AI 研究arXiv cs.LG19 Oct, 12:00 pm SGTSIGNAL 28E95A

CastFlow: Role-Specialized Agentic Framework Bridging Evidence Acquisition and Accurate Time Series Forecasting

ORIGINAL / CastFlow: Bridging Evidence Acquisition and Accurate Prediction in Agentic Time Series Forecasting

CastFlow addresses two key challenges in LLM-based agentic time series forecasting—acquiring forecast-relevant evidence through adaptive tool interactions and translating evidence into accurate numerical predictions—by proposing a role-specialized framework: a frozen LLM, guided by experience memory, orchestrates a multi-view toolkit for evidence acquisition, while a lightweight specialized LLM performs evidence-grounded temporal reasoning to refine an ensemble forecasting pr

01 ABSTRACT

Fact: CastFlow is a role-specialized agentic forecasting framework comprising experience-memory-guided tool acquisition and specialized LLM evidence reasoning, outperforming competitive baselines on nine of ten real-world benchmarks. Author's view: bridging autonomous evidence acquisition and specialized temporal reasoning is important for accurate and adaptive agentic forecasting.

02 KEY FINDINGS

Proposes a role-specialized framework: a frozen LLM orchestrates a multi-view toolkit for evidence acquisition, and a lightweight specialized LLM performs evidence-grounded temporal reasoning to refine an ensemble forecasting prior.
Introduces experience memory that distills successful tool-use trajectories from training data to guide the frozen LLM's tool interactions.
Designs a two-stage training strategy combining supervised fine-tuning and reinforcement learning with verifiable rewards, where overall forecasting outcomes guide the optimization of the specialized model.

03 RELEVANCE

This work provides an actionable framework for how LLM agents can adaptively acquire evidence and translate it into numerical predictions in time series forecasting, and demonstrates superior performance over competitive baselines on most benchmarks, offering reference value for advancing the accuracy and adaptivity of agentic forecasting.

04 EVIDENCE BOUNDARY

Evidence boundary: conclusions are based on outperforming baselines on nine of ten real-world benchmarks, not winning on all datasets; experience memory relies on successful tool-use trajectories from training data, and its generalizability is not fully discussed; the two-stage training involving reinforcement learning and verifiable reward design lacks quantification of computational cost and stability; although code is public, cross-domain transfer or real-time deployment validation is not mentioned.

Return to the primary source

AI GENERATED SUMMARY / DISCOVERED BY ARXIV CS.LG

Read original ↗