Best of Deep Learning — March 2025

1. 1
   Article

   Sebastian Raschka·1y

   The State of LLM Reasoning Models

   The post explores recent research advancements in reasoning-optimized large language models (LLMs), focusing on inference-time compute scaling methods. It discusses how various techniques, such as chain-of-thought reasoning and test-time preference optimization, improve the reasoning abilities of LLMs without altering underlying model weights. The article highlights the importance of increasing computational resources during inference to enhance performance, making even smaller models more capable. It also touches on other methods like reinforcement learning and supervised fine-tuning that contribute to improved reasoning in LLMs.

   53
2. 2
   Article

   Sebastian Raschka·1y

   First Look at Reasoning From Scratch: Chapter 1

   Sebastian Raschka shares the first chapter of his new book focused on reasoning in large language models (LLMs). This chapter defines reasoning within LLMs, differentiates it from pattern matching, and explains key methods to enhance LLM reasoning abilities. It also covers basic training stages of LLMs and introduces reasoning methodologies like inference-time scaling and reinforcement learning.

   16
3. 3
   Video

   Artem Kirsanov·1y

   How Brains & Machines Master Probability

   Brains and AI systems both face the challenge of reasoning under uncertainty with incomplete data. Variational inference is a mathematical tool used to create efficient models from limited clues. By using concepts like the evidence lower bound (ELBO) and latent variables, both natural and artificial systems can effectively manage high-dimensional data. Techniques such as important sampling and Jensen's inequality play crucial roles in this process, enabling more accurate and computationally feasible models.

   10