AgentFrontier Engine: ZPD-Calibrated Data Synthesis

• AgentFrontier Engine is an automated data synthesis framework guided by the Zone of Proximal Development (ZPD) that calibrates training tasks using an LKP–MKO asymmetry.
• It employs a three-stage pipeline—seed generation, agentic refinement, and ZPD-based filtering—to produce research-grade training data and dynamic benchmarks.
• Empirical evaluations demonstrate state-of-the-art performance with significant gains in multi-step reasoning and tool orchestration compared to existing baselines.

Searching arXiv for the cited papers to ground the article. AgentFrontier Engine is an automated, Zone of Proximal Development (ZPD)-guided data synthesis framework for training LLM agents on frontier-level reasoning and deep research tasks. Its central objective is calibration: to generate training and evaluation items that are neither trivially solvable nor effectively unsupervisable, but instead lie at the capability frontier where a base model cannot solve the task alone while a stronger, tool-augmented agent can. In the formulation introduced with AgentFrontier, the “Less Knowledgeable Peer” (LKP) denotes the base model without tools, and the “More Knowledgeable Other” (MKO) denotes a stronger tool-augmented agent; tasks unsolved by the former and solvable by the latter are treated as the most informative supervision for capability growth (Chen et al., 28 Oct 2025).

1. Conceptual definition and motivation

AgentFrontier Engine was introduced to address what the paper characterizes as a calibration problem in synthetic agent training: existing pipelines often do not target the actual capability frontier of the model being trained. Query-centric methods largely rephrase or mutate existing question-answer pairs, while document-centric methods derive questions from a single source or localized context; both regimes often remain within the model’s comfort zone and emphasize retrieval or shallow comprehension rather than cross-document synthesis, multi-step planning, or tool orchestration (Chen et al., 28 Oct 2025).

The framework’s pedagogical framing is explicit. ZPD is operationalized by distinguishing between tasks a model can solve unaided and tasks it can solve only with structured support. In AgentFrontier, the most useful supervision is not generic difficulty, but frontier difficulty: tasks beyond solitary solving yet achievable through tools, guidance, or demonstrations. This yields a dual training logic. Knowledge-intensive items already solvable by the LKP are routed toward continued pre-training, while frontier-level items solvable only by the MKO become post-training data for deep reasoning, tool use, and decomposition skills (Chen et al., 28 Oct 2025).

A common misconception is to treat AgentFrontier Engine as a general-purpose agent runtime or serving stack. In the paper’s own framing, it is instead a curriculum-generation and calibration pipeline. It transforms raw corpora into synthesized tasks, filters them by solvability, and then uses the resulting data both for training and for constructing a benchmark, the ZPD Exam (Chen et al., 28 Oct 2025).

2. ZPD formalization and solvability criterion

The engine does not introduce a single scalar “ZPD score.” Instead, it defines ZPD behaviorally through a solvability asymmetry between two agents. A task is in the model’s ZPD if the LKP cannot solve it alone, but the MKO can solve it with tools or guidance. This criterion is operationalized through automated judging and best-of-$N$ verification (Chen et al., 28 Oct 2025).

The LKP–MKO distinction is central. The LKP is the base model without tools. The MKO is a stronger agent with access to a tool suite and, by design, greater effective competence. The engine tests a candidate item by first checking whether the LKP can solve it unaided. If the answer is yes, the task is treated as too easy for frontier post-training and is instead sent to the continued pre-training pool. If the LKP fails, the system next evaluates whether the MKO can solve the task within $N=3$ independent attempts; acceptance into the frontier dataset requires that at least one of these attempts be correct (Chen et al., 28 Oct 2025).

This produces a sharp partition of synthesized data into three destinations. The first is $\mathcal{D}_{\text{pretrain}}$, for knowledge-intensive items already within the base model’s reach. The second is $\mathcal{D}_{\text{ZPD}}$, for frontier-level items that require assisted competence. The third is $\mathcal{D}_{\text{human}}$, for tasks that remain unsolved by the MKO and therefore require human review (Chen et al., 28 Oct 2025).

The same principle underlies the benchmark version, ZPD Exam-v1. There, the engine selects questions that are unsolvable by the baseline model in three unaided attempts but consistently solvable in three attempts with tools. The benchmark is therefore not a static set of manually authored hard questions; it is a dynamic frontier benchmark whose content is defined relative to current model capability (Chen et al., 28 Oct 2025).

3. Three-stage synthesis pipeline

AgentFrontier Engine is organized as a three-stage pipeline that transforms a raw corpus into calibrated training data (Chen et al., 28 Oct 2025).

Stage I: seed question generation for knowledge fusion

The process begins from a corpus of one million public documents. These documents are cleaned and chunked by Qwen3-235B-A22B, used as a chunking function $\Phi_{\text{chunk}}$, yielding semantic chunks $\mathcal{C}_{\text{chunk}}$ (Chen et al., 28 Oct 2025).

