AI Researcher @ FAIR (Meta) | Ph.D. Candidate @ UCL
Oct 14 • 7 tweets • 2 min read
🧵 New paper from FAIR (Meta) on recursion + latent reasoning: "Encode,Think,Decode (ETD): Scaling reasoning through recursive latent thoughts".
ETD improves the reasoning of base model by training it to iterate over a subset of reasoning-critical layers during mid-training.(1/n)
@nicola_cancedda @AldoLipani Motivation:
- Depth, not just parameter count, is key to stronger compositional reasoning.
- Chain-of-Thought does not fully capture varying amounts of internal computation.
- Interpretability studies show that reasoning is concentrated in a deeper layers. (2/n)
