ResearchBenchmarksSteering Validation on Real Agents LLMAgent, Claude Haiku 4.5 N-scaling Phase 4 (val)Follow83.7Haiku Judge AccuracyDACS61.75667.45373.1578.847Apr 9, 2026Evaluation ResultsMethodMethodLinksHaiku Judge AccuracyGPT-4o-mini Judge AccuracyContamination RateAvg Context TokensP-valueDACSN=5, Scenario=ra2_n5N=5, Scenario=ra2_n52026.0483.789.740.37990.0008DACSN=3, Scenario=ra1_n3N=3, Scenario=ra1_n32026.0479.885.459.96540.0023BaseN=5, Scenario=ra2_n5N=5, Scenario=ra2_n52026.0463.368.295.82,2750.0008BaseN=3, Scenario=ra1_n3N=3, Scenario=ra1_n32026.0462.667.795.81,3610.0023