Rather than generating questions from isolated chunks, the engine constructs composite units from thematically related chunks. It builds a vector index over all chunks, retrieves $k_{\text{nn}}=10$ nearest neighbors for each chunk, searches local neighborhoods for coherent triplets $(c_i,c_j,c_k)$, and keeps only triplets whose pairwise semantic similarity exceeds the thematic coherence threshold $\tau_{\text{theme}}=0.8$ (Chen et al., 28 Oct 2025). These composite units are then passed to a generator model $N=3$0 to synthesize seed question-answer pairs.

The purpose of this stage is knowledge fusion. By constructing seed items from semantically coherent multi-chunk units rather than single passages, the engine biases generation toward cross-source reasoning rather than localized extraction. This suggests a deliberate move away from conventional synthetic QA and toward tasks that require integration.

Stage II: agentic refinement and complexity escalation

Seed QA pairs are then refined by an agent $N=3$1 with access to the tool suite

$N=3$2

For each current pair $N=3$3, the engine applies an escalation operator to obtain $N=3$4, with refinement running up to $N=3$5 steps (Chen et al., 28 Oct 2025).

The escalation process enriches tasks along four dimensions specified in the paper. Knowledge Expansion uses search, scholar, and browser tools to broaden factual scope and connect sources. Conceptual Abstraction lifts questions from surface facts to principles, relations, and analogies. Factual Grounding cross-validates claims across sources and tightens precision. Computational Formulation introduces quantitative reasoning or simulation through the code tool (Chen et al., 28 Oct 2025).

This stage is what turns seed QA into research-style tasks. The paper describes the process as iterative and self-reinforcing: each round’s output becomes the next round’s input. A representative example begins from a biomedical seed item and is escalated into diagnostic reasoning and then a practical computational challenge through web search and numerical validation (Chen et al., 28 Oct 2025).

Stage III: ZPD-based filtering and redundancy control

Refined items are then filtered by the LKP–MKO criterion. Automated judging is used to test whether the LKP can solve a candidate item unaided; the paper specifies GPT-4o for this stage. Items solvable by the LKP are routed to continued pre-training, while items that defeat the LKP proceed to MKO verification (Chen et al., 28 Oct 2025).

The MKO generates $N=3$6 independent solutions. A candidate is accepted into the frontier dataset if at least one solution is correct. After acceptance, the engine performs semantic redundancy filtering: a new QA item is discarded if it is too similar to an already accepted ZPD item, with threshold $N=3$7, using the Qwen3 embedding/reranking model for similarity measurement (Chen et al., 28 Oct 2025).

These controls impose three kinds of quality constraint: difficulty calibration, correctness verification, and diversity maintenance.

4. Data products, training pipeline, and benchmark construction

AgentFrontier Engine produces two major artifacts: synthesized training corpora and the ZPD Exam benchmark (Chen et al., 28 Oct 2025).

For training, the paper reports a 50B-token continued pre-training mixture that includes 1 million summarized text chunks and 20 million knowledge-intensive QA pairs, alongside a post-training set of 12,000 frontier-level QA trajectories (Chen et al., 28 Oct 2025). The underlying model is Qwen3-30B-A3B-Thinking-2507. Training is implemented with Megatron-LM. For the MoE model, the reported configuration is 3 epochs, max sequence length 40,960, batch size 256, learning rate $N=3$8, minimum learning rate $N=3$9, linear decay, tensor parallel 4, expert parallel 2, and pipeline parallel 1. For the dense model, the configuration is 3 epochs, max sequence length 40,960, batch size 64, learning rate $\mathcal{D}_{\text{pretrain}}$0, cosine decay, and warmup ratio 0.1. Reported generation hyperparameters are temperature $\mathcal{D}_{\text{pretrain}}$1 and top-p $\mathcal{D}_{\text{pretrain}}$2 (Chen et al., 28 Oct 2025).

The benchmark artifact, ZPD Exam-v1, is generated from a corpus of 30,000 recent scientific papers published during 2023–2025 across mathematics, computer science, physics, history, humanities, chemistry, biology/medicine, engineering, and geography (Chen et al., 28 Oct 2025). It contains 1,024 public questions plus a corresponding private set. The questions are open-ended short-answer, enabling automated grading (Chen et al., 28 Oct 2025).

The benchmark is described as dynamic because it can be regenerated as model capabilities advance. Its evaluative interpretation is organized into three zones: Zone 1: Intrinsic Competence for scores < 20, Zone 2: The Reasoning Bottleneck for scores 20–60, and Zone 3: Emergent Mastery for scores > 60 (Chen et al., 28 Oct 2025).

