FusionEval is a newly released model-fusion evaluation benchmark, which was posted on arXiv on October 24, 2026 (arXiv:2610.30162v1). Its core objective is to systematically assess the performance of combining the capabilities of multiple domain-specific expert models into a single model, with particular emphasis on long-chain reasoning and tool‑calling abilities that have been overlooked in existing evaluations.
This benchmark constructs a controlled pool of expert models, with all experts independently trained on the Qwen3 and Llama foundation models, covering five major domains: mathematics, programming, science, instruction following, and tool usage. The evaluation suite integrates both fine-tuned and zero-shot fusion approaches, and uniformly assesses capability retention and computational overhead across eight public benchmarks, reporting normalized average scores, worst-case performance, runtime, and peak GPU memory usage.
Experiments show that while the training-free fusion approach can maintain reasoning capabilities reasonably well, it suffers a significant degradation on tool-use tasks; in contrast, the MOPD method better preserves tool‑calling capabilities and achieves a more balanced distribution of abilities, but requires additional data preparation, inference adaptation, and training pipelines. How to jointly preserve reasoning, instruction following, and tool‑use capabilities at low computational cost remains an open challenge. The project’s code has been made open source: https://github.com/Baicaihaochi/FusionEval_Benchmark。