Research map / Large language models / Reasoning and chain of thought
A$^2$FM: An Adaptive Agent Foundation Model for Tool-Aware Hybrid Reasoning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes A²FM, a 32B agent foundation model that routes each query to instant, reasoning, or agentic mode under one backbone, trained with route-then-align SFT and an adaptive policy optimization (APO) RL stage with cost-regularized rewards.
Problem
Reasoning-centric LLMs cannot invoke tools, while agentic LLMs lag on deep reasoning; both waste compute on simple queries. Existing adaptive-reasoning methods use binary think/no-think switches and do not learn tool-aware mode selection end-to-end.
Approach
A single backbone with three execution modes (instant, reasoning, agentic) and a learned router, trained by SFT on mode-conditioned trajectories then RL (APO) that forces rollouts in each mode and rewards correct, cheap mode choices via a multiplicative accuracy × adaptive × format reward.
Key results
- Adaptive A²FM reaches 70.4% on AIME25 and 74.5% on AIME24, and 56.0% on XBench-DS, the best reported among compared systems.
- On SuperGPQA, adaptive execution lowers cost-of-pass to $0.00487 per correct answer, a 45.2% reduction versus reasoning mode and 33.5% versus agentic mode.
- APO reaches 53.8% average accuracy with a 77.1% non-instant ratio, versus 55.4% for the best-mode oracle at 100% non-instant.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| adaptive routing (instant/reasoning/agentic) vs forced agentic mode (ablation) | A2FM · GAIA / XBench-DeepSearch / BrowseComp | accuracy: GAIA -3.4, XBench-DS +2.0, BrowseComp -1.0 (adaptive vs agentic) | Mixed |
| adaptive routing vs forced reasoning mode (ablation) | A2FM · AIME24 / AIME25 / MATH500 | accuracy: 0.0 on AIME24/25; +0.2 forced reasoning on MATH500 | No clear effect |
| adaptive routing vs forced reasoning mode (ablation) | A2FM · GPQA-d / SuperGPQA / MMLU-Pro / HLE | accuracy: GPQA-d -1.6, SuperGPQA +3.5, MMLU-Pro -3.2, HLE +3.3 | Mixed |
| adaptive routing vs forced agentic mode (ablation) | A2FM · GPQA-d / SuperGPQA / MMLU-Pro / HLE | accuracy: GPQA-d -4.6, SuperGPQA -1.3, MMLU-Pro -2.0, HLE -3.9 | Mixed |
| adaptive routing with APO vs forced reasoning and forced agentic modes (ablation) | A2FM · SuperGPQA cost of pass | cost of pass ($ per correct answer): $0.00487 vs $0.00889 (reasoning) and $0.00732 (agentic); -45.2% and -33.5% | Helped |
| instant mode for easy queries in adaptive routing | A2FM · SuperGPQA mode allocation by difficulty | instant-mode usage rate: 61.1% on easy vs 5.3% on difficult questions | Helped |
| APO reinforcement learning stage (vs. no APO) | A2FM · Accuracy-efficiency trade-off on general benchmarks | non-instant ratio and accuracy: non-instant ratio 77.1% vs 100% for best-mode oracle; accuracy 53.8 vs 55.4 | Helped |
| adaptive routing vs forced reasoning on AIME24/AIME25 (ablation) | A2FM · AIME24/25 | accuracy: 0.0 | No clear effect |
Surprising findings
- Forced agentic mode scores higher than adaptive mode on GAIA (60.7 vs 57.3) and BrowseComp, attributed to classification noise.
- Instant-mode accuracy stays around 55% across all difficulty levels on SuperGPQA, including hard questions.
Limitations
- Routing errors on boundary queries (e.g., GAIA classification noise) cause adaptive mode to underperform forced agentic mode.
- BrowseComp errors occur when the model overestimates internal knowledge and skips tool calls.
- Cost analysis relies on assumed token prices for the 32B model.
Experimental setup
Datasets: WebDancer WebShaper TaskCraft 2Wiki MuSiQue HotpotQA Skywork-Math DAPO-Math-17K GSM8K Skywork-Code Codeforces
Benchmarks: GAIA BrowseComp XBench-DeepSearch AIME24 AIME25 MATH500 GPQA-d SuperGPQA MMLU-Pro HLE
Metrics: accuracy (avg@1; avg@32 for AIME) cost-of-pass
Baselines: GPT-4.1 o1 Claude 4 Sonnet DeepSeek-R1 Qwen3-32B QwQ-32B OAgents DeepDive WebSailor AFM-Search
Builds on: GRPO Search-R1 Agent Foundation Models DeepSeek-R1
Closest papers by meaning
- AutoTool: Dynamic Tool Selection and Integration for Agentic Reasoning (ICML 2026 · similarity 0.90)
- MARS: Multi-Agent Adaptive Reasoning with Socratic Guidance for Automated Prompt Optimization (AAAI 2026 · similarity 0.90)
- Reasoning via Test-Time Instance-Level Policy Gradient in Latent Space (NeurIPS 2026 · similarity 0.90)
- Joint Optimization of Tool Creation and Use for Large Language Model Agents (NeurIPS 2026 · similarity 0.90)
- Adaptive Social Learning via Mode Policy Optimization for Language Agents (ICLR 2026 · similarity 0.89)
- Tools as Continuous Flow for Evolving Agentic Reasoning (NeurIPS 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.