PaperlayerView exact revision on arXiv

Source-linked AI summary

Know What You Don't Know: Agentic Data Synthesis by Evolving Generation Space for Closed-Loop Learning

Wei Yang, Qing Ping, Youzhi Luo, Yawei Wang, Abhishek Kumar, Panpan Xu, Jun Huan

arXiv:2610.12593v1cs.AIcs.MA

TL;DR

Interactive-agent synthesis must address a feedback-to-decision gap: learner weaknesses do not by themselves specify how future generation should change or whether the current synthesis space can represent them. EVOSYN uses a revisable synthesis state and specialized agents to turn feedback into reallocation or new generation controls. Across τ^2-Bench and CAR-Bench, it improves average task success by 13.3 points over the strongest synthesis baseline and produces structurally broader experience.

  • Problem

    Learner feedback reveals weaknesses but does not directly specify actionable synthesis decisions or whether the current synthesis space can represent them.

    Source: §1 · p. 2

  • Method

    EVOSYN uses a revisable synthesis space and specialized agents to translate learner feedback into effort reallocation or new controllable interaction factors, then grounds and verifies synthesized experience.

    Source: §1 · p. 2, §4.1 · p. 4, §4.2 · p. 5

  • Results

    13.3 points: EVOSYN improves average task success over the strongest synthesis baseline on τ^2-Bench and CAR-Bench, with consistent performance and synthetic-data-quality improvements reported in the conclusion.

    Source: §1 · p. 3, §6 · p. 9

  • Takeaways & Limitations

    Learner feedback reveals useful interaction factors missed by feedback-free synthesis, while iterative evolution broadens the structural support of synthesized experience.

    Source: §6 · p. 9

Abstract

from arXiv · show

The continued improvement of large language models increasingly depends on synthetic experience, yet effective data synthesis for interactive agents requires not only executable and verifiable trajectories, but also interactions that provide learning value for the current task agent. Existing methods increasingly incorporate environment feedback, data validation, or learner failures, but still face a feedback-to-decision gap: observed weaknesses do not directly specify how future synthesis should change, or whether they can be represented by the current synthesis space. To address this gap, we propose EvoSyn, a learning-feedback-driven multi-agent framework that organizes synthesis around a revisable state specifying what interaction factors can be deliberately generated and where synthesis effort should be allocated. A role-specialized multi-agent system translates learner feedback into synthesis revisions, while a Designer allocates synthesis directions and a Planner grounds them into executable experience. This process enables within-space adaptation toward weak or under-covered regions and synthesis-space expansion that promotes newly discovered factors into explicit generation controls. Experiments on $\tau^2$-Bench and CAR-Bench show that EvoSyn improves average task success by 13.3 points over the strongest synthesis baseline, while producing higher-quality and structurally broader synthetic experience and discovering controllable factors missed without learner feedback.

1 INTRODUCTION

Agentic synthesis must produce not only valid, executable tasks but also interactions useful to the current learner. EVOSYN addresses the feedback-to-decision gap by revising both how synthesis effort is allocated and which interaction factors can be generated.

1 INTRODUCTION

  • A valid, executable task may repeat behavior an agent already handles, while interactions exposing its remaining weaknesses are absent from training.Thus, agentic synthesis must consider learning utility as well as task validity.Source: §1 · p. 1
  • Observed learner failures identify weaknesses but do not directly specify future synthesis decisions or whether the current synthesis space can represent them.The paper calls this missing translation the feedback-to-decision gap.Source: §1 · p. 2
  • When weaknesses fit existing factors, synthesis should reallocate effort; when they do not, it must add actionable interaction factors as controls.The introduction notes that generating more tasks from the same categories may not address a capability gap.Source: §1 · p. 2
  • EVOSYN uses a revisable synthesis space and specialized agents to map learner feedback into generation decisions and verified experience.Its Reflector drives within-space reallocation or synthesis-space expansion.Source: §1 · p. 2
  • 13.3 points: EVOSYN improves average task success over the strongest synthesis baseline on τ^2-Bench and CAR-Bench.Analyses also find that learner feedback shifts allocation and reveals useful synthesis factors missed without feedback.Source: §1 · p. 3

2 RELATED WORK

Prior synthetic-data research spans offline construction, verifiability, learner-aware generation, and feedback for interactive agents. The related work leaves open how feedback becomes reusable synthesis decisions rather than valid samples or broader trajectory collections.

