PaperlayerView exact revision on arXiv

Source-linked AI summary

What Makes Good Agentic Data? An ACE Lens on Data Generation for LLM Agents

Xingshan Zeng, Zishan Xu, Boju Zhang, Yuzhou Wu, Lingzhi Wang, Jianghao Lin, Liangyou Li, Yasheng Wang, Lifeng Shang, Xin Jiang, Weinan Zhang, Yong Yu, Qun Liu, Weiwen Liu

arXiv:2608.27260v1cs.AIcs.CL

TL;DR

Agentic data generation remains fragmented across domains and often mixes candidate construction with verification and allocation. This paper introduces a factorized data model and ACE framework, concluding that useful scaling requires accurate, learner-relative, and behaviorally diverse experience rather than simply more samples.

  • Problem

    Existing agentic-data research is fragmented across domains and often conflates candidate construction with verification, selection, and allocation.

  • Method

    The paper factorizes agentic data as (E,q,τ,v) and frames generation as constrained distribution design balancing accuracy, learner-relative complexity, and diversity.

  • Results

    The literature supports prioritizing consistent, execution-grounded data, calibrating complexity to the learner and configuration, and measuring diversity through valid behavioral coverage rather than quantity.

  • Takeaways & Limitations

    Useful agentic-data scaling depends on allocating reliable, informative, and non-redundant experience as agents and environments evolve.

  • Takeaways & Limitations

    Executable checks and single success conditions cannot fully capture semantic accuracy, especially for trajectories generated by LLM simulators.

Abstract

from arXiv · show

LLM agents increasingly rely on generated interaction data to learn how to interact with external environments. Agentic data generation must maintain consistency among environments, tasks, interactions, and success signals while producing experience that is useful rather than merely abundant. Existing work spans many agent domains, but domain-centered organization and heterogeneous evaluation often obscure common generation mechanisms and conflate candidate construction with verification and selection. This work develops a two-level framework for the field. First, we represent agentic data as a common factorized object $(E,q,τ,v)$, comprising an environment specification, task signal, interaction realization, and optional verifier. We organize generation paradigms by their primary anchor and dependency structure. Second, we formulate generation as constrained distribution design through the Accuracy-Complexity-divErsity (ACE) lens. Accuracy establishes the feasible support of grounded and internally consistent data. Within this support, Complexity places learning mass relative to the capability of a declared learner and execution configuration, while divErsity controls coverage and redundancy of data. Using this framework, we explore how prior work verifies generated experience, constructs and calibrates difficulty, and expands behavioral coverage. The literature reveals a shift toward execution-grounded accuracy, learner-relative complexity, and diversity beyond surface variation or dataset size. We further discuss broader directions and emerging trends in agentic data generation through the ACE lens, including their implications for scaling, data sources, training regimes and adaptive learning. Overall, the central challenge is not simply to generate more data, but to continually allocate valid, informative, and non-redundant experience as agents and environments evolve.

1 Introduction

The paper frames agentic data generation as designing valid, learner-useful, and behaviorally diverse experience for LLM agents acting in external environments. It introduces a common factorization, a mechanism-oriented taxonomy, and the ACE objective to organize fragmented cross-domain work.

  • Motivation and framework: Agentic data must connect actionable environments, grounded tasks, valid interactions, and outcome or process supervision rather than merely generate instructions and responses.The framework represents each sample as d = (E, q, τ, v), with E defining the world, q the task, τ the realized interaction, and optional v providing supervision.
  • ACE lens: ACE treats generation as constrained distribution design: Accuracy defines valid support, Complexity targets learner-relative challenge, and Diversity promotes broad, non-redundant coverage.These requirements apply to consistency among E, q, τ, and v, probability mass relative to learner and execution configuration, and distinct environments, tasks, and behaviors.
  • Literature trends: The literature is shifting toward execution-grounded accuracy, learner- and configuration-relative complexity, and behavioral coverage beyond surface variation or dataset size.Agentic supervision is also expanding from fixed post-training trajectories toward pre-training, mid-training, and closed-loop experience.
  • Contributions: The paper synthesizes prior work across domains by evaluating how pipelines ensure Accuracy, construct and calibrate Complexity, and expand effective Diversity.Its contributions include a cross-domain data object and a mechanism-oriented taxonomy that makes consistency relations and measurement choices explicit.
  • Generation taxonomy: The factorization organizes generation by primary anchor and dependency structure, separating candidate construction from verification and selection across domains.The taxonomy includes forward, task-first, trajectory-first, and structure-first pipelines, while accommodating joint, iterative, and self-improving systems.

