Multi-Agent Data Generation (MAGEN)

• MAGEN is a framework that coordinates specialized agents to collaboratively generate synthetic data with enhanced diversity and task alignment.
• It employs architectural patterns such as synchronous DAG pipelines, self-play loops, and decentralized message passing to optimize generation and control risk.
• Empirical evaluations show that MAGEN systems effectively improve dataset quality, privacy masking, and performance in applications like image synthesis, QA, and tool usage.

Multi-Agent Data Generation (MAGEN) denotes a class of data-generation systems in which multiple specialized agents coordinate, critique, verify, transform, or compete to produce datasets that are more diverse, controllable, robust, or task-aligned than those produced by a single monolithic generator. Across the recent literature, MAGEN appears in unsupervised image generation, privacy-preserving question answering set construction for retrieval-augmented generation (RAG), medical vision-language pretraining, automated math problem generation, real-world image dataset construction, synthetic tool-use trajectory generation, persuasion dialogue simulation, and multi-agent trajectory synthesis. The coordinating mechanisms range from synchronous directed acyclic graph pipelines and self-play loops to learned message passing, authoritative state-grounding, and decentralized peer-to-peer runtimes; in some works the acronym is explicit, while in others the architecture is identified as MAGEN because distinct agents are assigned complementary generation roles (Driouich et al., 26 Aug 2025, Ghosh et al., 2016, Li et al., 3 Dec 2025).

1. Definition, genealogy, and scope

An early explicit operationalization of the MAGEN idea appears in "Message Passing Multi-Agent GANs," where two DCGAN-style generators, a single discriminator, a shared message generator, and, in the conditioned variant, an encoder, exchange learned messages of dimension $d_m = 50$ in order to improve image generation through coupled cooperation and competition objectives (Ghosh et al., 2016). In that formulation, data generation is still the classical unsupervised task of sampling from a latent distribution, but the generative process is no longer attributable to a single agent.

Subsequent work generalized the concept from adversarial image synthesis to agentic workflow design. In "Diverse And Private Synthetic Datasets Generation for RAG evaluation: A multi-agent framework," the acronym MAGEN is not used, yet the design is explicitly described as fully embodying the MAGEN paradigm through a Diversity agent, a Privacy agent, and a QA curation agent arranged in a coordinated pipeline for synthetic QA construction (Driouich et al., 26 Aug 2025). In "APIGen-MT: Agentic Pipeline for Multi-Turn Data Generation via Simulated Agent-Human Interplay," MAGEN takes the form of a blueprint generator, execution-backed checkers, policy validators, an LLM review committee, a simulated human, and a simulated agent, all organized to synthesize verifiable multi-turn tool-use data (Prabhakar et al., 4 Apr 2025). In "State-Grounded Multi-Agent Synthetic Data Generation for Tool-Augmented LLMs," the paradigm is extended further to include a persona-conditioned user simulator, an agent under test, a state-grounded tool simulator, a multi-axis judge, and an authoritative state manager that never speaks in the conversation but governs world-state consistency (Khedar et al., 15 Jun 2026).

The scope of MAGEN is correspondingly broad. Some systems generate fully synthetic artifacts from scratch; others transform or curate existing corpora. This breadth is visible in medical caption recaptioning and verification, real-world image dataset construction, question-set diversification, synthetic segmentation pretraining corpora, and interaction trajectories for downstream planners. A plausible implication is that MAGEN is best understood not as a single model family, but as a systems pattern for decomposing data generation into specialized roles whose interaction structure matters as much as the base models they invoke.

2. Architectural patterns and coordination topologies

MAGEN architectures differ primarily in how they divide roles, how they exchange intermediate state, and where they place authority for acceptance or rejection. Representative systems illustrate the range of design patterns.

