One static model does not fit all😭
We just dropped our latest work: Functional Neural Memory. Instead of static models, we generate custom "parameters" for every single input.
✅Prompt your model anytime
✅Instant personalization
✅Better instruction following
✅Flexible & dynamic memory (w/o memory bank✌️)
(🧵1/6)
🧠Our method is inspired by two key observations in model optimization:
1️⃣Shared LoRA/SFT is forced into compromise when optimizing heterogeneous objectives.
2️⃣Single LoRA targeting specific domain generalizes poorly.
(🧵2/6)
We propose HY-WU: a scalable framework for on-the-fly conditional generation of low-rank (LoRA) updates.
HY-WU synthesizes instance-conditioned adapter weights from hybrid image–instruction representations and injects them into a frozen backbone during the forward pass, producing instance-specific operators without test-time optimization.
(🧵3/6)
🎨Showcases of HY-WU's personalization ability, now you can try our cute dolls and cool clothes.
(🧵4/6)
LoRA — A cheap way to fine-tune a model by training a small add-on layer instead of changing all of the model's weights.
fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
Why it matters
A concrete alternative to fine-tuningTaking a trained model and training it a bit more on your own examples so it gets better at one specific job.Full definition → for personalization: the adapter is generated per request rather than trained per user, so there is nothing to store per user and no optimization step at inferenceRunning a trained model to get answers — the phase where AI is actually used, as opposed to trained.Full definition →.
Read the full thread on X ↗
