PaperlayerView exact revision on arXiv

Source-linked AI summary

AgentFrontier: Expanding the Capability Frontier of LLM Agents with ZPD-Guided Data Synthesis

Xuanzhong Chen, Zile Qiao, Guoxin Chen, Liangcai Su, Zhen Zhang, Xinyu Wang, Pengjun Xie, Fei Huang, Jingren Zhou, Yong Jiang

arXiv:2510.24695v1cs.CL

TL;DR

LLMs still struggle with cross-domain knowledge fusion, while scalable training data and frontier-level assessments remain limited. AgentFrontier uses ZPD-guided synthesis to create training data and an evolving exam; its trained model reports state-of-the-art results on expert-level multidisciplinary benchmarks, surpassing some larger proprietary agents.

  • Problem

    Existing datasets seldom target knowledge fusion across diverse sources, and expert-crafted frontier evaluations are costly and difficult to scale.

    Source: §1 · p. 2

  • Method

    AgentFrontier uses LKP–MKO calibration to synthesize ZPD-targeted training data and an automated benchmark for frontier reasoning.

    Source: §1 · p. 3

  • Results

    AgentFrontier-30B-A3B achieved 28.6% on HLE and state-of-the-art performance on ZPD Exam-v1, R-Bench-T, and xBench-ScienceQA.

    Source: §1 · p. 3

  • Takeaways & Limitations

    The paper concludes that ZPD-guided data synthesis is an effective approach for cultivating advanced reasoning in a data-efficient manner.

    Source: §7 · p. 14

    The agent’s problem-solving capacity is bounded by its predefined, static toolset; it cannot create tools.

    Source: Limitations and Future Work · p. 14

Abstract

from arXiv · show

Training large language model agents on tasks at the frontier of their capabilities is key to unlocking advanced reasoning. We introduce a data synthesis approach inspired by the educational theory of the Zone of Proximal Development (ZPD), which defines this frontier as tasks an LLM cannot solve alone but can master with guidance. To operationalize this, we present the AgentFrontier Engine, an automated pipeline that synthesizes high-quality, multidisciplinary data situated precisely within the LLM's ZPD. This engine supports both continued pre-training with knowledge-intensive data and targeted post-training on complex reasoning tasks. From the same framework, we derive the ZPD Exam, a dynamic and automated benchmark designed to evaluate agent capabilities on these frontier tasks. We train AgentFrontier-30B-A3B model on our synthesized data, which achieves state-of-the-art results on demanding benchmarks like Humanity's Last Exam, even surpassing some leading proprietary agents. Our work demonstrates that a ZPD-guided approach to data synthesis offers a scalable and effective path toward building more capable LLM agents.

1 Introduction

The paper targets the shortage of scalable training data and benchmarks for cross-domain knowledge fusion and agentic reasoning. It proposes ZPD-guided synthesis and evaluation, then reports strong results for a model trained on the synthesized data.

1 Introduction

  • Existing datasets and synthesis approaches seldom target knowledge fusion across diverse sources, while expert-crafted frontier evaluations are costly and difficult to scale.The introduction also identifies limited unified training support for tool use, self-reflection, planning, and multi-step reasoning.Source: §1 · p. 2
  • The AgentFrontier Engine synthesizes multidisciplinary reasoning data within an LLM’s ZPD, allocating knowledge-intensive tasks to continued pre-training and frontier-level tasks to post-training.The framework uses adversarial calibration between a base model without tools and a superior tool-augmented agent.Source: §1 · p. 3
  • ZPD Exam is an automated benchmark for advanced capabilities, using multidisciplinary questions that require cross-document fusion and deep research.The paper presents it alongside the Engine as a contribution for probing agent capabilities at the frontier.Source: §1 · p. 3
  • 28.6% on HLE; AgentFrontier-30B-A3B also achieved state-of-the-art performance on ZPD Exam-v1, R-Bench-T, and xBench-ScienceQA.The model was continually pre-trained on 50 billion tokens of knowledge-intensive data and post-trained on 12,000 frontier-level QA trajectories.Source: §1 · p. 3

2 AgentFrontier Data Engine

The AgentFrontier Engine turns a document corpus into calibrated training data through seed-question generation, agentic complexity escalation, and ZPD-based filtering. Its pipeline combines knowledge fusion with checks that distinguish learnable frontier tasks from simpler or potentially flawed ones.