2 Formulation

The formulation represents agentic data as an environment, grounded task, interaction realization, and optional verifier, building on partially observable multi-turn interaction. This factorization distinguishes executable environments, trajectory records, and success assessment across agent domains and training paradigms.

  • 2.1 Agentic Multi-turn Interaction: Multi-turn tool-augmented interaction is formulated as a partially observable process in which policies act on observable history while environments retain latent state.Actions induce state transitions and new observations, with reward or evaluation optional.
  • 2.1 Agentic Multi-turn Interaction: A trajectory records observations and actions, including text responses, tool calls, user messages, and environment feedback, supporting both simple calls and stateful multi-turn interactions.The executable environment may maintain additional latent state for transitions and reward computation.
  • 2.2 Environment Parameterization: Environment specification E can range from textual tool schemas to an executable substrate containing state, actions, constraints, observation interfaces, and termination conditions.Executable environments can support new rollouts, unlike purely static interface specifications.
  • 2.3 From Interaction to Agentic Data: The task signal q specifies objectives and constraints, but is grounded only when its entities, required actions, and completion condition are meaningful under E.Tasks may be explicit instructions, target states, hidden intents, or signals revealed across turns.
  • 2.3 From Interaction to Agentic Data: The interaction realization τ describes how a task is attempted or solved, as a fixed demonstration for SFT or online sampling for reinforcement learning.Pipelines may retain successful, failed, exploratory, and recovery trajectories for one environment–task pair.
  • 2.4 A Common Data Object: The optional verifier v evaluates consistency or outcome through mechanisms such as schema checks, executable tests, terminal predicates, policy rules, proof assistants, LLM judges, or hybrids.This interface separates the generated object from any particular training paradigm and enables comparison across SFT, RL, and environment-based evaluation.
  • 2.4 A Common Data Object: Agentic data is modeled as d = (E, q, τ, v), linking an actionable environment, grounded task, realized interaction, and optional validity or success signal.This factorization is conceptual rather than a serialization requirement and requires compatible relations among all factors.

3 Generation Paradigms

This section organizes agentic data-generation pipelines by the order and anchor through which they construct environments, tasks, and interactions, separating forward from reverse generation. It then distinguishes candidate construction from acceptance and frames accepted-data shaping through the ACE lens.

  • Paradigm overview: The dominant factorization constructs an environment, then a compatible task, then an interaction, but task-first and trajectory-first pipelines reverse this order and relocate consistency recovery.The ordering determines which factors are directly controlled and where consistency must be recovered.
  • Paradigm overview: Generation paradigms specify how environments, tasks, and interactions are produced, whereas data objectives determine which generated instances are accepted or emphasized; the literature distinguishes forward and reverse generation.Figure 3 presents the forward-versus-reverse distinction, while ACE is introduced as a generation principle rather than a separate paradigm family.
  • Forward generation: Forward generation grounds tasks and trajectories in an existing environment, using real, synthesized, or programmatic environments, but risks cascading errors when environments are narrow or unreliable.Verification increasingly checks environment components, task feasibility, intermediate transitions, and terminal outcomes throughout generation rather than only at the end.
  • Reverse generation: Reverse generation changes which artifact anchors the others, while newer adaptive and self-evolving systems revise generators using accumulated experience, verification outcomes, model behavior, and coverage gaps.These systems move beyond independent samples from a fixed generator toward continually revised tasks, workflows, and strategies.
  • ACE lens: Under ACE, accuracy is an admission gate defining the feasible set, while complexity and diversity determine which valid data are most useful; ACE analyzes mechanisms rather than partitioning methods.If no candidate passes the accuracy gate, utility is zero, and a single pipeline may improve multiple ACE dimensions simultaneously.

4 Accuracy

Accuracy is the admission condition for agentic data: environment, task, interaction, and verifier/outcome must be jointly valid before complexity or diversity can make examples useful. Reliable verification increasingly combines execution-grounded signals with structural, process, semantic, and selective human checks, while accounting for cost, bias, loopholes, and residual uncertainty.