2 RELATED WORK

  • Synthetic-data work progresses from offline instruction and reasoning corpora to verifiability methods and learner-aware generation driven by failures or capability boundaries.Examples include instruction synthesis and evolution, answer checking, tool-integrated traces, and self-evolving task generation.Source: §2 · p. 3
  • Agentic synthesis uses environment, data-quality, and agent-evolution feedback, but how to convert these signals into reusable synthesis decisions remains open.Prior directions target executable tasks, repaired samples, or broader trajectory collections.Source: §2 · p. 3

3 PROBLEM FORMULATION

The formulation treats agentic synthesis as an iterative process that generates verified task–trajectory data and trains an interactive agent. Its central state records both what can be synthesized and how effort is distributed, then uses learner feedback to revise that state.

3 PROBLEM FORMULATION

  • At each round, synthesis generates verified, executable, successful task–trajectory pairs to update the task agent in an interactive tool-use environment.The environment includes policies, tool interfaces, state constraints, and an automatic evaluator.Source: §3 · p. 3
  • The revisable synthesis state S_r = (O_r, α_r) represents both the interaction properties available for synthesis and the effort allocated within that space.Valid data can provide little learning signal when it misses interaction patterns where the current agent remains weak.Source: §3 · p. 3
  • Learner feedback must be mapped to a new synthesis state, capturing the feedback-to-decision gap between observed behavior and actionable generation decisions.The closed-loop overview places this revision between learner diagnosis and subsequent generation.Source: §3 · p. 4, Figure 2 · p. 4

4 METHOD

EVOSYN organizes synthesis around a revisable environment-grounded space, uses learner feedback to revise that space or its allocation, and translates resulting directions into verified training experience. The updated agent is evaluated and its feedback informs the next synthesis round.

4 METHOD

  • EVOSYN maintains a revisable synthesis state that guides grounded generation and is updated from learner feedback in a closed synthesis–learning loop.The framework connects learner-guided synthesis to agent learning over successive rounds.Source: §4 · p. 4

4.1 REVISABLE ENVIRONMENT-GROUNDED SYNTHESIS SPACE

  • The environment-grounded synthesis space represents controllable interaction factors and serves both learner diagnosis and generation control.Its coordinates support performance and coverage measurement; the space can be expanded when a capability-relevant property is not represented.Source: §4.1 · p. 4

4.2 FEEDBACK-GUIDED SYNTHESIS EVOLUTION

  • The Reflector uses performance, coverage, and failure diagnostics to decide whether to reallocate effort or expand the synthesis space.New factors must be evidence-grounded, interpretable, finite-valued, non-redundant, and actionable; accepted factors become explicit generation coordinates.Source: §4.2 · p. 5

4.3 ROLE-SPECIALIZED MULTI-AGENT SYNTHESIS

  • The Designer allocates generation directions, the Planner grounds them in task blueprints, and the Generator, rollout teacher, and Verifier produce and check executable experience.Verification covers executability, tool-call validity, policy compliance, state consistency, and task success.Source: §4.3 · pp. 5–6

4.4 TASK-AGENT LEARNING AND CLOSING THE LOOP

  • The task agent trains on accumulated verified data, then held-out evaluation, coverage, and failure diagnostics form feedback for revising the next synthesis state.This closes the loop from synthesis through learning back to synthesis revision.Source: §4.4 · p. 6

5 EXPERIMENTS

Across τ 2-Bench and CAR-Bench, EVOSYN combines learner feedback with a revisable synthesis space to improve task performance and adapt generation toward changing capability gaps. The experiments assess overall effectiveness, component contributions, and how synthesis evolves across rounds and compares with existing corpora.

