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

• Reasoning-centric models emphasize internal chain-of-thought but lack tool use, while agentic systems use tools yet can trail on deep reasoning.External orchestration can add computational and latency overhead, require workflow engineering for new domains, and leave multi-tool decisions outside end-to-end training.Source: §1 · p. 3
• Prior thinking-versus-non-thinking approaches generally offer binary reasoning control, not adaptive tool use, leaving broader reasoning–acting trade-offs unaddressed.Adding agentic behavior can further increase runtime and cost of pass.Source: §1 · p. 3
• A2FM unifies instant direct answering, explicit-CoT reasoning, and tool-aware agentic action under one backbone with a self-adaptive router.Its route-then-align principle routes by task, then aligns mode-conditioned trajectories using tool-use, reasoning, and direct-answer data.Source: §1 · pp. 3–4
• APO uses adaptive sampling across modes and a dynamic cost-regularized reward to balance accuracy and efficiency in mode selection.It favors minimal-effort instant solutions for easy queries and escalates to reasoning or agentic modes when evidence or extended deliberation is needed.Source: §1 · p. 4
• At 32B, A2FM reports 13.4% on BrowseComp, 70.4% on AIME25, and 16.7% on HLE, with state-of-the-art results across benchmarks.On SuperGPQA, its cost of pass is $0.00487 per correct answer, 45.2% below reasoning and 33.5% below agentic while maintaining comparable accuracy.Source: §1 · p. 4

2.1 Agent Systems and Frameworks

• Agent systems demonstrate tool-integrated workflows, but often incur communication and reasoning overhead, need workflow engineering to generalize, and rely on fixed rather than learned routing.Tool-focused single- and few-agent systems add structured browsing, verification, and planning, yet routing among direct answers, reasoning, and tools is often manually designed.Source: §2.1 · pp. 4–5

2.2 Adaptive Reasoning Approaches

• Adaptive reasoning approaches regulate either reasoning depth through length-aware control or whether extended reasoning is invoked through capability-aware routing.Source: §2.2 · p. 5

2.2.1 Length-Aware Control

• Length-aware methods use token penalties or efficiency rewards, or supervised CoT compression, to reduce redundant reasoning while maintaining task performance.Source: §2.2.1 · p. 5

2.2.2 Capability-Aware Routing

• Capability-aware routing uses signals such as uncertainty or query semantics to trigger reasoning, but generally focuses on binary short-versus-long CoT rather than adaptive tool use.Source: §2.2.2 · p. 5

3.1 Problem Formulation

• The system routes each query to an execution mode, then uses that mode's policy to generate a direct answer, chain-of-thought, or tool-interaction trajectory.The formulation jointly optimizes routing and mode policies for expected accuracy across a task mixture.Source: §3.1 · pp. 5–6

3.2 Stage 1: Route-then-Align Fine-tuning

• Route-then-align training first classifies queries as instant, reasoning, or agentic, then generates trajectories consistent with the selected mode.The modes produce direct answers, chain-of-thought followed by an answer, or tool-interleaved reasoning, respectively.Source: §3.2 · p. 6
• Agentic trajectories use a one-time plan to split queries into parallel subgoals and dynamically summarize solved tasks, enabling simultaneous multi-tool execution.The summary can terminate completed threads and open new ones as the process proceeds.Source: §3.2 · p. 6
• Difficulty-based sampling downweights always-solved examples to emphasize moderately challenging queries, while ambiguous queries receive the label of the trajectory with highest accuracy.Figure 3 depicts the resulting J-shaped distributions for the agentic and reasoning datasets.Source: §3.2 · p. 7, Figure 3 · p. 7

3.3 Stage2: Adaptive Policy Optimization

• For each query, APO samples forced rollouts in all three modes and additional adaptive rollouts in which the model selects its own mode.The forced rollouts support estimates of relative mode success; adaptive rollouts allow correct self-routing to be rewarded.Source: §3.3 · pp. 7–8
• APO rewards correctness, penalizes reasoning or tool use on easy queries solvable instantly, and assigns zero reward to outputs that violate the mode-specific format.Easy queries are those for which instant-mode accuracy exceeds threshold τ; the total reward combines accuracy, adaptivity, and format constraints.Source: §3.3 · pp. 8–9
• APO uses on-policy training and omits KL-divergence computation to address classification instability and expedite training and exploration of efficient mode choices.The policy optimizes groups of outputs sampled from the old policy.Source: §3.3 · p. 9
• APO's rollout strategy and adaptive reward are reported to improve adaptive routing while boosting accuracy and efficiency across diverse query scenarios.The method is designed to encourage accurate yet cost-effective routing.Source: §3.3 · pp. 7, 9

4.1 Dataset Composition

• Training data combines challenging agentic, mathematical-reasoning, and code sources while retaining simple examples to support instant-mode learning.Route-then-align SFT uses 5,289 agentic, 2,829 reasoning, and 2,890 instant trajectories; APO uses 6,000 question-answer pairs.Source: §4.1 · p. 9

4.2 Experimental Setup

• Evaluation spans tool-use, mathematics, and general-knowledge benchmarks, using a Qwen2.5-32B-Instruct backbone and route-then-align SFT followed by APO.The reported benchmark set includes BrowseComp, GAIA, XBench-DeepSearch, AIME24/25, Math500, GPQA-d, SuperGPQA, MMLU-Pro, and HLE.Source: §4.2 · p. 10