  • Accuracy as an admission condition: Accuracy gates ACE: environment, task, interaction, and verifier validity must all hold before complexity or diversity are evaluated.Invalid examples cannot become useful merely by being difficult or different.
  • Conjunctive validity: Validity requires both component-level checks and relational consistency among the environment, task, interaction, and intended outcome.For example, tasks must refer to available objects and operations, observations must follow executed actions, and verifiers must accept the intended terminal condition.
  • Verification stack: A layered verification stack assigns complementary roles to rules, execution and state signals, trajectory critics, semantic review, and selective human escalation.Human review is costly, so it is best reserved for ambiguous goals or disagreement among automated judges.
  • Verification evidence: Execution-grounded verification checks generated actions and trajectories through tool responses, state transitions, tests, or formal objectives rather than plausibility alone.Execution prevents fabricated observations and produces reproducible failure signals, but must be supplemented with process checks and semantic review because terminal success can conceal loopholes or unintended effects.
  • Limitations and trade-offs: Verification must balance coverage, cost, and bias because fixed or narrow checks can reward verifier-friendly shortcuts, reject valid alternatives, or miss semantic errors and side effects.Pipelines should use failure types for targeted regeneration, preserve alternative valid paths, and distinguish verifier acceptance from independently audited correctness.
  • Residual semantic uncertainty: Residual uncertainty remains when executable or simulator-based checks cannot capture incorrect dynamics, inefficiency, unsafe behavior, or inconsistency with user intent.This risk is especially pronounced when observations come from an LLM simulator rather than an executable environment.

5 Complexity

Complexity is learner-relative: it measures the difficulty induced by a grounded environment, task, and interaction protocol for a specified execution configuration, not merely length, tool count, or linguistic obscurity. Useful generation therefore calibrates behavioral difficulty within a moving learnable band while preserving accuracy and diversity.

  • Complexity definition: Complexity is model-aware and learner-relative, varying with the model, tools, environment protocol, verifier, sampling settings, and inference budget.Structural attributes such as horizon, dependency depth, branching, partial observability, and memory demand are controls and explanatory variables, not universal difficulty scores.
  • Complexity control: Useful examples occupy a moving learnable band between redundant easy cases and inaccessible all-failure cases, so generation may increase burdens or add scaffolding.Figure 6 depicts this band shifting with learner capability; for base-versus-agent comparisons, informative tasks are unsolved reliably by the base model but solvable with assistance.
  • Complexity factors: Agentic complexity decomposes across environment structure, task obligations, interaction requirements, and verifier-conditioned completion criteria.These factors include state and action dependencies, compositional or implicit goals, consequential serial or branching interactions, and requirements such as constraints, optimality, or forbidden actions [Tu et al., 2026, Lu et al., 2025b, Barres et al., 2025; Prabhakar et al., 2025, Shim et al., 2025, Zhang et al., 2026; Shen et al., 2024, Chen et al., 2025b, Xie et al., 2026b; Xu et al., 2025a, Wu et al., 2026b, Xia et al., 2026].
  • Behavioral calibration: The decreasing solver pass rate across verified recursive-synthesis rounds shows how structural growth can be checked against observed model difficulty [Li et al., 2026c].Difficulty assessment must follow schema, execution, semantic, and verifier checks, because invalid environments or infeasible tasks can otherwise masquerade as hard cases.
  • Limitations and adaptation: Complexity-aware selection must preserve coverage, avoid pushing beyond learnability, and periodically recalibrate against held-out domains as learners improve.Frontier methods seek mixed outcomes or stronger-scaffold advantages [Acikgoz et al., 2026, Wolf et al., 2026], while adaptive generators risk narrow feedback loops, saturation, and non-stationary distributions [Zeng et al., 2026b; Guo et al., 2025, Kang et al., 2026, Chen et al., 2025d].
  • Behavioral calibration: Structural proxies are inexpensive but unreliable predictors of behavioral difficulty, whereas calibration through executable rollouts is more informative but costly and configuration-sensitive.Longer or more constrained tasks may become easier through parallelization, narrowed search, or explicit decomposition; calibration remains sensitive to model, scaffold, inference budget, and stochasticity.

6 Diversity

Diversity measures the breadth of accurate, appropriately complex agentic data by covering distinctions that change what an agent can observe, decide, and do, rather than by counting samples, prompts, tools, or domain labels.

