Fetching from the wire…
01
02
03
04
05
06
07
08
Models2026-03-23 · source-backed
MiroThinker (open weights) achieves 81.9% on GAIA-Text via "interactive scaling" — training the model to systematically leverage environment feedback as a third scaling dimension beyond size and context length. Supports 256K context and up to 600 tool calls per task. Unlike isolated test-time compute, interactive scaling uses external corrections to keep trajectories on track.
Each link below shares sources, entities, or timing with this story.
Anthropic benchmarked against GAIA / Shared entity: GAIA / Shared topic / What happened next
Linked by a graph relationship (Anthropic benchmarked against GAIA); both cover GAIA; overlapping topics (call, context).
Anthropic benchmarked against GAIA / Shared entity: Text / Earlier coverage
Linked by a graph relationship (Anthropic benchmarked against GAIA); both cover Text; earlier Text coverage from 2026-03-12.
Anthropic benchmarked against GAIA / Same source domain
Linked by a graph relationship (Anthropic benchmarked against GAIA); reported by the same outlet (huggingface.co).
Anthropic benchmarked against GAIA / Shared topic
Linked by a graph relationship (Anthropic benchmarked against GAIA); overlapping topics (context, isolated).
Linked by a graph relationship (Anthropic benchmarked against GAIA); overlapping topics (call, context).
Linked by a graph relationship (Anthropic benchmarked against GAIA); overlapping topics (context, environment).
Linked by a graph relationship (Anthropic benchmarked against GAIA); overlapping topics (compute, interactive).
Linked by a graph relationship (Anthropic benchmarked against GAIA); overlapping topics (context, external).