RECAST: Active Evidence Construction via Adaptive Routing and Computation
RECAST：透過自適應證據路由，為 LLM 計算出正確的上下文
Traditional RAG struggles when answers require aggregating or computing data across heterogeneous sources. RECAST solves this by modeling evidence construction as a sequential decision process. A lightweight RouterLM iteratively selects primitive operations or tasks a frozen CompilerLM to generate executable code. Once sufficient evidence is derived and accepted, a frozen AnswerLM outputs the final answer. Trained via SFT and GRPO, RECAST outperforms strong baselines by up to 15.9% and demonstrates exceptional zero-shot generalization across held-out tasks.
Key points
Active Evidence Derivation
Moves beyond passive retrieval to actively construct and synthesize evidence via iterative filtering, aggregation, and multi-step computation.
Router-Compiler-Answer Trio
Leverages a lightweight RouterLM for decisions, a frozen CompilerLM for code execution, and a frozen AnswerLM for final generation.
Reinforcement Learning via GRPO
Trained using SFT followed by Group Relative Policy Optimization (GRPO), enabling a 9B parameter RouterLM to outperform larger closed models.
Zero-Shot Generalization
Achieved a 15.0% average improvement over the strongest baseline on held-out benchmarks, demonstrating robust zero-shot task adaptation.
How it works
Why it matters
RECAST shifts the RAG paradigm from passive retrieval to active computation, which is essential for tasks requiring numerical reasoning and cross-source data aggregation. By proving that a 9B model optimized with GRPO can outperform a larger Gemini model, this research offers a cost-effective and highly accurate architectural blueprint for deploying specialized enterprise agents.
Who it affects
- AI Developer
- AI Researcher
- Enterprise Leader
How to use it
- 1Financial analysis and multi-source database querying
- 2Complex Q&A requiring cross-document logic and numerical operations
- 3Automated data cleaning and multi-step code-generation reasoning
Limitations & caveats
- Highly dependent on secure code execution environments; syntax or logical errors in compiled code can disrupt the reasoning loop.
- The iterative routing and code compilation process introduces higher latency and computational overhead compared to single-shot retrieval.
Related
Microsoft Releases Agent Lightning v1.0: A 3,500-Line Lightweight Agentic RL Framework for Real-Harness Training
微軟開源 Agent Lightning v1.0：僅 3,500 行程式碼，直接用生產環境 Harness 訓練 Agent 的強化學習框架
Microsoft Research Asia has introduced Agent Lightning v1.0, a lightweight open-source framework that trains AI agents using their actual deployment harnesses via a transparent LLM proxy.
Agent in a Bottle: Can LLM Agents Package Their Capabilities into Cheap, Scalable Artifacts?
打造低成本 AI 工件：LLM Agent 是否具備「能力封裝」的本領？
The study introduces the BOTTLED benchmark to evaluate if LLM agents can autonomously package their capabilities into low-cost, task-specific artifacts, revealing that strong zero-shot performance does not guarantee successful bottling.
AdvSim2Real: Co-Evolving Web Agents and Adversaries inside a Web World Model
利用 Web 世界模型進行對抗訓練：AdvSim2Real 提升 Web Agent 抵禦適應性提示詞注入之能力
AdvSim2Real co-evolves a task curriculum, an injection adversary, and a web agent inside a frozen world model, boosting the robustness of a 4B agent against adaptive prompt injections by 33.6%.