RECAST: Learning to Compute the Right Context through Adaptive Evidence Routing
RECAST turns evidence gathering into a learned routing problem, choosing when to retrieve, compute, or synthesize tools.
The paper says fixed RAG misses cases where the answer has to be derived across sources, not found in one passage. Its RouterLM selects operations step by step, can ask a frozen CompilerLM to generate executable code, then sends accepted evidence to a frozen AnswerLM. Trained with SFT and GRPO, RECAST reports a 75.6% mean success rate across six benchmark families, 15.9 points above the strongest large-model baseline. On three held-out benchmarks, it reports a 15.0-point average gain over the strongest baseline.
ArXiv · AI/CL/LG's note
The paper says fixed RAG misses cases where the answer has to be derived across sources, not found in one passage. Its RouterLM selects operations step by step, can ask a frozen CompilerLM to generate executable code, then sends accepted evidence to a frozen AnswerLM. Trained with SFT and GRPO, RECAST reports a 75.6% mean success rate across six benchmark families, 15.9 points above the strongest large-model baseline. On three held-out benchmarks, it reports a 15.0-point average gain over the strongest baseline.
ArXiv · AI/CL/LG's note
score 5