We introduce Seed1.5-Thinking, capable of reasoning through thinking before responding, resulting in improved performance on a wide range of benchmarks. Seed1.5-Thinking achieves 86.7 on AIME 2024, 55.0 on Codeforces and 77.3 on GPQA, demonstrating excellent reasoning abilities in STEM and coding. Beyond reasoning tasks, the method demonstrates notable generalization across diverse domains. For instance, it surpasses DeepSeek R1 by 8% in win rate on non-reasoning tasks, indicating its broader applicability. Compared to other state-of-the-art reasoning models, Seed1.5-Thinking is a Mixture-of-Experts (MoE) model with a relatively small size, featuring 20B activated and 200B total parameters. As part of our effort to assess generalized reasoning, we develop two internal benchmarks, BeyondAIME and Codeforces, both of which will be publicly released to support future research. Model trial link: https://www.volcengine.com/experience/ark.
核心贡献 · Key contributions
Seed1.5-Thinking 在 AIME 2024 上达到 86.7，在 Codeforces 上达到 55.0，在 GPQA 上达到 77.3，匹配或接近 o3 级别性能。 Seed1.5-Thinking achieves 86.7 on AIME 2024, 55.0 on Codeforces, and 77.3 on GPQA, matching or approaching o3-level performance.
该模型在非推理任务上的胜率超过 DeepSeek R1 8%，展现出广泛的泛化能力。 The model surpasses DeepSeek R1 by 8% in win rate on non-reasoning tasks, demonstrating broad generalization.
提出了 VAPO 和 DAPO 两种强化学习框架，用于推理模型的稳定训练。 Introduces VAPO and DAPO, two RL frameworks for stable training of reasoning models.
开发了 Seed-Verifier 和 Seed-Thinking-Verifier，用于可验证任务中的精确奖励建模。 Develops Seed-Verifier and Seed-Thinking-Verifier for accurate reward modeling in verifiable tasks.
构建了包含 100 道难题的 BeyondAIME 基准，以更好区分推理模型。 Constructs BeyondAIME benchmark with 100 challenging math problems to better differentiate reasoning models.
提出了流式推出系统 (SRS) 用于高效的长思维链生成，实现 3 倍迭代加速。 Proposes Streaming Rollout System (SRS) for efficient long-CoT generation, achieving 3× faster iteration.
局限 · Limitations
Seed1.5-Thinking 在 AIME 2025 和 BeyondAIME 基准上落后于 o3 级别模型。 Seed1.5-Thinking lags behind o3-level models on AIME 2025 and BeyondAIME benchmarks.
该模型在面向记忆的基准 SimpleQA 上表现不佳。 The model shows less impressive results on SimpleQA, a memory-oriented benchmark.
强化学习训练仍不稳定，需要仔细调整超参数和数据分布。 RL training remains unstable and requires careful tuning of hyperparameters and data distribution.
基于验证器的奖励模型在复杂边缘案例上仍可能遭受奖励破解。 The verifier-based reward model may still suffer from reward hacking on complex corner cases.
对非可验证任务的泛化依赖于人类偏好数据，可能引入偏差。 Generalization to non-verifiable tasks relies on human preference data, which may introduce bias.
论文章节 · Sections（共 20）
摘要Abstract
1 引言1 Introduction
2.1 强化学习训练数据2.1 RL Training Data
2.2 高等数学基准2.2 Advanced Math Benchmark
3 奖励建模3 Reward Modeling
3.1 可验证问题的奖励建模3.1 Reward Modeling for Verifiable Problems
3.2 不可验证问题的奖励建模3.2 Reward Modeling for Non-verifiable Problems