ResearchBenchmarksDownstream behavior evaluation on GPT-2 small (layer 6) residual-stream activationsFollow0.092Delta CERSAE0.088480.112240.1360.15976Jun 12, 2026Evaluation ResultsMethodMethodLinksDelta CEPerformance RateRSAEActivation Function Fa...Activation Function Family=TopK2026.060.09298.55RSAEActivation Function Fa...Activation Function Family=ReLU2026.060.12398.07TopK SAEActivation Function Fa...Activation Function Family=TopK2026.060.13697.89ReLU SAEActivation Function Fa...Activation Function Family=ReLU2026.060.1897.17