System	Domain	Agentic configuration
RAG evaluation framework (Driouich et al., 26 Aug 2025)	RAG QA datasets	Diversity $\rightarrow$ Privacy $\rightarrow$ QA curation in a LangGraph DAG
"Message Passing Multi-Agent GANs" (Ghosh et al., 2016)	Unsupervised image generation	Two generators, one discriminator, shared Msg/Enc, learned message passing
"APIGen-MT" (Prabhakar et al., 4 Apr 2025)	Multi-turn tool-use data	Blueprint generator, execution checker, policy checker, review committee, simulated human, simulated agent
"StateGen" (Khedar et al., 15 Jun 2026)	Tool-grounded conversations	User simulator, agent under test, tool simulator, judge, authoritative state manager
"DatasetAgent" (Sun et al., 11 Jul 2025)	Real-world image datasets	Demand Analysis, Image Process, Data Label, Supervision
Medical MAGEN (Li et al., 3 Dec 2025)	Dermatology vision-language pretraining	Diagnostic prior tool, Captioning Agent, Summary Agent, Verification Agent
"SynthSeg-Agents" (Wu et al., 17 Dec 2025)	Zero-shot weakly supervised segmentation	Self-Refine Prompt Agent, Image Generation Agent, relabeling classifier
"MADS" (Li et al., 30 Sep 2025)	Persuasive dialogue generation	User Agents, Dialog Agent, Optimization Agent
"Matrix" (Wang et al., 26 Nov 2025)	Scalable synthetic data runtime	Peer-to-peer agent actors exchanging serialized orchestrator messages

These examples show that MAGEN coordination can be synchronous and staged, as in the RAG evaluation pipeline; iterative and self-refining, as in math generation, geometry, and dialogue simulation; differentiable and end-to-end, as in MPM GANs; or message-driven and decentralized, as in Matrix. This suggests two broad orchestration families: workflows in which a central or semi-centralized plan explicitly sequences stages, and workflows in which the task state itself travels among agents and effectively becomes the locus of orchestration.

A second axis is authority structure. Some MAGEN systems treat one agent as the final arbiter: the Privacy agent in the RAG pipeline enforces masking constraints before QA synthesis, the Verification Agent in dermatology can emit "No definitive diagnosis," and StateGen’s authoritative state manager enforces a backend-is-truth invariant. Others rely on aggregation or filtering: a CEO or Aggregator in math generation, an Optimization Agent in persuasion simulation, a review committee in APIGen-MT, or an ML verifier in MAG-V. The resulting design space is closer to workflow engineering than to a single algorithmic template.

3. Objectives, control signals, and formal views

A recurring MAGEN formulation treats generation as constrained optimization over utility and risk. In the RAG evaluation framework, the end-to-end design is cast conceptually as maximizing semantic diversity while keeping privacy risk below threshold:

$$\max_{S \subseteq D} D(S) \quad \text{s.t.} \quad C(S) \le \epsilon,$$

with a Lagrangian relaxation

$\mathcal{L}(S) = D(S) - \lambda C(S).$

Operationally, the system uses k-means over 1536-dimensional text-embedding-3-small embeddings, cosine-based distances, silhouette-style clustering diagnostics, and topic-coverage entropy on the diversity side, while the privacy stage detects and pseudonymizes PII, PWI, and PHI through policy-driven masking before QA curation (Driouich et al., 26 Aug 2025).

In adversarial image generation, MAGEN appears as multi-agent coupling inside the objective itself. MPM GANs augment the standard GAN loss with competing or conceding generator objectives and a learned communication channel. Messages $m_{i \to j}^{(t)}$ are computed from a generator’s latest image, and in the conditioned variant they are further encoded together with the generator’s input noise and the previous incoming message. The receiver concatenates the incoming message with its own noise. The hinge term $f(x)=\max(x,0)$ makes only positive score differences contribute, and the paper interprets message passing as a regularizer because the generators are pushed toward complementary subspaces of the data distribution (Ghosh et al., 2016).

