Base Models Can Reason By Taking a Cue From Training Data
This writeup gives a high-level overview of our paper and aims to build intuition for what it means for a base language model to reason. We walk through some interesting nuggets from our findings, illustrate them with interactive figures, and discuss implications raised by our results. Details can be found in the paper.
Tl;dr: Two opening tokens can bring a base model’s reasoning performance close to that of its RL-trained counterpart. These token cues come from associations learned during training, and RL makes effective cues more likely. Changing those associations can make even “Chicken” elicit reasoning. This writeup gives a high-level overview of our paper and aims to build intuition for what it means for a base language model to reason. We walk through some interesting nuggets from our findings, illustrate them with interactive figures, and discuss implications raised by our results. Details can be…
saved by
related reading
- behind the scenes of paper writingsophielwang.com
- Base Models Know How to Reason, Thinking Models Learn Whenarxiv.org
- [2201.11903] Chain of Thought Prompting Elicits Reasoning in Large Language Modelsarxiv.org
- Why Does Reinforcement Learning Generalize? A Feature-Level Mechanistic Study of Post-Training in Large Language Modelsarxiv.org
- Language Models can Solve Computer Tasksarxiv.org
- [2511.08579] Training Language Models to Explain Their Own Computationsarxiv.org
- The State of Reinforcement Learning for LLM Reasoningmagazine.sebastianraschka.com
- Training Language Models to Explain Their Own Computationsarxiv.org
- Training Large Language Models to Reason in a Continuous Latent Spacearxiv.org
- [2502.19402] General Reasoning Requires Learning to Reason from the Get-goar5iv.labs.arxiv.org
- [2203.14465] STaR: Bootstrapping Reasoning With Reasoningarxiv.org
- [2506.19733] Breaking Barriers: Do Reinforcement Post Training Gains Transfer To Unseen Domains?arxiv.org