App diversity mattered more at larger synthetic tool-use training budgets
Source check: Source checked. Not independent scientific confirmation.
Finding: Across 849 Qwen3.5 training runs, tool-use performance depended on how demonstrations were split among apps, tasks and database configurations. A composition-aware fit predicted the highest-compute quarter with mean absolute error 0.029, versus 0.171 from task count alone. A separate synthetic dataset also favored greater app diversity at its higher task budget.
Technology: Qwen3.5; GLM-5.2 demonstrations; nonlinear scaling regression
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
- AI contribution
- GLM-5.2 supplied verifier-screened synthetic tool-use episodes for supervised training of Qwen3.5 students. Performance under controlled dataset composition trained size-specific nonlinear scaling fits.
- Test and result
- Across 849 SynthAPI runs, a composition-aware fit had 0.0285 mean absolute error on the highest-compute quartile versus 0.1706 for the task-count-only fit. A separate SimSAT subsampling analysis observed a switch toward higher app diversity at 1,024 versus smaller task budgets.
- Limitation / still unknown
- The full fit was validated within the SynthAPI pipeline, and the second dataset checked one app-diversity trend. Both used synthetic environments. App coverage and benchmark weighting can affect the fitted relationship.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/abs/2610.37648v1 · https://arxiv.org/html/2610.37648v1
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review date: 21 Nov 2026. Review scope: arXiv v1 §§3.2, 4.1–4.3 and Appendix C, E.1 and E.3.
No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.
Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.