Originalarbeiterreichbar
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
Deep
geprüft 10.11.2026
Worauf sich diese Seite beruft, wörtlich, abgerufen am 04.09.2026:
Here we show that the reasoning abilities of LLMs can be incentivized through pure reinforcement learning (RL), obviating the need for human-labeled reasoning trajectories.
bestätigt 10.11.2026surpassing its counterparts trained via conventional supervised learning on human demonstrations
bestätigt 10.11.2026Moreover, the emergent reasoning patterns exhibited by these large-scale models can be systematically harnessed to guide and enhance the reasoning capabilities of smaller models.
bestätigt 10.11.2026using a curated dataset comprising 800,000 samples generated with Deep
bestätigt 10.11.2026Seek-R1 In total, we collect about 600k reasoning-related training samples.
bestätigt 10.11.2026In the end, we collected a total of approximately 200k training samples that are unrelated to reasoning.
bestätigt 10.11.2026For distilled models, we apply only SFT and do not include an RL stage, even though incorporating RL could substantially boost model performance.
bestätigt 10.11.2026Notably, it is remarkable that a model with only 1.5 billion parameters achieves superior performance compared to the best closed-source models.
bestätigt 10.11.2026Deep
bestätigt 10.11.2026Seek-R1-Distill-Qwen-32B, which is distilled from Deep Seek-R1, performs significantly better than Qwen2.5-32B-Zero across all benchmarks. First, distilling more powerful models into smaller ones yields excellent results
bestätigt 10.11.2026advancing beyond the boundaries of human intelligence may still require more powerful base models and larger-scale reinforcement learning
bestätigt 10.11.2026Besides, a public model is also vulnerable to further fine-tuning that could compromise inherent safety protections.
bestätigt 10.11.2026generate reasoning trajectories by performing rejection sampling from the checkpoint of the first-stage RL training
bestätigt 10.11.2026we adopt the Deep
bestätigt 10.11.2026Seek-V3 pipeline and reuse portions of the SFT dataset of Deep Seek-V3 to surpass non-reasoning baselines on mathematical benchmarks
bestätigt 10.11.2026
KI-GeschichteWas offene Gewichte wert sind, wenn sie 1,56 Terabyte wiegenGroup Relative Policy Optimization im GlossarReasoning-Modell im GlossarReinforcement Learning im Glossar02 Maschinelles Lernen03 Wie ein Modell von einem anderen lernt