Long Text to Predictive Features: LLM-Guided Blockwise Feature Engineering via Executable Program Search
arXiv cs.CL · 9 Oct 2026, 03:39 AEST
A new framework called LLM-BlockFE has been proposed to convert long text into executable feature programs for risk-control systems. This method aims to enhance prediction efficiency by avoiding real-time LLM calls during online inference.
Key facts
- Industrial risk-control systems typically rely on structured-data models for efficient prediction, yet substantial valuable information remains embedded in unstructured long text.
- To address this challenge, we propose LLM-BlockFE, an LLM-guided offline feature construction framework that converts long text into executable feature programs, thereby avoiding LLM calls during online inference.
- Across two public and two private datasets, LLM-BlockFE achieves absolute AUC improvements of 0.0069 to 0.0358 over the strongest baseline on each dataset in the full-dataset comparison.
- Post-launch monitoring across five deployed financial risk-control applications shows absolute KS improvements of 0.02 to 1.56 percentage points over the existing manually designed strategy.
Read aloud
A new framework called L L M-BlockFE has been proposed to convert long text into executable feature programs for risk-control systems. This method aims to enhance prediction efficiency by avoiding real-time L L M calls during online inference. Industrial risk-control systems typically rely on structured-data models for efficient prediction, yet substantial valuable information remains embedded in unstructured long text. To address this challenge, we propose L L M-BlockFE, an L L M-guided offline feature construction framework that converts long text into executable feature programs, thereby avoiding L L M calls during online inference. Across two public and two private datasets, L L M-BlockFE achieves absolute AUC improvements of zero point zero zero six nine to zero point zero three five eight over the strongest baseline on each dataset in the full-dataset comparison. Post-launch monitoring across five deployed financial risk-control applications shows absolute KS improvements of zero point zero two to one point five six percentage points over the existing manually designed strategy.
Sources in this cluster
- Long Text to Predictive Features: LLM-Guided Blockwise Feature Engineering via Executable Program Search
arXiv cs.CL · 9 Oct 2026, 03:39 AEST