5.1 EXPERIMENTAL SETUP

  • The evaluation uses executable environments and automatic task-level success evaluators on τ 2-Bench and CAR-Bench, with Qwen3-8B as the default task-agent backbone.The study also uses Qwen3-14B and Qwen3-30B for task-agent scaling and several models, including Qwen3-30B, Qwen3-235B, and Claude Opus 4.5, as synthesis LLMs.Source: §5.1 · p. 6

5.2 RQ1: OVERALL EFFECTIVENESS

  • 77.8% average success rate: EVOSYN-Final-SFT with Opus 4.5 exceeds the reported synthesis baselines, and the benefit persists across synthesis backends and task-agent scales.EVOSYN-R1 reaches 70.9%; with Qwen3-14B and Qwen3-30B task agents, EVOSYN-Final-SFT reaches 80.9% and 86.0%, respectively.Source: §5.2 · p. 6, Table 1 · p. 7

5.3 RQ2: ABLATION STUDY

  • Removing learning feedback lowers success from 0.713 to 0.660 on Airline and from 0.726 to 0.613 on CAR-Bench, while component ablations also reduce data quality.Without the Reflector, CAR-Bench factor coverage falls from 0.893 to 0.803 and trajectory diversity from 0.938 to 0.800; without the Designer, CAR-Bench success is 0.540 and factor coverage 0.634. Single-Agent Synthesis has Airline trajectory diversity of 0.077 and CAR-Bench success of 0.406.Source: §5.3 · p. 7, Table 2 · p. 8

5.4 RQ3: HOW DOES LEARNER FEEDBACK EVOLVE THE SYNTHESIS SPACE?

  • After write-heavy R1 failures, EVOSYN introduces num write and shifts R2 toward high-write trajectories; later read-heavy weaknesses prompt num read and read–decide–act synthesis in R3.The feedback-conditioned Reflector recovers both subsequently adopted factors (2/2), versus 0/6 for the no-feedback control; promoted values achieve 99.6–100% exact accuracy and Spearman ρ = 0.997–1.000.Source: §5.4 · pp. 7–8, Figure 3 · p. 8

5.5 RQ4: EVOSYN ADAPTS SYNTHESIS TOWARD CAPABILITY GAPS

  • R2 synthesis shifts toward multi-write and high-write tasks after write-heavy weaknesses, and those targeted slices improve after training; R3 redirects toward read-heavy interactions as weaknesses change.The R3 focus includes multi-read, navigation, and read-then-set interactions, which also improve in the next evaluation.Source: §5.5 · pp. 8–9, Figure 5 · p. 9

5.4 RQ3: HOW DOES LEARNER FEEDBACK EVOLVE THE SYNTHESIS SPACE?

  • 0.726: EVOSYN rises from 0.607 to 0.655 to 0.726, while Static Multi-Round changes from 0.607 to 0.627 to 0.613.EVOSYN also has higher final factor coverage (0.893 vs. 0.787) and trajectory-pattern diversity (0.938 vs. 0.750), indicating gains beyond static sample accumulation.Source: §5.4 · p. 8, Figure 4 · p. 9

5.6 RQ5: HOW DOES EVOSYN DIFFER FROM EXISTING SYNTHETIC CORPORA?

  • 73.0%: under matched Airline data budgets, EVOSYN supplies the largest unique decisive interaction support, versus AReaL-SEA at 69.0%, APIGen-MT at 49.8%, and WRIT at 25.5%.Baseline-unseen structural support grows from 13.6% in R1 to 79.4% in later rounds.Source: §5.6 · p. 9

6 CONCLUSION

EVOSYN couples environment-grounded synthesis with task-agent learning, using feedback to reallocate effort across known interaction factors or expand the synthesis space with new factors. Its formulation treats synthesis revision as the decision of how feedback should change future experience, not merely as a record of learner behavior.

