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

(zenodo.org) by lvrzhn | Aug 14, 2026 | 0 comments on HN