Compressed (caveman) reasoning traces train models to reason in ~2-3x fewer tokens at equal or better accuracy. For full ablation: sibling collection.
👋 Open to Work
Marco De Santis
marcodsn
AI & ML interests
LMs, datasets, cats?
Recent Activity
new activity about 1 month ago
marcodsn/catmind-1.2b:Weight merging and layers splicing updated a model about 1 month ago
marcodsn/remora-4b-v0 published a model about 1 month ago
marcodsn/remora-4b-v0Organizations
Compressed (caveman) reasoning traces train models to reason in ~2-3x fewer tokens at equal or better accuracy. For full ablation: sibling collection.
All ablation arms + dataset variants behind the caveman reasoning-compression study. For headline artifacts: see the main collection.
Synthetic Online Conversations
Synthetic Persona Bank
My submissions for the Agents-MCP-Hackathon