6 CONCLUSION

  • EVOSYN uses post-training feedback to reallocate synthesis effort within known factors and expand the space with newly discovered factors.The conclusion reports consistent improvements in task-agent performance and synthetic-data quality, and broader structural support than existing synthesis pipelines.Source: §6 · p. 9

A.1 PROBLEM FORMULATION

  • A synthesized example can be valid yet provide little useful experience if the learner has already mastered its interaction pattern.The formulation measures the resulting agent by held-out task success in the target environment.Source: §A.1 · p. 13
  • The synthesis state S_r = (O_r, α_r) separates which interaction properties can be generated from how effort is allocated across them.Generation realizes these high-level decisions as concrete tasks and trajectories in the environment.Source: §A.1 · p. 13
  • When a weakness is expressible in the current synthesis coordinates, the space stays fixed and generation effort is redistributed toward insufficiently covered or poorly targeted experience.This is within-space adaptation; the interaction factor is already controllable.Source: §A.1 · p. 14
  • When existing coordinates cannot represent a systematic weakness, reallocation alone cannot directly target it, so the synthesis representation must expand with new controllable factors.The expanded allocation is defined over the new set of synthesis directions.Source: §A.1 · p. 14
  • At each round, synthesis revision decides whether to change effort allocation within the current space or the factors the process can explicitly control.The formulation frames this choice as a feedback-conditioned process under a total synthesis budget, aimed at maximizing final task-agent performance.Source: §A.1 · p. 14

A.2 COMPARISON WITH EXISTING DATA SYNTHESIS FRAMEWORKS

Prior synthesis frameworks separately address environment grounding, data quality, or learner adaptation, while some closed-loop systems remain confined to existing task representations. EVOSYN makes the synthesis representation revisable and assigns specialized agents to revise, plan, ground, realize, and verify new experience.

A.2 COMPARISON WITH EXISTING DATA SYNTHESIS FRAMEWORKS

  • Earlier closed-loop systems incorporate task-agent feedback but remain confined to available task representations, predefined dimensions, or free-form generation.The comparison distinguishes environment-grounded executable synthesis from learner-aware reasoning methods that operate outside executable, stateful agent environments.Source: §A.2 · p. 14
  • EVOSYN is the only framework in the comparison that jointly supports all eight listed capabilities, including learning-effect feedback and explicit new-factor discovery.Table 3 marks capabilities as present only when explicitly supported and demonstrated; partial or implicit support is conservatively marked absent.Source: §A.2 · p. 14, Table 3 · p. 15
  • EVOSYN reallocates effort toward weak or under-covered known factors, but promotes a new reusable interaction factor when existing coordinates cannot represent the weakness.This changes both where synthesis is concentrated and what it can explicitly control, rather than only generating harder instances or resampling weak categories.Source: §A.2 · p. 15

A.3 CLOSED-LOOP ROLE-SPECIALIZED SYNTHESIS

  • A Reflector revises the synthesis state from learner feedback, while the Designer sets generation directions and budgets and the Planner grounds them in environment-consistent blueprints.The Generator and rollout teacher instantiate tasks and trajectories; a Verifier filters invalid supervision against environment and task constraints.Source: §A.3 · pp. 15–16

B EXPERIMENTS

  • The appendix describes additional experiments on scaling, Designer ablations, structural comparisons, feedback-driven reallocation and factor discovery, controllability, residual weaknesses, and case studies.The analyses trace how learner feedback is converted into synthesis-space revisions and subsequent changes in generated experience.Source: B EXPERIMENTS · p. 16

B.1 EXPERIMENTAL DETAILS

The experimental details specify interactive benchmarks, synthesis and training configurations, verification and comparison protocols, and intrinsic measures of data quality and structural coverage. Domain-specific ontologies constrain generated tasks, while held-out diagnostic feedback is separated from official test evaluation.

