We introduce Phi-4-reasoning, a 14-billion parameter reasoning model that achieves strong performance on complex reasoning tasks. Trained via supervised fine-tuning of Phi-4 on carefully curated set of "teachable" prompts-selected for the right level of complexity and diversity-and reasoning demonstrations generated using o3-mini, Phi-4-reasoning generates detailed reasoning chains that effectively leverage inference-time compute. We further develop Phi-4-reasoning-plus, a variant enhanced through a short phase of outcome-based reinforcement learning that offers higher performance by generating longer reasoning traces. Across a wide range of reasoning tasks, both models outperform significantly larger open-weight models such as DeepSeek-R1-Distill-Llama-70B model and approach the performance levels of full DeepSeek-R1 model. Our comprehensive evaluations span benchmarks in math and scientific reasoning, coding, algorithmic problem solving, planning, and spatial understanding. Interestingly, we observe a non-trivial transfer of improvements to general-purpose benchmarks as well. In this report, we provide insights into our training data, our training methodologies, and our evaluations. We show that the benefit of careful data curation for supervised fine-tuning (SFT) extends to reasoning language models, and can be further amplified by reinforcement learning (RL). Finally, our evaluation points to opportunities for improving how we assess the performance and robustness of reasoning models.
核心贡献 · Key contributions
提出 Phi-4-reasoning，一个 14B 参数模型，通过对经过筛选的可教学提示和 o3-mini 演示进行监督微调，实现强大的推理能力。 Introduces Phi-4-reasoning, a 14B model achieving strong reasoning via SFT on curated teachable prompts and o3-mini demonstrations.
通过基于结果的强化学习增强的 Phi-4-reasoning-plus，生成更长的推理轨迹并实现更高性能。 Phi-4-reasoning-plus enhanced by outcome-based RL yields longer reasoning traces and higher performance.
两个模型在数学和科学任务上超越更大的开放权重模型（如 DeepSeek-R1-Distill-70B），接近完整 DeepSeek-R1 的性能水平。 Both models outperform larger open-weight models like DeepSeek-R1-Distill-70B and approach DeepSeek-R1 on math and science.
显著的改进迁移到通用基准（IFEval、FlenQA、ArenaHard），表明广泛的能力提升。 Significant improvements transfer to general-purpose benchmarks (IFEval, FlenQA, ArenaHard), indicating broad capability gains.
精细的数据筛选用于监督微调，其益处扩展至推理模型，强化学习进一步放大了收益。 Careful data curation for SFT extends benefits to reasoning models, and RL further amplifies gains.
全面评估强调了因小型基准（如 AIME）的高方差而需要更稳健的评估实践。 Comprehensive evaluation highlights need for robust practices due to high variance on small benchmarks like AIME.
局限 · Limitations
小型基准（如仅包含 30 道题的 AIME）上的评估方差较大，导致比较不可靠，需要多次运行。 Evaluation variance on small benchmarks (e.g., AIME with 30 problems) makes comparisons unreliable; many runs needed.
强化学习仅关注数学，在编程、规划和空间任务上的改进不那么明显。 RL focused only on math; improvements less pronounced on coding, planning, and spatial tasks.
强化学习期间响应长度限制在 32K 词元，限制了更长推理可能带来的收益。 Response length capped at 32k tokens during RL, limiting potential gains from longer reasoning.
安全隐患：思维链可能泄露安全性指南，需要进一步研究。 Safety concern: chain-of-thought may reveal safety guidelines, requiring further research.
在生物学和化学上的提升小于数学和物理学，表明推理的迁移不均衡。 Smaller gains in biology and chemistry compared to math and physics, indicating uneven reasoning transfer.
论文章节 · Sections（共 3）
摘要Abstract
95 AIME 2025 并行测试时间算力95 AIME 2025 Parallel Test Time Compute