An arXiv preprint dated 20 Aug 2026 reports higher tool-use scores for Qwen3-4B-Base and Qwen3-8B-Base configurations that included a dedicated MidTool-Mix mid-training stage alongside supervised fine-tuning (SFT) and optional reinforcement learning (RL). The comparisons covered BFCLv3, tau2-Bench and MCP-Universe, with MidTool-Mix configurations above SFT-only comparisons on the reported overall measures.
The pattern had a clear boundary. On MCP-Universe’s web-search subset, the reported score remained at 0.00 even as browser automation, financial analysis and location improved in the reported comparisons.
The training idea
The study asks whether general tool-use ability can be shaped earlier through dedicated mid-training rather than being left almost entirely to post-training. Its pipeline collects web pages, PDFs, code repositories and structured tool artifacts, then converts them into two synthesis branches aimed at grounding and execution.
The final MidTool-Mix contained 20.3B tokens and 11.22M samples: 42% web, 26% code, 23% PDF and 9% native agentic trajectories.
In the main experiments, the team used a sampled 100K TOUCAN subset for SFT and 526 synthetic tool-use environments for RL. At each scale, raw-base and MidTool-Mix mid-trained models were compared under SFT and optional RL, with released post-trained Qwen3 models serving as references.
The scores moved together
On BFCLv3, MidTool-Mix plus SFT scored 50.25% at 4B and 51.12% at 8B overall, compared with 39.73% and 47.62% for SFT-only. The MidTool-Mix plus SFT plus RL configurations had listed overall scores of 54.18% and 55.12%, with multi-turn averages of 27.63% and 37.63%.
On tau2-Bench, MidTool-Mix plus SFT and RL had overall Pass@1/Pass@4 results of 19.96%/38.49% at 4B and 21.31%/39.57% at 8B. The SFT-only comparisons were 8.54%/20.50% and 10.43%/28.06%, respectively.
On MCP-Universe, the MidTool-Mix plus SFT and RL configurations had overall scores and pass rates of 23.80 and 10.06% at 4B, and 25.16 and 9.50% at 8B. The corresponding SFT-only figures were 13.20 and 1.68%, and 15.18 and 3.35%.
In a fixed 4B SFT ablation, the complete MidTool-Mix condition scored 50.25% overall on BFCLv3, versus 39.73% with no mid-training. On tau2-Bench Pass@1, it scored 12.23% versus 8.54%; on MCP-Universe, its score and pass rate were 18.66 and 5.03%, compared with 13.20 and 1.68% without mid-training.
The two synthesis branches showed complementary patterns in that ablation. Native-only had a larger BFCLv3 overall gain over no mid-training, +7.9 percentage points, than context-grounded-only, +4.9 percentage points; context-grounded-only was stronger on tau2-Bench and MCP-Universe. The combined mixture was the only configuration higher than no mid-training on all eight ablation metrics.
Signals beyond the main tests
A separate exploratory visual-tool pilot reported higher figures for MidTool-Mix than for baseline: tool success was 0.7231 versus 0.5863, while the overall rubric score was 0.0661 versus 0.0567. The pilot was small and no uncertainty estimates were reported.
During RL, MidTool-Mix models began from higher reward and improved more rapidly during the early training stage, according to the paper’s qualitative reward curves.
What the comparisons do not settle
These results remain comparisons among the tested training recipes. The manuscript reports no confidence intervals, formal significance tests or run-to-run variability estimates, so it does not establish that MidTool-Mix alone caused the benchmark differences.
Paper data and sources
Original title: MidTool: Mid-training Data Synthesis for Agentic Tool Use
Authors: Fengqing Jiang, Yite Wang, Boyi Liu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text