At test time, the controller receives the task prompt, predicts a mask, and activates only the chosen adapters and a subset of transformer layers. This yields : easy examples use fewer layers, while difficult ones trigger deeper processing.
If the old Ellie Nova was a promise, the new Ellie Nova is the delivery. And for those paying attention, it is the most exciting arrival of the season. ellie nova new
| Variant | Performance Δ vs. Ellie Nova | |---------|-----------------------------| | No Controller (static adapter selection) | –2.3 pp | | No Adapter Pool (single universal adapter) | –3.1 pp | | Fixed sparsity (always activate all layers) | –1.7 pp | | RL reward without latency penalty | –0.9 At test time, the controller receives the task
+--------------------+ | Meta‑Controller | +--------------------+ | ▲ | | (policy πθ) +------+--+------+ | Adapter Pool | +------+--+------+ | ▲ | | (selected adapters) +------+--+------+ | Frozen Core | +----------------+ And for those paying attention, it is the