Multi-hop reasoning collapses by 1M tokens; with a learning system it doesn't

 (zenodo.org)

2 points | by lvrzhn 7 hours ago ago

1 comments