Reasoning Statism
From the AI Research Compass wiki
Reasoning Statism: Long chains of thought, trained with reinforcement learning on verifiable problems, are the main road to stronger intelligence.
Beliefs
Reasoning Statism holds that intelligence is not just what a model knows but how long and how carefully it can think. Give the model room to write out intermediate steps, reward it when the final answer checks out, and it learns to plan, backtrack, verify and try alternatives on its own. The state program is simple: build ever larger pools of math, code and logic problems with checkable answers, and let reinforcement learning grow ever longer, ever better chains of thought.
It sits in the top-left corner, at the Language edge. The language lean is natural, since the reasoning happens in written tokens. The Monolith lean reflects that the reasoning is trained into one model’s weights rather than orchestrated by an external search or agent framework. Its paradigm level is Mainstream: the approach quickly became part of how frontier models are built, though its limits are still being probed.
Arguments
- Prompting models to produce intermediate steps substantially improves performance on arithmetic, commonsense and symbolic reasoning tasks.
- Reinforcement learning with verifiable rewards has produced models whose reasoning behaviors, such as self-verification and backtracking, emerged without step-by-step supervision.
- Process supervision of individual steps can outperform supervising only final answers on hard math problems.
Critiques
- Written chains of thought are not guaranteed to reflect the computation that actually produced the answer.
- Gains concentrate in domains with checkable answers; open-ended, physical or social problems are harder to reward.
- Long reasoning traces are costly at inference time, and rivals note that some problems need tools or perception rather than more thinking.
Representative works
- Chain-of-Thought Prompting Elicits Reasoning in Large Language Models2022
- Let's Verify Step by Step2023
- DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning2025
Relationships
Friends
How to get it
Land on this cell in the full or pro test: x ≈ −1.00, y ≈ +0.88, paradigm ≈ −0.54.