2 AgentFrontier Data Engine

  • The engine targets complex reasoning data at the LLM capability frontier through a three-stage agentic synthesis pipeline.It transforms a raw corpus into a calibrated, high-value dataset.Source: §2 · p. 3, Figure 3 · p. 4

2.1 Stage I: Seed Question Generation for Knowledge Fusionn

  • The pipeline retrieves thematically coherent chunk triplets, then uses a generator model to create seed question-answer pairs for later refinement.The triplets must exceed a thematic-similarity threshold for every distinct pair.Source: §2.1 · p. 3

2.2 Stage II: Escalating Complexity through Agentic Refinement

  • A tool-augmented refinement agent iteratively enriches each QA pair through knowledge expansion, abstraction, factual cross-validation, and computational formulation.Each iteration becomes input to the next, producing increasingly intricate reasoning paths.Source: §2.2 · p. 4

2.3 Stage III: ZPD-based Filtering and Calibration

  • Calibration sends base-model-solvable pairs to continued pre-training and tests harder pairs with the tool-augmented agent; pairs it solves at least once across three attempts enter the ZPD training set.Pairs unsolved in all three attempts are routed to human review; a semantic redundancy filter further removes overly similar questions.Source: §2.3 · pp. 4–5

3 ZPD Exam: A Self-Evolving Benchmark for LLM Agents

The ZPD Exam is an automated, evolving assessment intended to test agentic research rather than static parametric knowledge. Its calibrated tasks and tiered interpretation distinguish unaided competence, assisted reasoning bottlenecks, and emergent mastery.

3 ZPD Exam: A Self-Evolving Benchmark for LLM Agents

  • The ZPD Exam is designed to assess deep research capabilities with tasks that cannot be solved using parametric knowledge alone.The benchmark is intended to co-evolve with rapidly advancing models, unlike static expert-crafted exams that are costly to create.Source: §3 · p. 5, §3.1 · p. 5

3.1 Benchmark Construction: From Frontier Knowledge to Agentic Research

  • Exam questions are generated from a training-disjoint corpus and included only when a baseline fails three unaided attempts but solves them consistently with tools.ZPD Exam-v1 contains 1,024 public questions and a corresponding private set; its short-answer format supports automated grading.Source: §3.1 · p. 6
  • ZPD Exam-v1 contains 1,024 questions across nine disciplines, spanning mathematics, computer science and AI, physics, history, humanities, chemistry, biology and medicine, engineering, and geography.The exam is grounded in recent scientific papers and uses questions calibrated to a baseline model’s assisted capability frontier.Source: Figure 5 · p. 6, §3.1 · p. 6

3.2 ZPD Exam: A Diagnostic Benchmark for Agentic Reasoning

  • Scores define three diagnostic zones: below 20 reflects intrinsic competence, 20–60 a reasoning bottleneck, and above 60 emergent mastery.The middle zone describes agents that can partially succeed with assistance but struggle to plan, synthesize across tool calls, and adapt their approach.Source: §3.2 · pp. 6–7
  • The benchmark is presented as a diagnostic instrument for distinguishing what agents know, can learn with support, and have mastered.Its tiered results are intended to help identify reasoning faculties that need improvement.Source: §3.2 · pp. 6–7

4 Experiments

Experiments compare AgentFrontier fine-tuning with other datasets on Qwen3 models across four reasoning benchmarks. AgentFrontier leads across benchmarks and shows broad subject-level strength on HLE.

