Long Text to Predictive Features: LLM-Guided Blockwise Feature Engineering via Executable Program Search
The paper proposes freezing LLM-generated feature programs offline so production risk models can use long-text signals without live LLM calls.
LLM-BlockFE builds executable feature code in immutable blocks, tests candidates against a downstream model, and uses rollback plus interleaved search paths to avoid weak greedy choices. The authors report full-dataset AUC gains of 0.0069 to 0.0358 over the strongest baseline across two public and two private datasets. In five deployed financial risk-control applications, post-launch monitoring showed KS gains of 0.02 to 1.56 percentage points over the existing manual strategy. Source: ArXiv · AI/CL/LG's note.
LLM-BlockFE builds executable feature code in immutable blocks, tests candidates against a downstream model, and uses rollback plus interleaved search paths to avoid weak greedy choices. The authors report full-dataset AUC gains of 0.0069 to 0.0358 over the strongest baseline across two public and two private datasets. In five deployed financial risk-control applications, post-launch monitoring showed KS gains of 0.02 to 1.56 percentage points over the existing manual strategy. Source: ArXiv · AI/CL/LG's note.
score 4