5. Empirical performance and analytical findings

The holistically trained model, AgentFrontier-30B-A3B (CPT+RFT), achieves 28.6% on HLE text-only, 93.4% on ZPD Exam-v1, 77.1% on RBench-T, and 61.0% on xBench-ScienceQA (Chen et al., 28 Oct 2025). The paper reports that CPT contributes additional gains over RFT-only of +2.9 on HLE, +2.0 on ZPD Exam, +2.7 on RBench-T, and +7.0 on xBench-ScienceQA (Chen et al., 28 Oct 2025).

In comparisons against baselines and agents including GPT-4o, Claude 4 Sonnet, Gemini 2.5 Flash, DeepSeek-V3.1-671B, OpenAI DeepResearch, Gemini DeepResearch, Kimi-Researcher, WebDancer, WebSailor, and WebShaper, the paper characterizes AgentFrontier-30B-A3B as state-of-the-art among open-source models on the evaluated benchmarks and as competitive with much larger proprietary agents (Chen et al., 28 Oct 2025).

Subject-level analysis on Humanity’s Last Exam is also emphasized. Within Qwen3 backbones, AgentFrontier fine-tuning yields Qwen3-8B as top in 6/8 subjects, Qwen3-32B as top in 7/8 subjects, and Qwen3-30B-A3B as best in all 8 subjects. The reported HLE averages are 18.80 for Qwen3-8B + AgentFrontier, 23.82 for Qwen3-32B + AgentFrontier, and 25.67 for Qwen3-30B-A3B + AgentFrontier (Chen et al., 28 Oct 2025).

The paper also argues that AgentFrontier improves orchestration quality rather than merely increasing tool invocation frequency. On HLE, for Qwen3-30B-A3B, the reported overall conditional tool accuracy is 26.3%, compared with competitors around 20–21% (Chen et al., 28 Oct 2025). Its reported tool-use profile is comparatively balanced: Search 0.73, Scholar 0.89, Browser 1.32, Code 0.63, with Overall rounds 4.57 (Chen et al., 28 Oct 2025).

Two additional analytical results are notable. First, the paper reports that 95% of problems are solved within a 15-round horizon, with AgentFrontier’s advantage strongest in the early-to-middle multi-step window (Chen et al., 28 Oct 2025). Second, on a held-out validation set of 300 samples, the paper finds pass@1 = 21.7% and pass@8 = 40.7%; the resulting 19-point gap is interpreted as evidence of substantial latent solution diversity and as an argument for future RL-style optimization beyond single-reference fine-tuning (Chen et al., 28 Oct 2025).

6. Position within the broader agent-systems landscape

AgentFrontier Engine belongs to a family of recent systems that reposition the “frontier” problem in agent research, but its role is distinct from runtime control, query execution, benchmark-only design, or domain-specific workflow orchestration.

A useful contrast is Hyperparam, a browser-native, JavaScript-first query engine for AI-native client applications that analyzes unstructured text at the point of use through Hyparquet, Squirreling, and Icebird; there, the central problem is interleaving analytic operators with model-based text interpretation in a tiny JS-native stack (Daniel, 27 May 2026). AEGIS addresses a different boundary entirely: it is a pre-execution firewall and audit layer for tool-using agents, interposing between tool proposal and tool execution with deep string extraction, content-first risk scanning, JSON-Schema policy validation, human escalation, and tamper-evident logging (Yuan et al., 13 Mar 2026). Frontier-Eng is not an engine product but a benchmark for self-evolving engineering agents working in propose–execute–evaluate loops under fixed budgets and hard feasibility constraints (Chi et al., 14 Apr 2026). Foam-Agent 2.0 is a modular multi-agent framework for OpenFOAM automation with six specialized agents, MCP-exposed services, hierarchical multi-index retrieval, dependency-aware generation, and iterative repair (Yue et al., 17 Sep 2025). Terminus-4B studies a still narrower setting: a compact execution subagent for terminal-heavy coding workflows, showing that a specialized 4B model can preserve benchmark performance while reducing main-agent token usage by up to approximately 30% in some settings (Garg et al., 4 May 2026).

Against this backdrop, AgentFrontier Engine is best understood as a frontier-data engine rather than an online control plane or embedded application substrate. Its core contribution is to turn capability-frontier calibration into a systematic synthesis procedure: generate tasks from real knowledge sources, escalate them into research-grade problems with tools, verify them by LKP–MKO asymmetry, and use the result both as a training curriculum and as a renewable benchmark (Chen et al., 28 Oct 2025).

The paper is explicit about its limitations. Current supervision is binary scaffolding, in the sense that MKO guidance is all-or-nothing rather than graded; rejection sampling fine-tuning captures only one or a few trajectories rather than broader exploration; and the toolset is static, so the agent can use tools but cannot create new ones (Chen et al., 28 Oct 2025). These limitations define the next research questions: graduated scaffolding, tighter integration with RL, more adaptive tool use, and dynamic tool creation.