RECAST: Learning to Compute the Right Context through Adaptive Evidence Routing
arXiv cs.AI · 8 Oct 2026, 03:51 AEST
RECAST introduces a learned framework for evidence construction in large language models. It allows for active evidence derivation through a sequential decision process over various retrieval and computation operations.
Key facts
- Conventional Retrieval-Augmented Generation (RAG) relies on fixed similarity-based retrieval, while agentic variants adapt queries and tool use but remain largely retrieval-centric.
- 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.
- Across six heterogeneous benchmark families, RECAST achieves a mean success rate of 75.6%, outperforming the strongest large-model baseline by 15.9%.
- Moreover, training enables the 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 15.0% on average, demonstrating strong zero-shot generalization across tasks and heterogeneous source representations.
Read aloud
RECAST introduces a learned framework for evidence construction in large language models. It allows for active evidence derivation through a sequential decision process over various retrieval and computation operations. Conventional Retrieval-Augmented Generation RAG relies on fixed similarity-based retrieval, while agentic variants adapt queries and tool use but remain largely retrieval-centric. 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. Across six heterogeneous benchmark families, RECAST achieves a mean success rate of seventy five point six percent, outperforming the strongest large-model baseline by fifteen point nine percent. Moreover, training enables the Qwen3.five-9B RouterLM to outperform a training-free Gemini three point five Flash RouterLM by five point zero percent. On three held-out benchmarks, RECAST improves over the strongest baseline by fifteen point zero percent on average, demonstrating strong zero-shot generalization across tasks and heterogeneous source representations.
Sources in this cluster
- RECAST: Learning to Compute the Right Context through Adaptive Evidence Routing
arXiv cs.AI · 8 Oct 2026, 03:51 AEST