Rethinking Reflection in Pre-Training

🤗 Upvotes: 52 | cs.CL, cs.AI

Authors:
Essential AI, :, Darsh J Shah, Peter Rushton, Somanshu Singla, Mohit Parmar, Kurt Smith, Yash Vanjani, Ashish Vaswani, Adarsh Chaluvaraju, Andrew Hojel, Andrew Ma, Anil Thomas, Anthony Polloreno, Ashish Tanwer, Burhan Drak Sibai, Divya S Mansingka, Divya Shivaprasad, Ishaan Shah, Karl Stratos, Khoi Nguyen, Michael Callahan, Michael Pust, Mrinal Iyer, Philip Monk, Platon Mazarakis, Ritvik Kapila, Saurabh Srivastava, Tim Romanski

Title:
Rethinking Reflection in Pre-Training

Arxiv:
http://arxiv.org/abs/2504.04022v1

Abstract:
A language model's ability to reflect on its own reasoning provides a key advantage for solving complex problems. While most recent research has focused on how this ability develops during reinforcement learning, we show that it actually begins to emerge much earlier - during the model's pre-training. To study this, we introduce deliberate errors into chains-of-thought and test whether the model can still arrive at the correct answer by recognizing and correcting these mistakes. By tracking performance across different stages of pre-training, we observe that this self-correcting ability appears early and improves steadily over time. For instance, an OLMo2-7B model pre-trained on 4 trillion tokens displays self-correction on our six self-reflection tasks.