Search papers, labs, and topics across Lattice.
3
0
4
Sign-marginalized matching boosts coordinate recovery in RMSNorm Transformers by over 30%, revealing critical flaws in traditional alignment methods.
Revoking learned states in language models can achieve second-order accuracy, drastically improving safety without sacrificing performance.
Head Fisher alignment can be efficiently estimated in LLMs, revealing critical insights into task similarity that traditional metrics miss.
