Towards Large Reasoning Models: A Survey of Reinforced Reasoning with Large Language Models

TL;DR

A survey paper systematically organizing methodologies for improving LLM reasoning through reinforcement learning.

Problem

Existing LLMs rely on simple autoregressive generation, showing limitations in complex reasoning tasks. Methods are needed to generate intermediate thought processes like humans, learn through trial and error, and invest more computation at test time to improve accuracy.

Approach

Focusing on the learning-to-reason paradigm using reinforcement learning, the paper analyzes three key technical axes: automated data construction, learning-based reasoning techniques (tree search, reflective thinking, etc.), and test-time scaling. The OpenAI o1 series is presented as a representative case.

Results & Contribution

Systematically organizes the overall flow of LLM reasoning research and presents the development direction toward Large Reasoning Models. Emphasizes that the combination of automated data generation, RL-based learning, and test-time scaling is key to improving reasoning performance.