Other MAGEN systems replace differentiable coupling with routing, judging, or reinforcement signals. "Agentic Feature Augmentation" introduces a selector, generator, and router agent, where the router is trained with offline PPO over actions such as generate, select, and terminate, supported by short-term and long-term memory for state summarization and demonstration retrieval (Gong et al., 21 May 2025). Socratic-Geo uses Group Relative Policy Optimization (GRPO) for its Solver, assigning binary verifiable rewards to candidate solutions and computing group-relative advantages before PPO-style clipped updates (Jiao et al., 3 Feb 2026). These are not data-generation objectives in the narrow GAN sense; they are control policies over agentic data-transformation decisions.

Dialogue-oriented MAGEN systems often embed evaluation directly into the loop. MADS models user attitude evolution through a 16-state Chain-of-Attitude (CoA) space and estimates first-order transition matrices with entropy

$H(T_i) = - \sum_j T_{ij}\log T_{ij},$

so that diversity becomes a property of attitude trajectories rather than only of utterance surface forms (Li et al., 30 Sep 2025). StateGen formalizes generation via explicit world-state transitions

$S_t = T(S_{t-1}, a_t, r_t),$

with state diffs $\Delta S_t$ logged whenever a tool has write authority. In that setting, the generated datum is not merely a transcript but a transcript-plus-state-evolution object, suitable for training tool-augmented agents against backend-consistent traces (Khedar et al., 15 Jun 2026).

4. Representative applications and empirical evidence

The most direct MAGEN-style instantiation for dataset quality appears in RAG evaluation. Using the EU AI Act as input, the diversity/privacy/QA framework outperformed RagasGen and direct prompting across 10, 25, 50, 75, and 100 QA settings: LLM-as-a-Judge diversity ratings for the proposed system rose from 7.8 to 9.0, whereas RagasGen ranged from 7.0 to 8.1 and DirPmpt from 6.2 to 7.6. On privacy masking, evaluation on AI4Privacy PWI-Masking-200K, PHI-Masking-200K, and PII-Masking-200K reported per-entity-type accuracy between 0.75 and 0.94 depending on label and domain (Driouich et al., 26 Aug 2025).

In medical vision-language pretraining, MAGEN functions as a data-quality booster before representation learning. The dermatology system first identified low-quality image-text pairs by a cosine-similarity threshold of 0.7, flagging 183,934 candidates out of 403,563 Derm1M pairs for recaptioning. The pipeline then produced 133,930 verified enriched captions, while 50,004 cases received "No definitive diagnosis" and therefore retained the Captioning Agent’s initial description. In the key multi-agent ablation, average accuracy across PAD, F17K, SD-128, SNU-134, and Daffodil was 0.493 on original Derm1M captions, 0.491 with captioning alone, 0.531 with captioning plus the foundation-model diagnostic prior, and 0.544 with the full Caption + Tool + Verification pipeline (Li et al., 3 Dec 2025).

In automated question generation for mathematics, multi-agent inference-time computation improved task control only when coupled to explicit curation. Collective Consensus achieved rubric scores $\rightarrow$0 with Avg $\rightarrow$1, and Teacher-Critic Cycle reached Avg $\rightarrow$2; both outperformed baseline single-agent settings on difficulty matching and overall score. By contrast, the non-curated CC_RC and TCC_RC variants underperformed the curated versions, indicating that generation diversity without filtering or aggregation was insufficient (Karbasi et al., 6 Nov 2025). APIGen-MT reported a Phase 1 task configuration success rate of 70% with agentic feedback versus 28% without, a Phase 2 trajectory simulation success rate of 67%, and a final corpus of 3,820 validated multi-turn trajectories. Models trained on these data reached 56.2% overall on $\rightarrow$3-bench for xLAM-2-70b-fc-r and 78.19% overall on BFCL v3 for the same model family (Prabhakar et al., 4 Apr 2025).

