DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

Reasoning ability in LLMs has depended on large volumes of human-annotated chains of thought, which does not scale and caps the model's ceiling on hard problems.

Pure reinforcement learning with verifiable rewards (GRPO) - no human reasoning traces - to incentivize emergent self-reflection, verification and strategy adaptation. Two mechanisms carry the distillation story: (1) the emergent reasoning patterns in R1-Zero come from RL alone, with no SFT cold start, which shows reasoning can be elicited rather than taught; (2) those patterns are then explicitly harvested to distill smaller models. The 1.5B, 7B, 14B, 32B and 70B DeepSeek-R1-Distill models are produced by fine-tuning open base models (Qwen and LLaMA) on ~800k R1-generated samples - NOT by training the base models on the reasoning data themselves.

Paper's own claim (abstract): the RL-trained model achieves superior performance on verifiable tasks (mathematics, coding competitions, STEM) 'surpassing its counterparts trained via conventional supervised learning on human demonstrations', and the emergent reasoning patterns 'can be systematically harnessed to guide and enhance the reasoning capabilities of smaller models'. No single inference speedup number appears in the abstract - the pay-off is that a 1.5B-70B distilled model reaches reasoning quality that previously required a far larger model, which is the real inference-cost reduction.

SLUG RENAMED 2026-10-04 (parent agent): was 'deepseek-r1', colliding with the models record of the same id, which made every bare-slug cross-reference ambiguous. Inbound refs (['family']) all mean the models record, so none needed repointing; papers are cited via `sources`, never by bare slug. arXiv ID 2501.12948 verified via /abs/. Venue: the /abs/ page carries journal-ref 'Nature volume 645, pages 633-638 (2025)' - this paper was PUBLISHED IN NATURE, which is not in this schema's venue enum, so venue='other' and Nature is named here. Semantic Scholar independently returns Nature / 2025. Year = 2025 (v1 2025-01-22). adoption=in-production: the R1-Distill checkpoints are among the most downloaded open weights in existence and the repo already has four hf-eco source records for them. arXiv ID and title verified by fetching https://arxiv.org/abs/2501.12948 and reading the citation_title meta tag on 2026-10-03. Venue and year from the Semantic Scholar graph API batch endpoint (paper/arXiv:2501.12948) unless the record's notes say otherwise. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'main'. Peer-reviewed venue confirmed as Nature 2025 (journal publication; venue enum has no member). This is an additive track annotation; `venue` is unchanged at 'other' and no status change is implied by the track alone. VENUE RECLASSIFIED 2026-10-10 (crawler/w12-venue-rest): venue 'other' -> 'nature', venue_kind 'journal'. Confirmed by Crossref DOI 10.1038/s41586-025-09422-z / Nature article page ('DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning', Nature volume 645, pages 633-638 (2025)) and agreeing source record notes.