DeepSeek-AI (2025) — ARXIV-2501.12948
Key takeaways
- The claim: reasoning can be incentivised rather than demonstrated. Pure reinforcement learning against verifiable outcomes — mathematics, competitive programming, STEM — develops the capability without any human-annotated reasoning trajectories, which is what every prior recipe depended on and what limited them.
- What emerges under that pressure is behavioural rather than merely numerical: self-reflection, verification of intermediate results, and dynamic strategy adaptation appear without being specified.
- The resulting model beats counterparts trained conventionally on human demonstrations at the same verifiable tasks.
- The emergent patterns transfer: they can be harnessed systematically to improve smaller models, which is the distillation half that put R1-style traces into everyone else’s training mixtures.
Standing in the anthology
The code SOTA-130 has been naming without one. That practice — skip the reasoning SFT stage and run RLVR directly on the base model — is called the “R1-Zero” style in its own body, with the note that “that work is not in the record”. It is now.
This is the origin of a great deal the record already holds downstream. LIT-119’s tiny reasoning models are pretrained on synthetic traces of this kind; LIT-123’s looping analysis is about models distilled from teachers like this one; LIT-130’s RL-Zero track is Ai2 reproducing the pathway as an open benchmark. Filing it late is the same pattern as the Muon blog post — the record accumulated the descendants first.
LIT-167 is the critical reading of it, and belongs next to it: the “Aha moment” it reports may be partly a property of the base model rather than of the RL.