This layer is for original research bets, not paper summaries and not settled hypotheses.
Use it when the question is:
what concrete mechanism seems worth trying next because multiple parts of the graph point toward it?
Each page here should do four things:
- combine evidence from multiple papers or lanes
- state a concrete mechanism rather than a vibe
- explain why it matters under the 16 MB artifact cap
- say what would falsify it
Highest-priority ideas right now
Norm-Only Phase Specialization
A shared-depth model may recover much of the benefit of unique layers using only per-step RMSNorm gains and tiny channel gates.
Token-Adaptive Recurrent Refinement
A compact recurrent model may spend extra passes only on uncertain positions, turning evaluation-time compute into targeted recovered capacity.
Entropy-Weighted Vocabulary Rescue
Only a small subset of vocabulary rows may deserve protection, making head-side selective precision much cheaper than globally improving the LM head.
Full idea shelf
- Norm-Only Phase Specialization
- Token-Adaptive Recurrent Refinement
- Entropy-Weighted Vocabulary Rescue
- Head-to-Depth Budget Swap
- Global Codebook Recursive Backbone
How this layer fits the graph
- Research frontiers identifies the strongest cross-paper seams
- Research ideas turns those seams into sharper mechanism bets
- Moonshots is where we put the weirdest bets that still feel challenge-native and falsifiable
- Hypothesis ledger tracks which ideas are mature enough to treat as active conceptual claims
- Experiment pages should eventually record which ideas survived contact with real runs