B.1 EXPERIMENTAL DETAILS

  • Evaluation uses held-out tasks from τ²-Bench Airline, Retail, and Telecom and CAR-Bench Base, with each benchmark’s binary task-success reward.Results are averaged over three random seeds unless otherwise specified.Source: §B.1 · p. 16
  • EVOSYN is instantiated with Claude Opus 4.5, Qwen3-30B, or Qwen3-235B synthesis backends, with the selected backend shared across synthesis roles.Claude Opus 4.5 also serves as the user simulator; Designer, Planner, and Generator use temperature 1.0 unless specified otherwise.Source: §B.1 · p. 16
  • Synthesis retains only trajectories that replay successfully and achieve reward 1, and the process runs for at most three rounds with at most one new factor introduced per round.Typical rounds use approximately 1,200–1,500 blueprint proposals; failed pseudo-ground-truth trajectories receive at most two repair attempts.Source: §B.1 · p. 16
  • The default downstream learner is Qwen3-8B, trained by full-parameter supervised fine-tuning on verified successful trajectories accumulated across rounds.The training recipe uses six epochs, a learning rate of 1 × 10^-5, and an effective global batch size of 16; Qwen3-14B and Qwen3-30B are also evaluated.Source: §B.1 · p. 17
  • Controlled comparisons hold the environment, verifier, task-agent initialization, training recipe, and evaluation harness constant while changing the synthesis mechanism.Final evaluation uses an official held-out set disjoint from the diagnostic set that drives synthesis revision; test labels do not guide synthesis or training.Source: §B.1 · p. 17
  • Intrinsic quality metrics assess coverage, factor balance, paired interaction coverage, and trajectory diversity on verified samples using realized tool actions.For CAR-Bench, joint factor-configuration coverage is used because marginal values saturate across methods; Airline, Retail, and Telecom tasks are built from domain-specific ontology axes.Source: §B.1 · pp. 17–18

B.2 ADDITIONAL OVERALL PERFORMANCE

EVOSYN’s accumulated multi-round experience raises average success across Qwen3 task-agent scales, and its 30B agent surpasses the strongest directly evaluated foundation-model average. Additional gains over first-round training persist at all three scales.

B.2 ADDITIONAL OVERALL PERFORMANCE

  • 77.8%, 80.9%, and 86.0% are the Final-SFT average success rates for Qwen3-8B, Qwen3-14B, and Qwen3-30B, respectively.The corresponding base-backbone averages are 39.4%, 48.6%, and 57.1%.Source: §B.2 · p. 21, Table 4 · p. 22
  • 86.0% for Qwen3-30B trained with EVOSYN exceeds Claude Opus 4.5’s directly evaluated average of 85.3%.The table reports direct foundation-model performance alongside EVOSYN task-agent scaling results.Source: Table 4 · p. 22, §B.2 · p. 22
  • 6.9, 7.6, and 5.7 points are the additional multi-round gains over R1 for Qwen3-8B, Qwen3-14B, and Qwen3-30B, respectively.The authors describe task-agent scaling and synthesis evolution as complementary.Source: §B.2 · pp. 21–22

B.3 ADDITIONAL ABLATION: WHY IS A DESIGNER NECESSARY?

The Designer separates global synthesis allocation from local task construction, preventing the task-space collapse observed when a Planner must do both. Structural comparisons then assess whether EVOSYN contributes unique support, order-sensitive variation, and balanced coverage.

B.3 ADDITIONAL ABLATION: WHY IS A DESIGNER NECESSARY?

