ResearchBenchmarksSAE Feature Interpretability on Gemma 3 270M SAE Features (Value/Output)Follow21.82Input Interpretability ScoreQL_KEY-0.13445.565311.26516.9647May 30, 2026Evaluation ResultsMethodMethodLinksInput Interpretability ScoreInterpretability ScoreOutput Interpretability ScoreQL_KEY2026.0521.82—1.43QL_VALUE2026.0510.27—10.63TCa=10a=102026.0510.05—9.74TCa=5a=52026.0510.03—8.87TCa=1a=12026.0510—5.87ZO2026.055.78—2.75TL2026.052.23—5.38LLVALUE2026.052.14—6.13LLKEY2026.050.71—1.79Logit Lensvariant=LL_VALUEvariant=LL_VALUE2026.05—4.1—Query Lensvariant=QL_VALUEvariant=QL_VALUE2026.05—5.34—