[论文] RECAST: Learning to Compute the Right Context through Adaptive Evidenc...

论文概要

研究领域: ML 作者: Yilun Hao, Krishna Sayana, Isabella Ye, James S Ren, Sukhdeep Sodhi, Craig Boutilier, Chuchu Fan 发布时间: 2026-10-07 arXiv: 2610.10507

中文摘要

大语言模型越来越多地应用于基于冗长异构信息源的任务。传统的检索增强生成（RAG）依赖固定的相似度检索，而智能体变体虽能调整查询和工具使用，但仍以检索为中心。然而，在许多任务中，解决方案所需的证据并未显式存在于任何单一来源条目中——而是必须通过跨多个条目的过滤、聚合或计算来推导。本工作引入RECAST（通过计算、访问和合成工具路由证据），一个学习框架，将证据构建表述为对异构检索和计算操作的顺序决策过程，使证据能够被主动推导而非仅仅检索。轻量级RouterLM迭代选择并表述原始操作，或为冻结的CompilerLM指定自定义操作以翻译成可执行代码。一旦判断证据充分，RouterLM将接受的证据传递给冻结的AnswerLM生成最终解决方案。我们通过监督微调（SFT）加组相对策略优化（GRPO）训练RouterLM。在六个异构基准家族中，RECAST平均成功率为75.6%，比最强的大模型基线高出15.9%。此外，训练使Qwen3.5-9B RouterLM比无需训练的Gemini 3.5 Flash RouterLM高出5.0%。在三个留出基准上，RECAST比最强基线平均提升15.0%，展示了跨任务和异构源表示的强零样本泛化能力。

原文摘要

Large language models are increasingly applied to tasks grounded in long, heterogeneous information sources. Conventional Retrieval-Augmented Generation (RAG) relies on fixed similarity-based retrieval, while agentic variants adapt queries and tool use but remain largely retrieval-centric. However, in many tasks, the evidence required for a solution is not explicitly present in any single source item. Instead, it must be derived through filtering, aggregation, or computation across multiple source items. In this work, we introduce RECAST (Routing Evidence through Computation, Access, and Synthesized Tools), a learned framework that formulates evidence construction as a sequential decision process over heterogeneous retrieval and computation operations, allowing evidence to be actively deri...

#论文 #arXiv #ML #小凯