4.1 Experimental Setup

  • Training trajectories use a multi-round, tool-augmented ReAct process in which reasoning reports summarize evidence, assess progress, and specify the next action.The evaluation uses Qwen3 dense and mixture-of-experts models across HLE, ZPD Exam, R-Bench, and xBench-ScienceQA.Source: §4.1 · pp. 7–8
  • Each dataset supplied 12,000 accepted trajectories, with training normalized to 25,600 rounds per dataset, capped at 40,960 tokens per round, for three epochs.AgentFrontier showed a more balanced and diverse tool-use distribution, with substantial scholar, browser, and code tool use.Source: §4.1 · p. 8, Table 1 · p. 8

4.2 Main Results

  • AgentFrontier-trained Qwen3-30B-A3B models outperform all other training datasets across every evaluated benchmark, and the trend holds across other backbones.Figure 6 presents the fine-tuning dataset comparison across four benchmarks.Source: §4.2 · pp. 8–9, Figure 6 · p. 9
  • On HLE, AgentFrontier ranks first in six of eight subjects with Qwen3-8B and seven with Qwen3-32B, leading the next-best dataset by 3.8 and 3.9 average-score points, respectively.With Qwen3-30B-A3B, it leads in every subject and reaches a 25.67% average score.Source: §4.2 · p. 9

5 Analysis

The analysis examines alternative-solution availability, reasoning and tool-use gains, holistic training results, and a clinical research case. The results include stronger benchmark performance after continued pre-training and effective synthesis of evidence in the case study.

5.1 BoN Analysis: Validating Difficulty Richness & Potential for RL Training

  • 21.7% pass@1 rose to 40.7% pass@8 on 300 validation samples, indicating successful alternatives remain available beyond initial failures.The authors interpret this 19.0-point improvement as evidence of frontier-level difficulty and potential for reinforcement learning.Source: §5.1 · pp. 9–10, Figure 7 · p. 10

5.2 Why AgentFrontier Excels: Deconstructing the Gains in Reasoning and Tool-Use

  • 95% of problems are solved within 15 rounds, where the RFT dataset consistently outperforms fine-tuning baselines.The authors attribute the advantage to tasks that require knowledge fusion rather than single-source retrieval.Source: Figure 8 · p. 11, §5.2 · p. 11
  • 26.3% macroaverage conditional tool accuracy exceeds competitors' 21% plateau with a comparable number of interactions.The authors frame the gain as tool-use efficacy rather than increased call volume.Source: Table 3 · p. 11, §5.2 · p. 11

5.3 Holistic Agentic Training

  • AgentFrontier sets a new open-source state of the art and competes with larger proprietary agents; CPT adds 2.9 points on HLE, 7.0 on xBench-ScienceQA, and 2.0 on ZPD Exam.The final row isolates CPT's gains, including on ZPD Exam where RFT-only performance was near saturation.Source: Table 4 · p. 12, §5.3 · p. 13

5.4 Case Study

  • In a clinical case, AgentFrontier identified worsening on prednisone as an anomaly and used targeted literature search to validate a latent-infection hypothesis.OpenAI DeepResearch instead focused on negative findings and misdiagnosed Charcot Arthropathy.Source: §5.4 · p. 13

6 Related Work

Related work addresses synthetic agent data, multidisciplinary benchmarks, and deep-research agents. The paper positions its approach against limits in difficulty calibration, static evaluation, and localized retrieval.

Data Synthesis for LLM Agents

  • Prior synthetic-data work grounds tasks in external knowledge, but tasks may remain solvable through localized retrieval and lack principled difficulty calibration.Existing approaches use heuristic constraint addition or probes to distinguish reasoning from recitation [Patel et al., 2025; Yan et al., 2025].Source: §6 · p. 13

Multi-disciplinary Benchmark

  • As frontier models saturate static multidisciplinary benchmarks, the paper introduces ZPD Exam as a self-evolving evaluation framework [Phan et al., 2025].Humanity’s Last Exam remains fixed despite its expert curation.Source: §6 · p. 13

Deep-Research Agents

  • Deep-research agents automate multi-step search and reasoning for cross-domain information synthesis, using planning, tools, and web navigation [Jin et al., 2025].The literature includes both proprietary and open-source systems.Source: §6 · pp. 13–14

7 Conclusion

The conclusion presents ZPD-guided synthesis as a way to co-develop agent training resources and an evolving exam. AgentFrontier-30B-A3B achieves strong results on expert-level multidisciplinary benchmarks.