4.3 Main Results

• On agentic benchmarks, adaptive A2FM scores 56.0% on XBench-DS, ranks second on GAIA, and reaches 12.4% on BrowseComp, behind DeepDive and OAgents.On BrowseComp, it exceeds AFM-Search’s 11.1%; the comparison table covers agentic, reasoning, and general-knowledge benchmarks.Source: §4.3 · p. 10, Table 1 · p. 10
• Forced A2FM-Agentic reaches 14.4% on BrowseComp, second behind DeepDive (14.8%), and 60.7% on GAIA, establishing a new SOTA.The paper reports these results as evidence that tool-focused execution can outperform specialized frameworks in complex deep-search scenarios.Source: §4.3 · pp. 10–11
• On reasoning benchmarks, A2FM scores 74.5% on AIME24, a new SOTA, and 70.4% on AIME25, second only to o1 (79.2%).Adaptive and reasoning-specific variants have identical AIME24/25 results, with nearly all AIME queries routed to reasoning.Source: Table 1 · p. 10, §4.3 · p. 11
• On general benchmarks, A2FM improves over Qwen2.5-32B by 13.6 points on GPQA-d (63.1 vs. 49.5) and 15.9 on SuperGPQA (54.7 vs. 38.8).The paper reports that its SuperGPQA score surpasses GPT-4.1 (50.8) and Claude 4 Sonnet (55.7).Source: Table 1 · p. 10, §4.3 · p. 11
• On SuperGPQA, instant mode handles 61.1% of easy questions but 5.3% of difficult ones, while its accuracy stays at 55% across difficulty levels.Figure 5 presents mode allocation by difficulty and token usage across modes.Source: §4.3 · p. 11, Figure 5 · p. 11
• $0.00487 per correct answer is A2FM’s adaptive cost on SuperGPQA, 45.2% below reasoning and 33.5% below agentic execution.The comparison is reported alongside the efficiency analysis in Figure 5.Source: §4.3 · p. 11, Figure 5 · p. 11

5 Analysis

• A2FM (RL) converges to 53.8% average accuracy with a 77.1% non-instant ratio, versus the Best Mode oracle’s 55.4% accuracy and 100% ratio.The 1.6-point accuracy gap accompanies a 22.9-point reduction in non-instant triggering; Figure 4 shows the trajectory toward the accuracy-efficiency Pareto frontier.Source: §5 · p. 12
• Routing accuracy is 92.2% on GAIA, 94.0% on BrowseComp, and 100% on AIME24.GAIA labels use majority votes from five baselines followed by manual verification; errors often involve plausible reasoning/tool-use alternatives, while BrowseComp errors include missed tool calls.Source: §5 · p. 12, Table 2 · p. 12
• Across four benchmarks, adaptive mode is competitive or superior to forced single modes, but trails agentic on GAIA (57.3 vs. 60.7), largely due to classification noise.Adaptive surpasses single modes on XBench-DeepSearch and AIME24, and matches reasoning on AIME24 because all queries route to reasoning; Figure 6 compares the modes across benchmarks.Source: §5 · p. 12

6 Conclusion

• A2FM matches or surpasses strong baselines across agentic, reasoning, and general benchmarks while achieving high accuracy and efficiency.The conclusion gives no specific benchmark values.Source: §6 · p. 12
• A2FM combines agentic, reasoning, and instant modes under a single backbone with a self-adaptive router.Source: §6 · p. 12
• The authors identify adaptive multi-mode modeling as a scalable path toward efficient and versatile LLM agents.Source: §6 · p. 12

Core Contributors

• The listed core contributors are Qianben Chen and Jingyi Cao.Source: Core Contributors · p. 13

A.1 System Prompt

• The system prompt offers agentic, reasoning, and instant options, selected according to whether a task needs external information or code, complex reasoning, or neither.Source: §A.1 · p. 18
• The prompt directs the model to classify a task into one of the three agents before formulating its processing trajectory.Source: §A.1 · p. 18
• Execution instructions require purposeful tool use, parallel work on independent goals, sequential paths within each goal, and verification before marking goals complete.Source: §A.1 · p. 18
• Planning instructions call for parallelizable goals, sequential fallback paths, and success criteria for each path.The prompt also specifies timing only for the first step.Source: §A.1 · p. 18
• Progress summaries recap goals and path status, identify what worked or failed, and specify next parallel sub-paths.Source: §A.1 · p. 19
• The prompt specifies web-search, page-crawling, and code-execution tools, requiring valid parameters and recommending page verification after web search.Source: §A.1 · p. 19
• Reasoning mode requires a reasoning phase before the answer, while instant mode uses only an immediate answer without planning or tool calls.The prompt sets a greater-than-1,000-word reasoning phase and an under-300-word instant trajectory.Source: §A.1 · p. 20

A.2 Summary Prompt

• A.2 includes a post-RL case-trajectory example and a web-page extraction prompt requesting three points on query information, clues, and relevant links.Source: §A.2 · p. 20

B Case Study

• The trajectory classifies the architecture comparison as requiring accurate factual lookup and selects toolcalling_agent to query external sources.The rationale is that the answer depends on precise model-definition information rather than common sense.Source: B Case Study · pp. 20–21
• 12 BERT Base encoder layers versus 6 base Transformer encoder layers yields a difference of 6 layers.The BERT passage identifies 12 encoder layers, while the Transformer passage states its encoder has N = 6 identical layers.Source: B Case Study · pp. 23–24
• BERT-Base has 12 encoder layers versus N = 6 in the Attention is All You Need base encoder, a difference of 6 layers.The post-RL model gives the same comparison: 12 - 6 = 6.Source: B Case Study · p. 25
• The question is classified as a straightforward factual comparison, so the default_agent is selected without web search or code execution.The passage says this query does not require external search or code and identifies the default agent as appropriate for basic factual queries.Source: B Case Study · p. 25

C Template

• Table 3 lists templates used in training.Source: Table 3 · p. 26