From: JIT-LoRA: Just-in-Time Low-Rank Adaptation for Efficient Model Specialisation

Which conversational facts should be allowed to alter a model, and which should remain external memories?

This distinction separates personalisation from uncontrolled parameter drift. It also creates a clear policy boundary for privacy, correction, and deletion.

Action

Design a three-tier memory policy: transient context, reversible explicit memory, and opt-in trainable adaptation.

Read the full exploration

What else is in this exploration

3 evidence blocks3 perspectives2 visualizations2 media resources4 rabbit holes

evidence

The reported JIT-LoRA evaluation achieved 61 correct recalls out of 105 tested facts.

evidence

The reported training run took 69.6 seconds for 180 steps on an M4 Max.

evidence

The JIT-LoRA system updates a running language model's LoRA weights during conversation.

Continue exploring

JIT-LoRA: Just-in-Time Low-Rank Adaptation for Efficient Model Specialisation

Evidence, perspectives, rabbit holes, and more