Researchers propose UltraX, a function-calling refinement framework designed to address the limitations of existing large-scale data processing methods by introducing insertion alongside deletion and modification. This approach enables fine-grained instance-level editing through a reliable program-supervision generation pipeline that converts text pairs into structured supervision.
- The framework utilizes dataset-adaptive prompt optimization to guide an expert LLM in producing high-quality end-to-end refined texts.
- Line Alignment Mapping and Dynamic Context Replacement convert original-refined text pairs into structured program supervision.
- Low-confidence example filtering and ratio-controlled sampling by operation combination improve supervision quality and stabilize the training distribution.
- Sliding-window prediction, global operation aggregation, and systematic post-processing normalize and validate model outputs during execution.
Experiments demonstrate that UltraX achieves the highest average performance across all corpora while matching or surpassing baselines with fewer training tokens, indicating stronger data efficiency and refinement reliability.