  • 6 Diversity: Diversity targets behavioral coverage: data should span distinctions that change what an agent can observe, decide, and do, not merely increase dataset size or surface categories.Pipelines pursue this through broader data sources and executable support, compatible recombination, reachable-experience exploration, controlled variants, and adaptation to remaining coverage gaps.

6.1 Diversity under the ACE Objective

Under the ACE objective, diversity is measured as a batch-level utility over accuracy-filtered data, emphasizing valid, behaviorally distinct, and learner-relevant coverage. Diversity should therefore avoid both redundant surface variation and novelty beyond the learner’s learnable region.

  • Diversity under the ACE Objective: Diversity is a batch-level utility computed over the accuracy-filtered set B_A, making validity a prerequisite for useful coverage.Inconsistent environments and unsupported trajectories enlarge the error space rather than the useful support.
  • Diversity under the ACE Objective: Structural diversity requires distinct behaviors, since renamed tools or paraphrased requests may induce the same behavior.
  • Diversity under the ACE Objective: Diversity is learner-aware: novelty for one model or corpus may be redundant for another, while unrestricted novelty can exceed the learnable region.

Diversity under the ACE Objective

The ACE diversity objective measures effective coverage across environment, task, and interaction factors after filtering for accuracy and useful complexity, while penalizing behavioral redundancy. It is a domain-specific design principle whose representations, weights, and targets must be specified per application.

  • Diversity under the ACE Objective: Diversity measures the breadth of valid, appropriately complex factor coverage after discounting behaviorally redundant samples.The factors are environment-, task-, and interaction-level variables, represented by Z_E, Z_Q, and Z_I.
  • Diversity under the ACE Objective: After accuracy filtering, entropy measures breadth and balance for each factor, weighted by relative importance, while λ penalizes behavioral redundancy.The filtered subset is B_A, and I_z denotes the useful complexity range for learner configuration z.
  • Diversity under the ACE Objective: The formulation is a design principle rather than a universal score, requiring domain-specific factor representations, estimators, redundancy measures, weights, learner configurations, and complexity targets.These choices should be specified separately for each domain.

6.2 Factor-Level Diversity

Factor-level diversity spans environment specifications, task signals, interaction realizations, and generator provenance, with meaningful variation defined by changed evidence, decisions, valid behaviors, or induced coverage rather than surface labels. These factors should be assessed through the capabilities, regimes, strategies, and source blind spots they expose.

  • Environment-specification diversity: Environment diversity varies actionable-world structure, including states, action spaces, observations, dynamics, rules, rewards, interfaces, verifiers, and other actors.Environment names or tool counts are weak proxies when preconditions, effects, information paths, and success conditions remain unchanged.
  • Factor-level diversity: Diversity has four complementary factors: environment, task, interaction, and generator/provenance.Figure 8 summarizes this factorization.
  • Task-signal diversity: Task diversity changes goals, capabilities, constraints, users, intent revelation, feasibility, temporal conditions, and difficulty, with clarification, refusal, partial completion, recovery, and success treated as distinct regimes [Shim et al., 2025, Wang et al., 2026e].Surface requests matter only when they alter evidence, decisions, or acceptable outcomes; broad difficulty coverage can still be narrow in goals, users, or capabilities.
  • Interaction-realization diversity: Interaction diversity covers alternative valid plans, action orders, tool substitutions, representations, horizons, assistance, recovery patterns, and outcomes for a fixed environment and task.Behaviorally distinct paths reveal strategy coverage, whereas repeated hesitation or accidental detours are rollout noise; failure traces also contribute.
  • Generator and provenance diversity: Generator and provenance diversity mixes corpora, teacher models, prompts, simulators, critics, decoding policies, and scaffolds to reduce blind spots, but must be evaluated by induced factor coverage rather than source names.Source heterogeneity is an intervention, not an outcome, so provenance should be recorded and its contribution assessed against repeated task or trajectory templates.

6.3 Expanding Agentic Diversity: Mechanisms and Domain Evidence

Agentic diversity is expanded through environment addition, composition, experience-first exploration, controlled variation, and adaptive coverage, but nominal scale is insufficient without valid behavioral coverage. Meaningful diversity is domain-specific: it must capture verified relations, state transitions, workflows, outcomes, and other behavioral units rather than surface variation alone.

