arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2604.07147cs.CLcs.AIcs.LG
收藏

动态上下文演化用于可扩展的合成数据生成

Dynamic Context Evolution for Scalable Synthetic Data Generation

Honda Research Institute, USA, Inc.(本田研究所美国公司)

Ryan Lingo, Rajeev Chhajer

AI总结:

本文提出动态上下文演化(DCE)框架,通过尾部采样、语义记忆和自适应提示演化机制,有效缓解跨批次模式崩溃问题,提升生成多样性与概念结构 richness。

摘要:

大型语言模型在独立提示下多次批量生成时会产生重复输出,我们称之为跨批次模式崩溃:当语言模型在没有访问先前生成的情况下被反复提示时,输出多样性逐渐丧失。从业者长期以来通过随意去重和种子旋转来缓解这一问题,但没有系统框架存在。我们引入动态上下文演化(DCE),包含三种机制:(1) 显式尾部采样(模型标记每个想法并猜测其明显性,明显的想法被丢弃),通过模型自我评估过滤高概率候选;(2) 语义记忆,维护一个持久的嵌入索引以在不同批次中拒绝近似重复;(3) 自适应提示演化,使用记忆状态和旋转多样性策略重建生成提示。在三个领域(可持续包装概念、教育考试问题和创意写作提示)和两种模型家族(gpt-5-mini和claude-haiku-4-5)的实验中,组件消融显示DCE在0.0±0.0%崩溃率(对比朴素提示的5.6±2.0%)和每种子17-18个HDBSCAN簇(对比朴素的波动2-17)方面表现优异。这些结果通过独立嵌入模型(all-MiniLM-L6-v2)验证,并在VTS阈值tau和去重阈值delta的敏感度扫描中保持稳定。去重和提示演化单独不足但联合有效,使用标准API调用约每1,000候选0.50美元,无需微调或定制架构。

英文摘要:

Large language models produce repetitive output when prompted independently across many batches, a phenomenon we term cross-batch mode collapse: the progressive loss of output diversity when a language model is prompted repeatedly without access to its prior generations. Practitioners have long mitigated this with ad hoc deduplication and seed rotation, but no principled framework exists. We introduce Dynamic Context Evolution (DCE), comprising three mechanisms: (1) verbalized tail sampling (the model labels each idea with a guess about how obvious it is, and obvious ideas are discarded), which filters high-probability candidates via model self-assessment; (2) semantic memory, which maintains a persistent embedding index to reject near-duplicates across batches; and (3) adaptive prompt evolution, which reconstructs the generation prompt each batch using memory state and rotating diversity strategies. In experiments across three domains (sustainable packaging concepts, educational exam questions, and creative writing prompts) and two model families (gpt-5-mini and claude-haiku-4-5), a component ablation across 2-3 random seeds per method shows that DCE achieves 0.0 +/- 0.0% collapse versus 5.6 +/- 2.0% for naive prompting, while producing 17-18 HDBSCAN clusters per seed versus naive's volatile 2-17, indicating reliably richer conceptual structure. These results are validated with an independent embedding model (all-MiniLM-L6-v2) and hold across sensitivity sweeps of the VTS threshold tau and dedup threshold delta. Deduplication and prompt evolution are individually insufficient but jointly effective, at approximately $0.50 per 1,000 candidates using only standard API calls, with no fine-tuning or custom architectures required.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