ToolRec: Calibrated preference alignment for tool-query recommendation¶
Core public-data preference mechanism.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | Huazhong University of Science and Technology. |
| First public release | 2026-06-07 (arXiv v1) |
| Author code | No public author code found as of 2026-09-02. |
| Local adapter or method key | toolrec |
| Local reproduction code | src/auto_research/reproductions/toolrec/ |
Paper method and evidence¶
Augment a KTO-style objective with user calibration and frequency-driven dynamic system calibration. Improve tool-triggering queries while limiting contextual relevance loss.
flowchart LR
A[User preference and tool frequency] --> B[Dual calibration]
B[Dual calibration] --> C[KTO-style query alignment]Original paper figure¶
Figure shows both calibration paths. From the original paper; copyright remains with the authors.
Core equation¶
\[
\mathcal L=-w_u r_y\log\sigma(z)-w_s(f_{tool})(1-r_y)\log\sigma(-z).
\]
Reported results¶
OPPO Xiaobu A/B, 2026-04-21 through 04-27, 5% per arm: clicks +4.74%, CTR +3.32%, relevance -1.44% relative.
Local reproduction¶
MovieLens-1M fixed catalog: NDCG@10 0.05401→0.04819 (-10.78%). Three-seed metrics, mean, SD and 95% CI. No private features, production traffic or serving. These local public-data results do not reproduce online lift.