D-REAL：用无偏梯度优化投机解码接受长度
D-REAL: Acceptance-Length Optimization with Unbiased Gradients for Speculative Decoding
投机解码的 drafter 训练一直缺一个对齐接受长度的目标函数，这篇论文把它做成了无偏梯度估计器。在 Qwen3-8B 上十个基准验证，一个 epoch 就有 9.4% 吞吐提升。做推理优化的工程师值得照着复现。
D-REAL（Drafter Reinforcement of Expected Acceptance Length）是一种后训练方法，直接以期望接受长度为优化目标，解决传统 token 级训练目标无法直接提升投机解码连续接受 token 数的问题。方法按前缀存活概率对位置级 draft-target 分布重叠度加权，沿 draft 采样前缀忠实刻画验证对早期提议的序列依赖，并推导出随机验证下单轮目标的无偏梯度估计器。
实验以 Qwen3-8B 为目标模型，在 DSpark 和 DFlash2 两种架构上跨十个基准评测。在基线平均接受长度已进入平台期的情况下，仅一个 epoch 的 D-REAL 后训练即可将基准平均接受长度提升最多 0.23 个 token，在贪婪采样（T=0.0）和随机采样（T=1.0）下均成立，对应端到端服务吞吐最高提升 9.4%。
对部署大模型推理服务的团队来说，这说明 drafter 模型的后训练仍有明确的优化空间，且目标函数可以直接对齐服务吞吐这一业务指标。
D-REAL: Acceptance-Length Optimization with Unbiased Gradients for Speculative Decoding
Speculative decoding efficiency depends on the number of consecutively accepted draft tokens, yet token-level training objectives do not directly optimize this quantity. We introduce D-REAL (Drafter Reinforcement of Expected Acceptance Length), a post-training method that directly optimizes expected acceptance length for a fixed decoding context. D-REAL weights position-wise draft-target distributional overlap by prefix survival probabilities along a draft-sampled prefix, faithfully capturing the sequential dependence of verification on earlier proposals. By accounting for the parameter dependence of the proposal distribution, we derive an unbiased gradient estimator for this single-round objective under stochastic verification. We evaluate D-REAL on DSpark and DFlash2 with Qwen3-8B as the target model across ten benchmarks. Starting from baselines whose mean acceptance length has plateaued, one epoch of D-REAL post-training consistently improves the benchmark-averaged mean acceptance length by up to 0.23 tokens across both architectures under greedy (T = 0.0) and stochastic (T = 1.0) sampling, translating into up to a 9.4% end-to-end serving throughput improvement.
本条情报归属 AITOP「论文」核心追踪栏目。在跨区域产业观察坐标中，该动态直接关联「AI 开发者工具链、Coding Agent 与工程化实践」研究主线，反映出当前技术演进与商业落地的最新风向。涉及核心实体与技术议题包括：D-REAL、speculative decoding、Qwen3-8B、推理优化，读者可通过下方标签进入对应专题页查阅历史脉络与关联对比。
信源核验说明：本文原始线索采编自「arXiv cs.AI」（@Meng Wei, Yefei Chen, Antong Li, Yanan Gao, Xingliang Shi, Xin Gao, Zhiling Ye, Yun Yue, Chunxiao Guo, Peng Wei），经 AITOP 多维评分管线结构化提炼。