What if language models could learn to "think harder" only when they need to—allocating deep computation to challenging tokens while breezing through simple ones? We're excited to have Reza tomorrow in our AI4Science community giving a talk on Mixture-of-Recursions!
RESEARCH
-

OpenAI Horizon Alpha Expected Release Date August 2026
By
–
Previously when OpenAI tested models (Quasar Alpha & Optimus Alpha) it took 4-12 days to release the model afterwards. So if we extrapolate this for Horizon Alpha, the dates are 3-11 August
-
LoRA Compatibility and Retraining Strategies for AI Models
By
–
they are compatible with LoRAs. that doesn’t mean that LoRAs are pluggable *as-is*. While we’ve found that LoRAs sometimes work well or even better plugging them as is, we’ve found better results by retraining them for the most part. The important part is the compatibility.
-
Weaver Combines Weak Verifiers to Boost LLM Output Quality
By
–
New from Snorkel AI: Weaver combines weak verifiers to boost LLM output quality: no fine-tuning needed. +14.5% accuracy 98.2% of gains, 99.97% less compute (via distillation) Outperforms naive voting, rivals GPT-4o Read the full paper
-
KL Divergence Measurement on Reinforcement Learning Outputs
By
–
no, because it's KL measured on the RL outputs
-
LLM Safety: Why Current Development Approaches Are Insufficient
By
–
That's the whole problem. We are developing software against LLMs in the same way we use SQL: we write our own instructions as a programmer, then we combine them with input from our users With SQL we can do that safely. With LLMs we can't
-
DreamScene: 3D Gaussian Text-to-3D Scene Generation
By
–
DreamScene
— AK (@_akhaliq) 31 juillet 2025
3D Gaussian-based End-to-end Text-to-3D Scene Generation pic.twitter.com/9fEDUEj0uRDreamScene 3D Gaussian-based End-to-end Text-to-3D Scene Generation
-
Scaling Prompt Optimization: Systems, Theory, and Benchmarks Needed
By
–
probably 10x more people should be working on prompt optimization systems (we need a vLLM for promptopt), theory, new techniques, benchmarks. the whole kit and caboodle
-
Building AutoRL: A Great Reinforcement Learning Achievement
By
–
It's really great. I built a lot of AutoRl w/ it 🙂
-
Natural Birth Outperforms Multibillion-Dollar LLM Models
By
–
You can make a baby for free, and the result is still better than a multibillion-dollar LLM.
