How Large Language Models Reason

Description

While Large Language Models (LLMs) dazzle us with their human-like reasoning—from cracking puzzles to constructing logical arguments—the question remains: "How exactly do they reason?". In this project students will uncover the sophisticated reinforcement learning techniques and other approaches, wich are fueling the reasoning in LLM models. Students will engage critically with core ideas, experiment with prompting methods, explore reinforcement learning from human feedback, and deepen their understanding of both theoretical principles and practical AI innovations driving today's improvements.

Prerequisites

Strong knowledge of Python. Familiarity with basic probability theory.

Some background material

• "Reasoning with Large Language Models: A Survey” (Plaat et al., 2024)" – A taxonomy of prompt-based reasoning methods and open challenges.
• "A Survey of Scaling in Large Language Model Reasoning" (Chen et al., 2025) – Explores how scaling context, steps, and rounds affects reasoning quality.
• "Advancing Reasoning in Large Language Models: Promising Methods and Approaches" (Patil, 2025) – Overview of prompting, neuro‑symbolic methods, RL tweaks, and hallucination mitigation.
• MIT News: "Like human brains, large language models reason about diverse data in a general way:" (Feb 2025) – Insights into how LLMs internally represent and reason about varied inputs.
• ITPro Podcast: "Are reasoning models fundamentally flawed?" – Critical discussion of limitations when LLMs handle complex problems.