  • Core mechanisms: Diversity pipelines add environments and components, compose compatible tools and workflows, derive tasks from reachable behavior, and perturb selected factors under controlled conditions.Graph- and blueprint-based composition reduces contradictions, while experience-first generation grounds tasks in observed behavior; controlled variation targets specific query, action, observation, and domain shifts.
  • Core mechanisms: Verified environments and marginally new capabilities matter more than nominal dataset scale, because broader distributions aid transfer while unreliable additions may be less useful.The relevant gain is a valid new capability, relation, state transition, or feedback structure.
  • Core mechanisms: Exploration and adaptation require multiple policies, seeds, resets, novelty or coverage signals, and broad replay because reachability and learner-frontier targeting can miss difficult, familiar, or foundational behaviors.Coverage constraints, stratified replay, and periodic broad exploration counter overfitting to transient model failures.
  • Domain evidence: Across domains, meaningful diversity is defined by behavioral units spanning states, workflows, users, strategies, outcomes, interfaces, dynamics, and verification regimes, not surface variation alone.Figure 9 summarizes these domain-specific instantiations, while Table 3 lists representative works.
  • Domain evidence: Coding diversity centers on verified problem–repository–feedback relations, while embodied and social diversity requires changes to transitions or rewards through coupled scenes, activities, policies, incentives, communication, and longitudinal outcomes.For embodied systems, procedural scene variation becomes useful when coupled with activities, physical constraints, task graphs, or policy changes; social simulations can couple personas with longitudinal behavior and outcomes.

6.4 Measuring Diversity

Agentic diversity requires staged, multidimensional measurement rather than a single score: studies should assess factor coverage, behavioral non-redundancy, ACE-conditioned validity and difficulty, transfer to held-out settings, and marginal learner utility. Coverage statistics should combine marginal and selected joint distributions, while evaluations report balance, failure modes, generalization gaps, and learner-relative value.

  • Measurement scope: Diversity measurement should first establish which environment, task, and interaction factors are covered, then test behavioral distinctness, validity, difficulty, transfer, and additional learner value.Each study should explicitly define which differences count as meaningful.
  • Factor Coverage and Balance: Coverage and normalized entropy are complementary: coverage detects represented categories, while entropy measures balance and reveals concentration that coverage alone can miss.Coverage may be high despite dominance by one category, whereas entropy does not identify absent important categories.
  • Factor Coverage and Balance: Marginal factor statistics should be supplemented with meaningful joint distributions to avoid template locking, where individually diverse factors occur only in fixed combinations.Full Cartesian coverage is usually impractical, so selected combinations should reflect expected deployment conditions.
  • Behavioral Non-redundancy: Category coverage overestimates diversity when samples require the same behavior, so comparisons should include action graphs, state transitions, dependency paths, clarification and recovery patterns, and final-state changes.These representations distinguish alternative strategies from paraphrases, renamed tools, or inconsequential rollout variation.
  • ACE-conditioned Coverage: Coverage should be measured after accuracy filtering and within declared complexity ranges, with factor-slice validity rates, failure types, and model-relative difficulty reported alongside diversity.Otherwise invalid samples or trivial tasks can make datasets appear broader than their useful coverage.
  • Transfer Coverage: Transfer evidence should use named held-out factors and report both held-out performance and its gap from matched in-domain settings [Lv et al., 2026, Wang et al., 2026d, Wu et al., 2026b, Li et al., 2026b].Relevant held-out factors include tools, schemas, states, policies, interfaces, user behaviors, dynamics, and horizons.

6.5 Diversity Tradeoffs and Limitations

Diversity improves structural coverage but creates tradeoffs with quantity, accuracy, learner-relative complexity, realism, controllability, and scale. Sustainable generation therefore requires coverage-aware evaluation, execution-grounded verification, adaptive allocation, and careful mixture management.

