LLM-BlockFE
An LLM-guided block-level feature engineering framework that builds executable feature programs by incrementally appending immutable code blocks. Features are generated offline and executed online, and the framework is used for long-text feature extraction and risk-control modeling.
Timeline (last 90 days)
The LLM-BlockFE framework lets the LLM compile long text into an executable feature extraction program in the offline stage, and online inference only executes that program without calling the LLM again.
LLM-BlockFE builds the feature program by incrementally appending immutable code blocks, and once each code block is written it is no longer modified.
LLM-BlockFE adopts a block-level rollback mechanism, allowing rollback at the granularity of code blocks, and combined with depth-calibrated credit assignment to judge which code block should be responsible for the current effect.
LLM-BlockFE adopts multi-trajectory interleaved search, advancing multiple independent search trajectories simultaneously, with different trajectories sharing a fixed direction description to reduce repeated exploration.
LLM-BlockFE was validated on four datasets (two public datasets plus two private datasets).
In the full-data comparison, relative to the strongest baseline, LLM-BlockFE's absolute AUC improvement on each dataset is between 0.0069 and 0.0358.
In five deployed financial risk control applications, after LLM-BlockFE went live, monitoring showed an absolute KS improvement of between 0.02 and 1.56 percentage points, compared against existing manually designed strategies.
After the LLM-BlockFE search ends, the program is frozen and deployed, and online inference only needs to execute this program to extract structured features, with inference overhead on par with handcrafted feature engineering.
All Facts (8)
The LLM-BlockFE framework lets the LLM compile long text into an executable feature extraction program in the offline stage, and online inference only executes that program without calling the LLM again.
50%UnverifiedLLM-BlockFE builds the feature program by incrementally appending immutable code blocks, and once each code block is written it is no longer modified.
50%UnverifiedLLM-BlockFE adopts a block-level rollback mechanism, allowing rollback at the granularity of code blocks, and combined with depth-calibrated credit assignment to judge which code block should be responsible for the current effect.
50%UnverifiedLLM-BlockFE adopts multi-trajectory interleaved search, advancing multiple independent search trajectories simultaneously, with different trajectories sharing a fixed direction description to reduce repeated exploration.
50%UnverifiedLLM-BlockFE was validated on four datasets (two public datasets plus two private datasets).
50%UnverifiedIn the full-data comparison, relative to the strongest baseline, LLM-BlockFE's absolute AUC improvement on each dataset is between 0.0069 and 0.0358.
50%UnverifiedIn five deployed financial risk control applications, after LLM-BlockFE went live, monitoring showed an absolute KS improvement of between 0.02 and 1.56 percentage points, compared against existing manually designed strategies.
50%UnverifiedAfter the LLM-BlockFE search ends, the program is frozen and deployed, and online inference only needs to execute this program to extract structured features, with inference overhead on par with handcrafted feature engineering.
50%