For visual data construction, MAGEN spans both curation of real data and fully synthetic generation. DatasetAgent improved downstream performance after expanding real-world datasets: CIFAR-10 expansions raised average accuracy by approximately 0.52% across eight architectures, STL-10 expansions by approximately 0.41%, and VOC2007 expansion raised YOLOv8 from 76.3/45.8 to 81.0/49.7 in [email protected]/[email protected]:0.95 (Sun et al., 11 Jul 2025). SynthSeg-Agents, by contrast, generated a fully synthetic corpus for zero-shot weakly supervised semantic segmentation and obtained 57.4% mIoU in the ToCo pipeline and 60.1% in the Seco pipeline on PASCAL VOC 2012, plus 30.2% mIoU on MS COCO 2014 without real training images (Wu et al., 17 Dec 2025). Earlier, MPM GANs showed that message-passing multi-agent training improved SVHN feature quality from 22.48% error for a DCGAN discriminator to 17.1% for Different Noise CMP discriminator features and 15.2% when discriminator and message features were combined, while qualitative analyses indicated trait specialization across generators (Ghosh et al., 2016).

Tool-grounded conversational MAGEN has emphasized verifiability and control. StateGen evaluated 64,698 conversations across three production corpora and reported a tool-call hallucination score of 9.66/10 on a mixed corpus of 49,331 samples; persona variation induced a 15.8 percentage-point spread in goal achievement (Khedar et al., 15 Jun 2026). MAG-V began from 19 seed questions, generated 190 zero-shot queries, filtered them to 45, and trained a deterministic verifier on reverse-engineered alternate questions and trajectory features; its k-NN verifier reached 82.33% accuracy and 71.73 F1, improving accuracy by 11% over a GPT-4o judge baseline and matching GPT-4 on accuracy (Sengupta et al., 2024).

Persuasion and geometric reasoning provide further evidence that MAGEN can materially alter downstream capability. MADS increased organic traffic conversion rate from 1.83% to 2.24%, corresponding to a 22.4% uplift, and increased user intention rate from 4.53% to 5.82% (Li et al., 30 Sep 2025). Socratic-Geo started from 108 seed geometry problems, scaled its validated curriculum to approximately 2.5k synthesized examples by Stage 3, and reported an Overall score of 49.11 across benchmarks excluding GeomVerse; its Socratic-Generator achieved a GenExam relaxed score of 42.4%, surpassing Seedream-4.0 at 39.8 and approaching Gemini-2.5-Flash-Image at 43.1 (Jiao et al., 3 Feb 2026).

5. Relation to adjacent paradigms and common misconceptions

MAGEN overlaps with multi-agent learning, workflow orchestration, synthetic data generation, and automated data curation, but the cited literature draws several useful boundaries. In generative modeling, MPM GANs differs from multi-discriminator GANs, CoGAN, and InfoGAN or conditional GANs because the defining mechanism is learned inter-generator messaging together with explicit cooperation or competition objectives, not merely the presence of multiple networks (Ghosh et al., 2016).

MAGEN is also not equivalent to data synthesis from scratch. DatasetAgent constructs datasets from real-world images rather than artificial images; the dermatology MAGEN system selectively recaptions and verifies low-quality pairs rather than replacing the entire corpus; and the RAG evaluation framework transforms existing document collections into private, diverse QA sets rather than sampling free-form synthetic text unconstrained by source material (Sun et al., 11 Jul 2025, Li et al., 3 Dec 2025, Driouich et al., 26 Aug 2025). A related misconception is that MAGEN must always target a training dataset. Several systems instead generate evaluation sets, verification corpora, or test trajectories.

Nor is MAGEN reducible to the claim that more agents automatically improve output quality. The math problem generation study explicitly shows that non-curated agentic variants can underperform baselines, while curated variants improve difficulty matching and overall score (Karbasi et al., 6 Nov 2025). This suggests that agent specialization only becomes useful when coupled to an acceptance mechanism such as filtering, aggregation, abstention, state validation, or deterministic execution checks.

