A$^2$FM: An Adaptive Agent Foundation Model for Tool-Aware Hybrid Reasoning

TL;DR

A2FM proposes a unified three-mode foundation model that jointly leverages instant, reasoning, and agentic execution under a single backbone with a self-adaptive router. The training proceeds in two stages: route-then-align supervised fine-tuning to align mode-specific trajectories, followed by Adaptive Policy Optimization to balance accuracy and efficiency through adaptive mode selection and a cost-regularized reward. Empirically, A2FM achieves strong performance across agentic, reasoning, and general benchmarks at 32B scale, while delivering notable reductions in computation and token usage, achieving a cost per correct answer of approximately $0.00487. The results demonstrate that adaptive multi-mode routing can match or exceed specialized models while significantly improving efficiency, highlighting a scalable path toward versatile, tool-aware LLM agents.

Abstract

Large language models split into two families: reasoning-centric LLMs, which strengthen internal chain-of-thought reasoning but cannot invoke external tools, and agentic LLMs, which learn to interact with environments and leverage tools but often lag in deep reasoning. This divide arises from fundamentally different training objectives, leading to mismatched strengths and inefficiency on simple queries, where both families tend to overthink or over-call tools. In this work, we present Adaptive Agent Foundation Model (A$^2$FM), a unified framework that follows a route-then-align principle: the model first learns task-aware routing and then aligns mode-specific trajectories under a shared backbone. To address the inefficiency gap, we introduce a third mode-instant-that handles simple queries directly, preventing unnecessary reasoning or tool calls while complementing the agentic and reasoning modes. To jointly enhance accuracy and efficiency, we propose Adaptive Policy Optimization (APO), which enforces adaptive sampling across modes and applies a cost-regularized reward. On the 32B scale, A$^2$FM achieves 13.4% on BrowseComp, 70.4% on AIME25, and 16.7% on HLE, setting new SOTA among comparable models and performing competitively with frontier LLMs across agentic, reasoning, and general benchmarks. Notably, the adaptive execution achieves a cost of pass of only $0.00487 per correct answer-cutting cost by 45.2% relative to reasoning and 33.5% relative to agentic, thus delivering substantially higher cost efficiency while maintaining comparable accuracy.

Figures (6)

• Figure 1: Average performance on agentic, reasoning, and general (ARG) benchmarks. Overall, A$^2$FM ranks 1st, 2nd, and 1st on the three categories, respectively. Note that we uses the best-suited mode for each dataset respectively.
• Figure 2: Overview of A2FM. Left: the framework integrates three execution modes—instant, reasoning, and agentic—under a unified backbone with task-aware routing. Right: mode allocation across six benchmarks (MMLU-Pro, GPQA-d, AIME25, MATH500, Xbench-DeepSearch (Xbench), and GAIA-text (GAIA), illustrating how A2FM adapts routing to task characteristics.
• Figure 3: Difficulty-based sampling produces a J-shaped distribution for both the agentic (left) and reasoning (right) datasets, reducing trivially solved cases while emphasizing moderately challenging queries.
• Figure 4: Overview of Adaptive Policy Optimization (APO). Left: Rollout and reward process. For each query, mode-specific rollouts are generated either by prefix injection (forced agentic/reasoning/instant) or by adaptive classification. Both prefix-injection tokens and tool-response tokens are excluded from loss since they are not model-generated. Right: Accuracy–efficiency trajectory under APO, showing how A2FM progressively approaches the Pareto frontier by improving accuracy while reducing non-instant triggering (excluding AIME24/25).
• Figure 5: Efficiency analysis on SuperGPQA: mode allocation vs. difficulty and token usage across modes.