Seed1.5-Thinking 技术报告解读:VAPO 与 DAPO 强化学习训练推理模型

解读字节跳动 Seed1.5-Thinking 的强化学习方案:VAPO 与 DAPO 两类算法提升推理模型的训练稳定性、带推理路径的验证器把准确率推到 99.3% 以消除奖励作弊,以及 Decoupled-GAE 与长度自适应优化。

Seed1.5-Thinking —技术报告摘要

论文:Seed1.5-Thinking: Advancing Superb Reasoning Models with Reinforcement Learning (arXiv:2504.13914)
团队:ByteDance Seed
发布时间:2025.04
页数:19页


1. 预训练 Recipe

项目详情
基座模型MoE,200B总参/ 20B激活
预训练细节原文未详述(focus在后训练)

2. 后训练 Recipe (SFT / Cold Start)

训练策略

项目详情
Cold Start流程①人工精标少量高质量长CoT样本 → ②训练初版reasoning model → ③用Seed-Verifier做rejection sampling生成更多CoT轨迹
数据特点以CoT数据为主;过多non-CoT SFT数据会显著降低模型探索能力
跨域泛化数学Cold Start流程可泛化到coding、logic puzzle、creative writing
训练配置截断32K tokens,2epochs,cosine decay lr 2e-5→2e-6

训练数据

项目详情
SFT数据高质量长CoT轨迹(rejection sampling生成)
来源域STEM、code、logic reasoning、non-reasoning

3. RL Recipe

训练策略

项目详情
统一RL框架融合verifiable data + general data + hybrid data(verifier+RM双评分)
核心算法VAPO(actor-critic SOTA)+ DAPO(policy-gradient without critic SOTA)
VAPO关键技术①Value-Pretraining:用π_sft采样+Monte-Carlo return初始化value model ②Decoupled-GAE:λ_value=1.0, λ_policy=0.95独立平衡bias/variance ③Length-adaptive GAE:λ_policy = 1 - α/(1/l),均匀化长短序列的TD error分布
数据类别①Verifiable(verifier反馈)②General(reward model打分)③Hybrid(verifier+RM双信号)
RL数据域STEM、code、logic reasoning、non-reasoning(creative writing/dialogue)
关键发现Math data有强跨域泛化能力;logic reasoning数据显著提升ARC-AGI

Reward Modeling

项目详情
Verifiable问题两级verifier:Seed-Verifier(LLM +人工原则,82.7%准确率)→ Seed-Thinking-Verifier(带reasoning path的verifier,99.3%准确率)
Seed-Thinking-Verifier作为verifiable task与数学推理联合优化;生成详细reasoning判断;解决reward hacking/prediction uncertainty/corner cases
Non-verifiable问题Pairwise generative reward model:比较两个response优劣,用P(“YES”/“NO”)作为reward score;缓解outlier score,稳定RL训练

Infra设计

项目详情
架构Decoupled streaming rollout:异步处理partial trajectory generations through prioritized sample pools
加速3× faster iteration cycles vs synchronous frameworks
稳定性Mixed-precision training + automatic fault recovery
规模支持大规模RL runs(具体GPU数未披露)

4. 关键亮点

  1. VAPO + DAPO:两种RL范式的SOTA算法,解决reasoning model训练不稳定问题
  2. Seed-Thinking-Verifier:带reasoning path的verifier,99.3%准确率,消除reward hacking
  3. Decoupled-GAE + Length-adaptive GAE:稳定长CoT序列的RL训练
  4. Pairwise Generative RM:non-verifiable问题的reward稳定方案
  5. Cold Start + Rejection Sampling:可跨域泛化的高质量CoT数据生成pipeline
  6. 20B激活模型超DeepSeek-R1:AIME 86.7, non-reasoning win rate +8%
本文结束 感谢您的阅读