Verification design is another point of divergence. StateGen centers correctness on backend state and an authoritative state manager; MAG-V relies on reverse-engineered alternate questions plus trajectory-similarity features and classical ML; APIGen-MT uses executable environments, policy unit tests, LLM review committees, and rejection sampling (Khedar et al., 15 Jun 2026, Sengupta et al., 2024, Prabhakar et al., 4 Apr 2025). MAGEN therefore does not imply a single judge architecture. Depending on the domain, verification may be formal, environment-backed, classifier-based, committee-based, or partially heuristic.

Finally, orchestration need not be centralized. "Matrix: Peer-to-Peer Multi-Agent Synthetic Data Generation Framework" argues that a centralized orchestrator becomes a practical bottleneck at scale and instead serializes both control and data flow inside messages passed through distributed queues, reporting 2--15$\rightarrow$4 higher throughput under identical hardware resources without compromising output quality (Wang et al., 26 Nov 2025). This point is important because many MAGEN discussions implicitly assume a central controller, whereas the runtime literature treats decentralization itself as a first-class design variable.

6. Limitations, risks, and research directions

Across the surveyed systems, several recurring limitations appear. The RAG evaluation framework does not claim formal $\rightarrow$5-anonymity or differential privacy, so privacy preservation remains a detection-and-mask approach vulnerable to false negatives, over-masking, and semantic drift in QA generation (Driouich et al., 26 Aug 2025). The dermatology MAGEN pipeline inherits knowledge-base gaps, residual hallucinations, domain shift, and privacy or consent concerns attached to web-sourced medical imagery (Li et al., 3 Dec 2025). SynthSeg-Agents highlights prompt brittleness, VLM and CLIP biases, and sample-quality variation in fully synthetic image corpora (Wu et al., 17 Dec 2025). StateGen reports strong control over tool-call hallucination by construction, but still identifies judge calibration, overall-score transparency, and state-manager fidelity as open issues (Khedar et al., 15 Jun 2026).

Scalability introduces a second class of constraints. MPM GANs notes sensitivity to message dimension and noise bottlenecks, and its extension to $\rightarrow$6 generators in a fully connected communication graph yields $\rightarrow$7 communication overhead (Ghosh et al., 2016). Matrix identifies network bottlenecks, object-store pressure, at-least-once rather than exactly-once semantics, and the need for careful concurrency tuning at cluster scale (Wang et al., 26 Nov 2025). These observations suggest that MAGEN systems face a dual systems burden: they must manage both model quality and distributed-systems behavior.

A third limitation concerns evaluation itself. Socratic-Geo relies on LLM-as-judge and rule-based verification but does not present formal guarantees for visual fidelity beyond binary admission checks (Jiao et al., 3 Feb 2026). MAGS emphasizes that router quality depends on sufficiently diverse offline trajectories and on reward design that properly balances usefulness against proliferation or redundancy (Gong et al., 21 May 2025). More broadly, several papers call for richer ablations, better human calibration, explicit factuality filters, or domain-specific auditing. The common pattern is that MAGEN pipelines can generate large quantities of structured data, but the scientific difficulty shifts toward validating whether the generated structure corresponds to the intended target behavior.

Future directions named in the literature are correspondingly heterogeneous. They include adding explicit factuality filters and richer privacy metrics in RAG evaluation; expanding medical MAGEN to authoritative ontologies such as UMLS, SNOMED CT, ICD, and HPO; strengthening abstention and human-in-the-loop auditing in safety-critical domains; upgrading prompt, image, or routing agents independently in modular systems; using hybrid deterministic state updates where LLM state managers are insufficient; and extending geometry-style programmatic generation to other domains with executable semantics. This suggests that MAGEN is likely to remain a family of domain-specific design patterns rather than converging quickly to a single canonical architecture. Its unifying principle is not a shared model class, but the decomposition of data generation into interacting specialist roles whose coordination, verification, and memory structures are deliberately engineered.