A method that uses a language model to write text-feature code improved risk-model separation across five financial applications, its authors report. The model does its work while the features are being designed; live systems run the resulting code without asking an LLM to process each new case. The researchers posted the study to arXiv on October 8.
Illustrative call-center photo by Kellie Walenciak, licensed under CC BY-SA 4.0.
Financial models can use structured fields such as counts and amounts, but longer notes and call records are harder to turn into consistent inputs. The researchers’ LLM-BlockFE method asks an LLM to propose code that converts text into numerical features. It tests each addition and keeps it only when the downstream model’s validation score improves. Once the search ends, the selected programs are frozen and run as ordinary feature-extraction code.
In tests on four datasets—two public and two private—the method’s area under the receiver operating characteristic curve, or AUC, beat the strongest tested baseline by 0.0069 to 0.0358. AUC measures how well a model ranks cases with different outcomes; it is not the percentage of correct decisions. The tests used separate training, validation and test data and averaged results across five random seeds.
The paper also describes five deployed financial risk-control applications monitored from September through November 2025. The authors compare the new and manually designed text features on the same samples, labels, structured fields and downstream model. Across all 15 application-month comparisons covering data from 2,694,342 users, the new features had higher Kolmogorov–Smirnov (KS) scores, a measure of how separated model scores are for the two outcome groups; the authors describe the higher scores as greater discrimination. Gains ranged from 0.02 to 1.56 percentage points.
These results come from an arXiv preprint, not an independent replication. The two private datasets are from Qfin Holdings credit-risk applications, and one author lists Qfin as an affiliation; the paper does not establish that the gains carry over to other financial institutions.