https://www.vishalbhoriwal.me/papers/Recent content in Papers on Vishal BhoriwalHugoenWed, 22 Jan 2025 00:00:00 +0000https://www.vishalbhoriwal.me/papers/deepseek-r1/Wed, 22 Jan 2025 00:00:00 +0000https://www.vishalbhoriwal.me/papers/deepseek-r1/<p><strong>Link:</strong> <a href="https://arxiv.org/abs/2501.12948" target="_blank" rel="noopener noreferrer">arxiv.org/abs/2501.12948</a> </p> <hr> <h2 id="abstract">Abstract</h2> <p>DeepSeek-R1 shows that strong reasoning capabilities in large language models can emerge from pure reinforcement learning — no human-labeled reasoning traces required. The resulting model matches OpenAI&rsquo;s o1 on a range of benchmarks, and the learned reasoning patterns can be distilled into much smaller models that still punch well above their weight.</p> <hr> <h2 id="background">Background</h2> <p>Getting LLMs to reason reliably has mostly meant one thing: more human supervision. Chain-of-thought prompting helps, but it requires annotated examples. RLHF aligns models to human preferences, but again depends on human feedback. OpenAI&rsquo;s o1 demonstrated that models could &ldquo;think&rdquo; before answering using extended reasoning chains — but the training details were never released.</p>