Research · Track 02
World Models and JEPA
Predictive architectures that learn structure from far less data.
6–10 contributors
Open — seeking a lead
5–15 hrs/week contributing · 20+ hrs/week core
Why now
Context
Joint Embedding Predictive Architectures point past purely generative AI toward sample-efficient learning, grounded reasoning and world models for embodied agents. For a research programme with limited compute, an architecture family whose central claim is sample efficiency is not a fashionable choice — it is the practical one.
There is a draft paper already in circulation among members on a JEPA-based small reasoning model with metacognition. It is not yet publication grade. Part of this track’s first job is to decide honestly whether it can get there, and to say so publicly either way.
Sub-themes
Scope
- I-JEPA and V-JEPA replications and improvements
- World models for planning and embodied agents
- Self-supervised learning for low-resource Indic content
- Predictive representation versus generative pretraining trade-offs
- Energy-based reasoning and inference-time search
- Evaluation suites for representation quality beyond linear probing
Year one deliverables
Output
- One to two architecture or position papers
- An open JEPA implementation with ablations
If replication does not reach published baselines within two quarters, the track narrows to the Indic low-resource question only, or closes.