7 Conclusion

  • The framework co-generates targeted training resources and a self-evolving ZPD Exam to enhance and evaluate agentic reasoning.The conclusion describes the synthesis paradigm as based on Zone of Proximal Development theory.Source: §7 · p. 14
  • AgentFrontier-30B-A3B achieves state-of-the-art results on expert-level multidisciplinary benchmarks, surpassing significantly larger proprietary agents.Source: §7 · p. 14
  • The authors conclude that pedagogical data synthesis is effective for cultivating advanced reasoning in a data-efficient manner.Source: §7 · p. 14

Limitations and Future Work

The section identifies three current capability boundaries and describes the refinement process, its stopping rule, and the configured parameters and cost-accounting basis of the data engine.

Limitations and Future Work

  • Current scaffolding supplies complete solution trajectories, imitation learning under-explores diverse valid solutions, and the static toolset cannot create new tools.The proposed directions are graduated hints, reinforcement learning, and hierarchical tool composition or program synthesis.Source: Limitations and Future Work · p. 14

A Data Engine Details

  • The engine details document the data-generation procedure, hyperparameters, and computational costs to support transparency and reproducibility.The reported cost estimates use model tokenizers and API pricing from the experiment date.Source: A Data Engine Details · p. 15, §A.3 · pp. 15–16

A.1 Hyperparameter Configuration

  • The thematic-coherence threshold is 0.8 and the nearest-neighbor count is 10 for forming coherent seed combinations.The threshold controls chunk similarity for composite units; the neighbor count searches for coherent triplets.Source: §A.1 · p. 15
  • Refinement is capped at 30 iterations, verification uses three independent attempts, and question diversity uses a 0.7 similarity threshold.A new question is discarded when its similarity to an existing ZPD question exceeds the threshold.Source: §A.1 · p. 15

A.2 Agentic Refinement and Stopping Criterion

  • The refinement agent raises cognitive load through enrichment, and stops when the baseline cannot solve the question or 30 iterations are reached.Enrichment can add concepts, abstract principles, introduce realistic context, or make a qualitative problem quantitative.Source: §A.2 · p. 15

A.3.1 Cost of Agentic Refinement (Stage II)

The refinement stage averages 7.81 iterations per candidate and costs approximately $0.24, with reported token throughput and tool-call rates.

A.3.1 Cost of Agentic Refinement (Stage II)

  • Processing a single candidate data point involves an average of 7.81 refinement iterations, with 18,613.82 input tokens and 11,643.22 output tokens.Source: §A.3.1 · p. 16
  • Tool usage averages 0.70 Search, 0.61 Scholar, 1.21 Browser, and 0.94 Code Interpreter calls per candidate.Code Interpreter runs locally without API cost.Source: §A.3.1 · p. 16
  • $0.24 is the average refinement cost per candidate, based on an estimated LLM cost of approximately $0.234.The cost calculation uses 7.81 iterations and the reported input and output token rates.Source: §A.3.1 · p. 16

A.3.2 Cost of MKO Verification (Stage III)

MKO verification uses three attempts to check task solvability, with an amortized verification cost of about $0.54 and an end-to-end cost of about $0.78 per verified pair.

A.3.2 Cost of MKO Verification (Stage III)

  • Three independent MKO attempts verify that a candidate is solvable by an expert-level agent; a 33% success rate yields $0.54 amortized verification cost.Source: §A.3.2 · pp. 16–17
  • Verification averages 3.32 API calls, with 20,181.57 input and 24,169.88 output tokens per API call.Source: §A.3.2 · p. 16
  • Verification averages 1.30 Browser, 0.53 Code Interpreter, 0.50 Search, and 0.92 Scholar calls.Code Interpreter runs locally without API cost.Source: §A.3.2 · p. 16
  • $0.78 is the amortized end-to-end cost per verified PhD-level QA pair, comprising $0.24 refinement and $0.54 verification.The paper characterizes this as a quality-over-quantity approach and contrasts it with manual expert curation.Source: §A.3.2 · p. 17

