LLM-BlockFE: LLM-Guided Offline Feature Construction from Long Text for Industrial Risk Control

Forum topic · 小凯 · 2026-10-12

Summary

Industrial risk-control systems rely on structured-data models for fast prediction, but much valuable information remains locked in unstructured long text. Manual feature engineering is labor-intensive, while calling an LLM for every real-time request is impractical. LLM-BlockFE is an LLM-guided offline feature construction framework that converts long text into executable feature programs, eliminating LLM calls at inference time. It builds feature programs by incrementally appending immutable code blocks and evaluates candidate features with a downstream model. To escape the local optima of greedy search, it introduces a block-level rollback mechanism based on depth-calibrated credit assignment and advances multiple independent search trajectories in an interleaved manner, sharing fixed descriptions of explored directions to reduce redundant exploration. On two public and two private datasets, LLM-BlockFE improves AUC by 0.0069 to 0.0358 in absolute terms over the strongest baseline. Across five deployed financial risk-control applications, it improves KS by 0.02 to 1.56 percentage points over existing manual strategies. arXiv: 2610.12390.

Paper Overview

Field: NLP
Authors: Ziming Dai, Dabiao Ma, Ziheng Guo, Jack Dong, Zimu Zhou
Published: 2026-10-08
arXiv: 2610.12390

Abstract (translated)

Industrial risk-control systems typically rely on structured-data models for efficient prediction, yet substantial valuable information remains embedded in unstructured long text. Extracting this information through manual feature engineering is labor-intensive, while requiring a large language model (LLM) to process every real-time input may not meet practical deployment requirements.

The authors 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. Key ideas:

Feature programs are constructed by incrementally appending immutable code blocks, with candidate features evaluated by a downstream model.
To address the tendency of conventional greedy search to converge to suboptimal solutions, a block-level rollback mechanism based on depth-calibrated credit assignment is introduced.
Multiple independent search trajectories are advanced in an interleaved manner, sharing fixed descriptions of each trajectory's explored directions to reduce redundant exploration.

Results

On 2 public and 2 private datasets, LLM-BlockFE achieves an absolute AUC improvement of 0.0069–0.0358 over the strongest baseline in full comparisons.
Across 5 deployed financial risk-control applications, it improves KS by 0.02–1.56 percentage points over existing manual strategies.

Original Abstract (excerpt)

> Industrial risk-control systems typically rely on structured-data models for efficient prediction, yet substantial valuable information remains embedded in unstructured long text. Extracting this information through manual feature engineering is labor-intensive, while requiring a large language model (LLM) to process every real-time input may not meet practical deployment requirements. 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. LLM-BlockFE constructs feature programs by incrementally appending immutable code blocks and evaluates candidate features using a downstream model. To address the tendency of conventional greedy se...

---

*Auto-collected on 2026-10-12.*