  • Diversity versus Quantity: Structural coverage can provide greater marginal value than repeated sampling from saturated modes, although additional samples still improve estimation and robustness in under-sampled regions [Chen et al., 2026a,b].Raw counts cannot distinguish support expansion from denser repetition, so scaling claims require coverage-controlled comparisons.
  • Diversity versus Accuracy: Broader recombination increases incompatible tools, unreachable goals, inconsistent states, and misaligned verifiers, requiring execution, procedural tests, constraints, and state-based checks [Liu et al., 2024, Song et al., 2026, Tian et al., 2026, Ivanov and Rana, 2026].A fixed checker may be narrower than the valid solution space and discard legitimate strategies.
  • Diversity versus Model-aware Complexity: A broad dataset may be poorly allocated for a learner: easy diversity adds little signal, while frontier-only generation can discard foundations and chase transient failures.Because useful complexity moves during training, generation needs frontier-focused allocation plus broad replay coverage, with diversity reported within difficulty bands.
  • Realism, Controllability, and Scale: Environment-free trajectory generation scales from API specifications alone but shifts accuracy assurance to simulator consistency and model-based filtering [Li et al., 2025e, Lee et al., 2026].Real systems provide authentic dynamics and long-tail behavior but are costly, unsafe, difficult to reset, and prone to temporal drift; learned simulators scale cheaply but may invent transitions.
  • Mixture Interference and Open-world Drift: Combining domains and generators creates interference from incompatible formats, action conventions, reward scales, and sampling frequencies, while normalization can erase domain-specific semantics [Chen et al., 2024, Zhang et al., 2024, Cai et al., 2025].Any finite mixture is only a snapshot because tools, interfaces, policies, and user behavior continue to change.

7 Discussion

The discussion reframes agentic data generation around effective support rather than raw quantity, hybrid real–synthetic allocation, earlier training exposure, and continual adaptation. Across these choices, ACE requires balancing grounded accuracy, learner-relative complexity, and behavioral diversity as agents and environments evolve.

  • Data scaling: Diversity-oriented scaling can outperform quantity scaling, while smaller collections of robust, behaviorally distinct environments can beat larger redundant or weakly verified sets [Chen et al., 2026a,b, Xu et al., 2026a, Tu et al., 2026].Effective support expands through valid experience, learner-useful complexity, and coverage across environments, states, tasks, policies, and interaction structures [Guo et al., 2025, Chen et al., 2025d, Wolf et al., 2026, Hao et al., 2026].
  • Data scaling: Static quantity scaling loses value as learners improve, because tasks saturate, complexity frontiers move, diversity becomes redundant, and new domains introduce accuracy and verifier gaps.Generation and allocation must therefore adapt dynamically with the learner rather than relying on fixed datasets.
  • Real and synthetic data: ACE favors hybrid real–synthetic allocation: real data anchor authentic interfaces and deployment needs, while synthetic data expand, balance, and calibrate verified coverage and model-aware difficulty.Real data can contain failures, hidden context, stale interfaces, privacy-sensitive content, and unverifiable outcomes; synthetic correctness depends on generator and environment fidelity.
  • Training regimes: Agentic pre-training and mid-training move interaction learning earlier, using executable or learned environments and interaction-relevant supervision from repositories, videos, documents, and relational structures [Chen et al., 2026c, Lu et al., 2025a, Xiong et al., 2026, Zhou et al., 2026a].Training units may be local state transitions or fragments rather than complete (E, q, τ, v) records; ACE then emphasizes transition correctness, source fidelity, fragment dependencies, and broad diversity.
  • Self-evolving agents: Self-evolving agents turn generation into a continual loop that targets capability gaps, adjusts difficulty, repairs rollouts, extracts skills, and preserves capabilities through replay.Dynamic ACE must prevent reinforced errors, track the moving capability frontier, and avoid concentration on recent failures or verifier-friendly strategies [Guo et al., 2025, Chen et al., 2025d].

8 Conclusion

The paper frames agentic data generation as jointly constructing environments, tasks, interactions, and trustworthy success signals, then shaping their distribution through the ACE lens. It concludes that effective generation requires learner-relative, behaviorally diverse, continually adapted experience rather than merely plausible or abundant samples.

  • Conclusion: Agentic data generation jointly constructs an actionable environment, grounded task, interaction process, and, when needed, trustworthy success signal, organized through factorized generation pipelines and ACE distribution shaping.The factorized account connects forward, task-first, trajectory-first, and structure-first pipelines.
  • Conclusion: Accuracy defines the feasible set through environment-task-interaction-verifier consistency, while Complexity is learner-relative and Diversity measures valid behavioral coverage rather than superficial variation or sample count.The framework treats these ACE dimensions asymmetrically: validity comes first, followed by calibrated difficulty and meaningful coverage.
  • Conclusion: The field is shifting from fixed post-training trajectories toward generated environments, broader supervision, and feedback-driven experience that continually discovers capability gaps and adapts allocation as learners change.Generation is presented as a continual process of constructing and verifying relevant experience, not a one-time production step.
Loading 2608.27260v1…