Abstract

In the second journal club, we will first continue with the discussion on mechanistic interpretability by diving into a specific example of mechanistic interpretability work from the following paper: ‘When Models Manipulate Manifolds: The Geometry of a Counting Task’ Anthropic 2025: https://transformer-circuits.pub/2025/linebreaks/index.html

Then, we will discuss some highly cited ML work on reasoning capabilities in LLMs.

Description: From CoT to R1 - An overview of reasoning in LLMs

Papers: Chain-of-Thought Prompting Elicits Reasoning in Large Language Models DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

Overview:  Reasoning in LLMs have shown to provide significant improvement in model performance. We provide an overview of how reasoning capabilities develop in LLMs and its wider implications

It is not necessary to read the above literature before the session!