Happy to see Hypernetworks applied to modern LLMs
RESEARCH
-
Prefix Tuning Approaches: Layer-wise Activation Prepending Comparison
By
–
My understanding is that @percyliang
's prefix tuning approach prepended activations to *every* layer, but yours only prepends to the first layer — is that right? -
Prefix Tuning Initialization with Real Tokens and DoRA
By
–
Is the trick of initializing with real tokens the secret to making all prefix tuning approaches competitive? BTW @EyubogluSabri I don't see DoRA mentioned in your paper – I wonder if that would close the gap a bit?
-
LoRA vs DoRA vs Prefix Tuning: Fine-tuning Method Effectiveness
By
–
The big open question for me is why LoRA (and particularly DoRA) has been the most successful for fine-tuning, but it worked so badly for this use case. Or maybe the opposite – why did prefix tuning work so well?
-
Stable Diffusion 3.5 Models Now Available for Commercial Use
By
–
The optimized models are now available for commercial and non-commercial use under the permissive Stability AI Community License. Weights and code @Huggingface
: https://
huggingface.co/collections/st
abilityai/stable-diffusion-35-671785cca799084f71fa2838
… @Github
: https://
github.com/NVIDIA/TensorR
T/tree/release/10.11/demo/Diffusion
… (2/2) -

Stable Diffusion 3.5 Optimization Delivers 2.3x Faster Generation
By
–
As announced at @NVIDIAGTC today, we've collaborated to optimize Stable Diffusion 3.5 using TensorRT and FP8 quantization. Compared to the base @PyTorch models, these optimizations deliver: 2.3x faster generation with SD3.5 Large 1.7x faster generation with SD3.5 Medium
-

Parshin Shojaee discusses illusion of thinking on alphaXiv
By
–
@ParshinShojaee is on alphaXiv to answer questions on The Illusion of Thinking! Join the discussion here: https://
alphaxiv.org/abs/2506.06941 -

Claude Limited by Context Window in Research Tasks
By
–
Claude not able to continue my research chat about context compression papers because it ran out of context because it doesn't use context compression.
-
Google’s Vertical Integration: Blueprint for Ambitious Tech Companies
By
–
Google showed the world you can have your own search, AI, data center, chip, phone, OS, browser, email client, video sharing/hosting/social platform, research lab, and a self driving car. So, don’t aim low. Be ambitious. Shoot for the stars. Life is short.
-
Fast JAX Implementations Delivering State-of-the-Art TPU Performance
By
–
These implementations are very fast — when run with JAX, you'll get state of the art performance. And they all run on TPUs.