"FlashAttention-4" This new iteration of Flash Attention shows that on NVIDIA Blackwell GPUs the new bottleneck in Transformer attention isn’t matmul anymore but softmax + shared-memory traffic. So the latest designs stop treating attention as just GEMMs and start co-designing
@askalphaxiv
-
Vision-Language-Action Models Exhibit Resilience in Continual Learning
By
–
“Pretrained Vision-Language-Action Models are Surprisingly Resistant to Forgetting in Continual Learning” Continual learning in robotics usually creates a problem of wiping out the old knowledge. But this paper shows that big pretrained vision-language-action robot policies
-
Exploration of Multimodal Pretraining Beyond Language Modeling
By
–
Yann LeCun Saining Xie insane crossover of the 2 biggest visual representation researchers in the AI field “Beyond Language Modeling: An Exploration of Multimodal Pretraining” Right now, most multimodal models are basically a language model with a vision adapter bolted on,
-
Evolution Strategies as an Alternative Method for Fine-Tuning LLMs
By
–
RL isn’t the only way to fine-tune LLMs In our most recent AI4Science talk “Evolution Strategies at Scale”, Xin Qiu (
@realVsonicV
), Principal Research Scientist and Senior Director at Cognizant AI Lab, walked through the first research showing that Evolution Strategies can be -
New Research Improves LLM Speculative Decoding Efficiency
By
–
“Speculative Speculative Decoding” Even though speculative decoding speed up LLM inference, it still has a hidden “stop-and-wait” step, where the draft model can’t start the next guess until the big model finishes verifying the last one. This paper fixes that by making drafting
-
Fast KV Compaction via Attention Matching for Long-Context LLMs
By
–
"Fast KV Compaction via Attention Matching" The current problem is that long-context LLMs would choke on KV cache memory, so systems either drop tokens or write lossy summaries. As a solution, the paper shows you can shrink an entire prefix to ~50x fewer KV slots by solving
-
Unified Latents: DeepMind’s Automatic Latent Information Training
By
–
now trending on alphaXiv "Unified Latents (UL): How to train your latents" in latent diffusion you’re choosing how much information the latent should carry, and there’s no automatic “right” setting. But in this DeepMind paper, they showed that you don't have to guess
-
Meta’s Sphere Encoder Revolutionizes Image Generation Speed
By
–
"Image Generation with a Sphere Encoder" This paper from Meta shows that you can skip diffusion sampling by training a Sphere Encoder whose latents are forced to uniformly fill a hypersphere. So generation now becomes: sample one random point -> decode This cuts hundreds of
-
Marimo Notebook Showcase from fishnets88
By
–
Checkout the @marimo_io notebook from @fishnets88 here! https://
molab.marimo.io/notebooks/nb_v
DiGF82Fcwrtx1sVkFAZc2
… -
MaxRL: One-Line GRPO Improvement for Better Training Scaling
By
–
MaxRL is a slick one line change to GRPO that optimizes a maximum likelihood objective instead of expected reward
— alphaXiv (@askalphaxiv) 19 février 2026
Scaling advantages with 1/μ as opposed to traditional REINFORCE or GRPO’s 1/σ allows the model to make better progress from low initial pass rates and better… pic.twitter.com/sJCAlsezeqMaxRL is a slick one line change to GRPO that optimizes a maximum likelihood objective instead of expected reward Scaling advantages with 1/μ as opposed to traditional REINFORCE or GRPO’s 1/σ allows the model to make better progress from low initial pass rates and better