SYNTHESIS NOTE
TopicsReinforcement Learningthis note

Can chain-of-thought reasoning be learned during pretraining itself?

Explores whether reasoning emerges more effectively when models treat thinking as an exploratory action during next-token prediction, rather than only after pretraining through reinforcement learning.

Synthesis note · 2026-02-22 · sourced from Reinforcement Learning
How should we spend compute at inference time?

The dominant paradigm separates pretraining (next-token prediction) from reasoning (RL post-training with verifiable rewards). RLP challenges this by bringing RL's core mechanism — exploration — into pretraining itself. The key idea: treat chain-of-thought as an exploratory action taken before predicting each next token, with reward computed from the information gain that thought provides.

The reward signal is elegant: measure the increase in log-likelihood of the observed token when conditioning on both context and a sampled reasoning chain, compared to context alone. This is verifier-free (no task-specific checkers needed), dense (assigns credit at every position), and applicable to ordinary web-scale text during pretraining. The model learns to think for itself before predicting what comes next, teaching independent thinking behavior earlier in training.

Results compound: pretraining with RLP on Qwen3-1.7B lifts the average across eight math-and-science benchmarks by 19%. With identical post-training, gains compound further. Applied to Nemotron-Nano-12B, overall average increases from 42.81% to 61.32%. The largest improvements are on reasoning-heavy tasks like AIME25 and MMLU-Pro.

This is significant because it reframes when reasoning should be learned. Since Do base models already contain hidden reasoning ability?, RLP suggests that pretraining itself can plant stronger reasoning seeds. And since Does RL teach reasoning or just when to use it?, RLP may teach the "how" during pretraining, leaving post-training to teach the "when" — a cleaner division of labor.

Unlike prior reinforcement pretraining (RPT) which uses sparse binary rewards and relies on proxy-model filtering, RLP provides continuous improvement signals at every position and trains on full documents, eliminating the need to preselect high-entropy tokens.

Inquiring lines that read this note 80

This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.

Do base models contain latent reasoning that training can unlock? Can prompting inject entirely new knowledge into language models? What actually drives chain-of-thought reasoning improvements in language models? How do training data properties shape reasoning capability development? Why do correct reasoning traces tend to be shorter than incorrect ones? What determines success in training models on multiple tasks? What pretraining choices and baseline capability constrain reinforcement learning gains? How does AI adoption affect human skill development and labor equality? Does reinforcement learning teach reasoning or just when to reason? What properties determine whether reward signals teach genuine reasoning? How does policy entropy collapse constrain reasoning-focused reinforcement learning? What capability tradeoffs emerge when scaling model reasoning abilities? What constrains reinforcement learning's ability to expand model reasoning? Why does supervised fine-tuning improve accuracy while degrading reasoning quality? Can inference-time compute substitute for scaling up model parameters? Can next-token prediction alone produce genuine language understanding? How does latent reasoning compare to verbalized chain-of-thought? How do self-generated feedback mechanisms enable effective model learning? Why do continual learning scenarios trigger catastrophic forgetting and interference? How do neural networks separate factual knowledge from reasoning abilities? Can self-supervised signals enable process supervision without human annotation? How should iterative research systems allocate reasoning per search step? Does fine-tuning modify underlying model capabilities or only behavioral outputs? Does recurrence enable reasoning capabilities that fixed-depth transformers cannot achieve?

Related concepts in this collection 4

This note in its neighbourhood — explore the map, then jump to a related concept in the list below.

Concept map
17 direct connections · 122 in 2-hop network ·medium cluster Open in graph ↗

Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph

your link semantically near linked from elsewhere

Related papers in this collection 8

Papers most semantically related to this note, ranked by cosine similarity in the embedding space.

Original note title

chain-of-thought as pretraining exploratory action with information-gain reward bridges next-token prediction and reasoning emergence