A2FM-32B-rl

🧠 A²FM: Adaptive Agent Foundation Model for Tool-Aware Hybrid Reasoning

A²FM (Adaptive Agent Foundation Model) unifies reasoning-centric and agentic paradigms into a single framework that adaptively selects among three execution modes — instant, reasoning, and agentic.
It follows a route-then-align training principle and introduces Adaptive Policy Optimization (APO) to jointly optimize accuracy and efficiency.

A²FM achieves state-of-the-art performance on major reasoning and agentic benchmarks:

• 13.4% on BrowseComp (agentic)
• 70.4% on AIME25 (reasoning)
• 16.7% on HLE (general)

Notably, its adaptive execution achieves a cost of pass of only $0.00487 per correct answer, cutting cost by 45.2% vs. reasoning and 33.5% vs. agentic, delivering substantially higher cost efficiency while maintaining comparable accuracy.

📄 Paper
💻 GitHub

---

🔑 Key Highlights

• ⚙️ Unified reasoning & agentic modeling
  Integrates direct reasoning, chain-of-thought, and tool-augmented actions within a single backbone.

• 🔄 Route-then-Align supervised fine-tuning
  Trains task-aware routing followed by mode-aligned trajectory learning.

• 🧩 Adaptive Policy Optimization (APO)
  Reinforcement learning with adaptive sampling and cost-regularized reward for efficiency–accuracy balance.

• 💡 Substantially lower inference cost
  Adaptive routing cuts redundant reasoning/tool use while preserving correctness.

---

📘 Citation

@article{chen2025textsuperscript,
  title={A$\backslash$textsuperscript $\{$2$\}$ FM: An Adaptive Agent Foundation Model for Tool-Aware Hybrid Reasoning},
  author={Chen, Qianben and Cao, Jingyi and Zhang, Jiayu and Qin, Tianrui and Li, Xiaowan and Zhu, King and Shi, Dingfeng and Zhu, He and Liu, Minghao and Liang, Xiaobo and others},
  journal={arXiv preprint arXiv:2510.12838},
  year={2025}
}