Reproduce “TopK Language Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Ryosuke Takahashi, Tatsuro Inaba, Kentaro Inui, Benjamin Heinzerling

Researchers from Tohoku University and RIKEN developed TopK Language Models, which incorporate a TopK activation function directly into transformer architectures to provide inherent interpretability. These models achieve competitive performance, with validation perplexity increasing from 11.76 to 14.96 for a D=1024, L=24 model, while offering enhanced neuron specialization, causal concept steerability, and the ability to trace concept formation over training.

Key insights

  • Intrinsic interpretability can be achieved in transformer LMs by directly incorporating sparse activation functions without significant performance degradation.
  • Enforcing TopK sparsity leads to highly specialized, token-selective, and semantically coherent neurons within the model.
  • The proposed architecture enables unprecedented traceability of individual neuron specialization and concept formation across different training checkpoints and model layers.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.