Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
1
0
3
LLMs exhibit a surprising "serial recall" bias, favoring tokens that immediately follow repeated tokens in the input sequence, and it's specifically driven by induction heads.