Setting synthetic-model hyperparameters without ground-truth LLM features

Determine appropriate values for the hyperparameters used to model large language model representations in the SynthSAEBench synthetic data framework—specifically the number of features, the correlation levels between features, the hierarchy degree, and the superposition level—given the current absence of ground-truth knowledge of LLM features, so that synthetic benchmarks can be calibrated to real neural network behavior.

Background

The paper’s synthetic framework aims to replicate real neural network phenomena such as superposition, correlation, and hierarchy while providing ground-truth features to evaluate sparse autoencoders. However, the authors emphasize that true ground-truth feature properties in LLMs are not known, which prevents exact calibration of key synthetic hyperparameters.

This uncertainty limits direct alignment of synthetic benchmarks to real models, making it difficult to choose values for number of features, correlation strengths, hierarchical dependencies, and the extent of superposition. Resolving this would allow more realistic and reliable benchmarking of SAE architectures against phenomena observed in LLMs.