Reproduce “Symmetry Breaking in Transformers for Efficient and Interpretable Training” 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
Eva Silverstein, Daniel Kunin, Vasudev Shyam
The authors test whether batchwise, untrained query and value biases can break attention symmetries that may constrain optimizer motion, without adding trainable parameters. In GPT-2 124M experiments, they compare ECD validation loss after 500 million FineWeb-Edu tokens across four seeds. With value-bias mean 0.5, loss falls from 3.88 to 3.39; puzzle gains are mixed and not statistically significant at this scale.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.