Search papers, labs, and topics across Lattice.
University of Copenhagen
1
0
2
Context sensitivity in LLMs evolves significantly across training stages, revealing that SFT biases models towards simpler contexts that can be both reinforced and reshaped later on.
