Language models integrate temporal anchors into factual recall
Temporal anchors resolve context-dependent facts via subject-mediated or direct MLP routes in open-weight models.
Stories per day · 30 days
1 RSS
2 stories, newest first · one card per story
Stories tagged #factual-recall
First-token cueing recovers closed-book facts LLMs fail to start
Prefilling only the gold answer's first token fixes a fifth to a half of multi-token English QA errors across 16 open models.