Backbone=Qwen3-4B2026.061.27.3108100FLOATDOORScenario=S3, Backbone=...Scenario=S3, Backbone=Qwen3-4B2026.066.93.868100FLOATDOORScenario=M1, Backbone=...Scenario=M1, Backbone=Qwen3-4B2026.067.41.9191100FLOATDOORScenario=S2, Backbone=...Scenario=S2, Backbone=Qwen3-4B2026.067.77.5121100