A breakdown of Microsoft’s SkillOpt paper, which applies gradient-descent-like optimization to continuously improve AI agent skills.
Microsoft’s SkillOpt: 2x Accuracy Without Touching the Model Read More »
In this post, we break down DeepSeek-V4, a new LLM from DeepSeek designed for highly efficient reasoning over million-token contexts
DeepSeek-V4 Explained: The End of Standard Attention in LLMs? Read More »
Google’s Nested Learning paper and Hope model explained: a new approach to continual learning in LLMs that addresses catastrophic forgetting.
GDPO is NVIDIA’s solution to GRPO’s limitations in multi-reward RL for large language models. We break down the paper in this post.
GDPO Explained: How NVIDIA Fixes GRPO for Multi-Reward LLM Reinforcement Learning Read More »
Manifold-Constrained Hyper-Connections (mHC) explained: How DeepSeek rewires residual connections in LLMs for next-gen AI
DeepSeek’s mHC Explained: Manifold-Constrained Hyper-Connections Read More »
Discover how reinforcement learning enables hierarchical reasoning in LLMs and how HICRA improves on top of GRPO.
Emergent Hierarchical Reasoning in LLMs Through Reinforcement Learning Read More »
In this post, we break down the TRM paper, a simpler version of the HRM, that beats HRM and top reasoning LLMs with a tiny 7M params model.
Less Is More: Tiny Recursive Model (TRM) Paper Explained Read More »
In this post we break down the Hierarchical Reasoning Model (HRM), a new model that rivals top LLMs on reasoning benchmarks with only 27M params!
In this post we break down Microsoft’s Reinforcement Pre-Training, which scales up reinforcement learninng with next-token reasoning
Microsoft’s Reinforcement Pre-Training (RPT) – A New Direction in LLM Training? Read More »
In this post we explain the Darwin Gödel Machine, a novel method for self-improving AI agents by Sakana AI
