First Passage Attention Uses Path Extrema to Improve Context Erasure

First Passage Attention (FPA) improves upon softmax attention by using a learned scalar field to selectively erase intermediate context through path extrema rather than monotonic decay. This allows a constant-dimension FPA head to maintain relevant tokens more efficiently than standard positional bias methods.