the open
    research
    Contribute
    Back to the reading lists

    Language, speech & LLMs· 2025

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

    DeepSeek-AI (first author) · 2025

    Read the original

    DOI 10.1038/s41586-025-09422-zarXiv 2501.12948OpenAlex W4414281281

    On the reading lists

    Why it matters

    Demonstrated powerful emergent reasoning through reinforcement learning with verifiable rewards.

    Condensed from the expert lists' notes.

    Contested

    The paper's claim that reinforcement learning incentivises new reasoning ability is disputed. A NeurIPS 2025 best-paper runner-up, “Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?”, reports that current RL with verifiable rewards does not elicit fundamentally new reasoning patterns.

    Source: NeurIPS 2025 best paper awards announcement (research report 02, C-03)

    Reading it

    Difficulty: Intermediate

    Read this first: Chain-of-Thought Prompting; policy gradients.

    The record

    First released
    2025 (from arXiv 2501.12948)
    Citations
    5,325 citations, as counted by Papers With Code (CC BY-SA 4.0)
    Named by
    both expert lists
    Also relevant to
    Reinforcement learning
    Score
    0.609 · computed rank 141 · consensus 0.64, must-read 1, citations percentile 0.906 (2020s · language), awards 0, existing lists 0. Method
    Versions merged
    DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning (editor)

    Editor's notes

    • Merged “DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning” into this entry: The 2025 arXiv report and the later Nature paper describe the same model and method; one entry, dated from the arXiv release. (AI draft (Claude), not yet reviewed, 2026-10-29)
    • Readability check: pass: Readable account of RL for reasoning; read with its contested note. (AI draft (Claude), not yet reviewed, 2026-10-29)

    Data sources

    Catalogue data from OpenAlex (CC0), Crossref (CC0 metadata; no abstracts are shown), arXiv, and Papers With Code (CC BY-SA 4.0). Each count names its source. See the method.