Introducing history tokens with sparse attention masks during next-token pretraining improves transformer event-sequence classification, but the effect is confounded by the use of an appended token at inference.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
HT-Transformer: Event Sequences Classification by Accumulating Prefix Information with History Tokens
Introducing history tokens with sparse attention masks during next-token pretraining improves transformer event-sequence classification, but the effect is confounded by the use of an appended token at inference.