RECAST: Learning to Compute the Right Context through Adaptive Evidence Routing for LLMs

Forum topic · 小凯 · 2026-10-09

Summary

RECAST (Routing Evidence through Computation, Access, and Synthesized Tools) is a learned framework from researchers including Yilin Hao and Craig Boutilier that reframes evidence construction for large language models as a sequential decision process over heterogeneous retrieval and computation operations. Unlike conventional RAG, which depends on fixed similarity-based retrieval, RECAST actively derives evidence that is not explicitly present in any single source item, using filtering, aggregation, and computation across multiple items. A lightweight RouterLM iteratively selects and formulates primitive operations or specifies custom operations for a frozen CompilerLM, which translates them into executable code; once evidence is deemed sufficient, a frozen AnswerLM produces the final solution. RouterLM is trained via supervised fine-tuning followed by Group Relative Policy Optimization (GRPO). Across six heterogeneous benchmark families, RECAST achieves an average success rate of 75.6%, outperforming the strongest large-model baseline by 15.9%. A trained Qwen3.5-9B RouterLM surpasses a training-free Gemini 3.5 Flash RouterLM by 5.0%, and on three held-out benchmarks RECAST improves over the strongest baseline by an average of 15.0%, demonstrating strong zero-shot generalization across tasks and heterogeneous source representations. Paper: arXiv 2610.10507.

Paper Overview

Field: ML Authors: Yilun Hao, Krishna Sayana, Isabella Ye, James S Ren, Sukhdeep Sodhi, Craig Boutilier, Chuchu Fan Published: 2026-10-07 arXiv: 2610.10507

Abstract

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.

This work introduces 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 derived rather than merely retrieved. A lightweight RouterLM iteratively selects and formulates primitive operations, or specifies custom operations for a frozen CompilerLM to translate into executable code. Once evidence is judged sufficient, the RouterLM passes the accepted evidence to a frozen AnswerLM that generates the final solution.

The RouterLM is trained with supervised fine-tuning (SFT) followed by Group Relative Policy Optimization (GRPO).

Key Results

Across six heterogeneous benchmark families, RECAST achieves an average success rate of 75.6%, exceeding the strongest large-model baseline by 15.9%.
Training enables a Qwen3.5-9B RouterLM to outperform a training-free Gemini 3.5 Flash RouterLM by 5.0%.
On three held-out benchmarks, RECAST improves over the strongest baseline by an average of 15.0%, demonstrating strong zero-shot generalization across tasks and heterogeneous source representations.

---

*Collected automatically on 2026-10-09*