Jayce: framework-free prototype memory lets local LLMs learn facts 1.6x-4x faster than backprop
kavanutz · reddit · 2026-09-22
Reddit user kavanutz open-sourced Jayce, an experimental project inspired by how toddlers learn words from a few examples and quick corrections.
The core idea is Adaptive Prototype Memory (APM): instead of touching model weights, it grabs the LLM's raw context vectors and drops them into a fixed pool of 4,096 prototype slots. When corrected, the closest prototype physically shifts toward the new data on the spot.
Reported benchmark results:
- Training updates run 1.6x–4x faster than standard Adam backprop
- On sequential MNIST tests, higher accuracy than backprop with the same number of examples
- Strict memory ceiling, fully offline, running with a local Qwen3-4B GGUF
The whole thing is framework-free — pure NumPy plus native Java — with full benchmark data, a vector-shifting breakdown, and a terminal script to try the learning loop yourself. Repo: Loophole-LLC/Jayce.
More from Research
- Xiaomi's MiMo-V2.6 lands: 1.02T-param MoE takes top open model spot on Artificial Analysis — _AndrewZhao · 2026-09-22
- O'Reilly builds a working data vocabulary for the semantic era, from warehouses to ontologies — rseroter · 2026-09-22
- Researchers Speed Up DSA Method for Comparing Neural Dynamics, Making It Generalizable Across Domains — GretaTuckute · 2026-09-22
- Glance reads structured visual answers from a frozen 4B VLM, cutting GPU cost up to 85% — multiply_matrix · 2026-09-22
- krea2-bbox-turbo Full-Rank Finetune Hits Epoch 14, Release Planned at Epoch 20 — Amazing_Painter_7692 · 2026-09-22
- OpenAI forms math advisory group after its AI resolves 100+ open problems — TechCrunch AI · 2026-09-22