B.1 Tools Implementation

  • The tool suite provides parallel web Search, scholarly metadata lookup, goal-directed Browser extraction, and sandboxed Python computation.Tools produce structured outputs for integration into the reasoning loop.Source: §B.1 · p. 17

B.2 Training Details

  • Supervised fine-tuning uses Megatron-LM, with separate hyperparameter tables for MoE and Dense models.Source: §B.2 · p. 17

B.3 More Results on on Fine-tuning Datasets

AgentFrontier fine-tuning achieves the highest reported overall conditional accuracy on ZPD-Exam and RBench-T, with strong tool-specific accuracy also reported.

B.3 More Results on on Fine-tuning Datasets

  • 87.6% on ZPD-Exam and 63.7% on RBench-T are the highest overall conditional accuracies reported for the AgentFrontier fine-tuning dataset.Source: §B.3 · p. 17, Table 7 · p. 18
  • 83.3% Code accuracy on ZPD-Exam and 78.6% on RBench-T are among the reported top-tier tool-specific results.The analysis covers tool usage and conditional accuracy after rejection-sampling fine-tuning across four datasets.Source: §B.3 · pp. 17–18, Table 7 · p. 18

C Case Study

The case study presents an AgentFrontier diagnosis example and reproducibility prompts, alongside instructions for refining question-answer pairs and a data pipeline for generating, escalating, and filtering training examples. It also addresses preserving mathematical and technical details during similarity filtering.

C Case Study

  • Negative initial radiographs and joint fluid do not rule out bone infection, while worsening during prednisone supports the chronic osteomyelitis diagnosis.The account explains that osteomyelitis can yield negative early radiographs and clean synovial fluid, and that immunosuppression may worsen an occult infection.Source: C Case Study · p. 22, Pathophysiological Mechanisms… · p. 23
  • AgentFrontier-30B-A3B correctly diagnoses chronic osteomyelitis, while OpenAI DeepResearch incorrectly identifies Charcot arthropathy.The stated ground truth is chronic osteomyelitis.Source: C Case Study · pp. 19, 23

D Prompts Used in Experiments

D.1 Evaluation Prompt

  • The evaluation prompt requires an extracted final answer, a correctness judgment against the reference, and a confidence score.It instructs the judge to focus on meaningful answer differences rather than solving the problem or introducing another answer.Source: Evaluation Prompt · p. 24

D.2 Similarity Filter Prompt

  • The similarity filter accepts a candidate QA pair only when its question and answer match the reference in specific information and technical details.Any difference in what is asked or answered means the pairs are not identical.Source: Similarity Filter Prompt · p. 24
  • The similarity-filter prompt instructs agents to pay special attention to mathematical expressions, symbols, and technical specifications.Source: Similarity Filter Prompt · p. 24

D.3 Agentic Refinement Prompt

  • Agentic refinement broadens QA pairs with retrieved background knowledge and raises reasoning demands through abstraction, synthesis, relationships, or analogies.The refined answer should incorporate the added information and explicitly state the abstract principle or relationship.Source: §D.3 · p. 25
  • The refinement prompt requires cross-validation of factual claims and computational questions solved with Python code and its output.Answers should add precise quantitative data, named entities, and references to authoritative sources; computational answers include executable code and a brief explanation.Source: §D.3 · p. 25
  • The pipeline generates seed QA pairs from thematically similar document chunks, then iteratively escalates and filters them for the ZPD dataset.It uses nearest-neighbor chunks and a thematic similarity threshold to form groups for seed generation.Source: §D.3 · p. 26
  • The ZPD filter rejects questions solvable by a less knowledgeable peer and keeps candidates when a more knowledgeable other produces a correct solution, subject to diversity filtering.The pipeline tests multiple solutions from the more knowledgeable other and checks question similarity against the existing ZPD set.Source: §D.3 · p. 26

Related papers

Loading 2510.24695v1…