Multi-hop reasoning collapses by 1M tokens; with a learning system it doesn't(zenodo.org)
2 pointsby lvrzhn7 hours ago1 comment