  • Adding a Designer restores coverage across all six Airline task families and reduces the largest family’s share from 93.6% to 23.4%, unlike Planner–Generator synthesis.Without the Designer, two families receive no coverage; flight change takes 93.6% of accepted data.Source: §B.3 · p. 22, Figure 6 · p. 23

B.4 COMPARATIVE STRUCTURAL ANALYSIS OF SYNTHETIC DATA

  • The structural analysis evaluates unique interaction support, order-sensitive contrast, and how evenly synthesis mass is distributed across interaction structures.These three matched-budget analyses distinguish breadth, local variation, and distributional balance.Source: §B.4 · p. 23

B.4.1 UNIQUE AND EXPANDING STRUCTURAL SUPPORT

Under matched data budgets, EVOSYN contributes interaction structures absent from existing corpora, and its baseline-unseen support grows through synthesis-space evolution. It also produces more order-sensitive variation than AReaL-SEA on Telecom.

B.4.1 UNIQUE AND EXPANDING STRUCTURAL SUPPORT

  • 73.0% of EVOSYN’s distinct Airline structural signatures are absent from other matched corpora, the largest unique-support fraction among the compared methods.The corresponding fractions are 69.0% for AReaL-SEA, 49.8% for APIGen-MT, and 25.5% for WRIT.Source: §B.4.1 · p. 23, Figure 7 · p. 24
  • EVOSYN’s support absent from all matched baselines rises from 13.6% in R1 to 79.1% in R2, remaining high in R3 and R4.The share of individual examples with baseline-unseen structures also increases from 1.7% to 15.0%.Source: §B.4.1 · pp. 23–24

B.4.2 CONTRASTIVE SUPPORT FOR ORDER-SENSITIVE INTERACTION STRUCTURE

  • On Telecom, EVOSYN’s tool-order contrastive support is 53.1%, versus 9.6% for AReaL-SEA, a 5.5× increase.The comparison captures alternative decisive-action orderings among examples with similar action compositions.Source: §B.4.2 · p. 24

B.4.3 BALANCED STRUCTURAL DIVERSITY ACROSS INTERACTION PATTERNS

EVOSYN spreads synthetic experience across a broader and more balanced set of interaction structures, reallocates effort toward learner weaknesses, and expands its generation controls as new failure-relevant factors emerge. Across the reported CAR-Bench rounds, its structural support and task success improve, while comparisons of marginal learning gain remain descriptive.

B.4.3 BALANCED STRUCTURAL DIVERSITY ACROSS INTERACTION PATTERNS

  • On Retail, EVOSYN’s trajectory-shape evenness is 0.460, exceeding ToolMind’s 0.342, AReaL-SEA’s 0.148, and APIGen-MT and WRIT at 0.088 each.On Telecom, EVOSYN also exceeds AReaL-SEA on factor balance, interaction coverage, trajectory richness, and trajectory evenness.Source: §B.4.3 · pp. 24–25, Figure 9 · p. 25

B.5 TARGETED SLICE STATISTICS FOR CAPABILITY-GAP ADAPTATION

  • After write-heavy capability gaps in R1→R2, synthesis shifts toward read-intensive slices in R2→R3, and the upweighted slices improve after training.The reported gains for navigation, multi read, and read then set are 0.148, 0.205, and 0.103, respectively.Source: §B.5 · p. 26, Figure 10 · p. 26

B.6 NO-FEEDBACK CONTROL FOR SYNTHESIS-SPACE EXPANSION

  • The feedback-conditioned Reflector recovers both adopted factors, num write and num read, whereas the no-feedback control recovers neither of its six proposed dimensions.The comparison is 2/2 adopted-factor recovery with feedback versus 0/6 without it.Source: §B.6 · pp. 26–27, Table 6 · p. 27

B.7 CONTROLLABILITY OF EVOLVED SYNTHESIS FACTORS

  • Requested and realized num write values agree with 99.6–100% exact-match accuracy and Spearman ρ = 0.997–1.000 across R2 synthesis runs.The study reports this agreement across multiple domains and synthesis backbones.Source: §B.7 · p. 27, Table 7 · p. 28

B.8 CAPABILITY-GAP ALIGNMENT OF SYNTHESIS REALLOCATION

  • Across seven CAR-Bench slices, residual capability gaps align with next-round synthesis shifts (Spearman ρ = 0.964), unlike the negatively correlated Fixed-State baseline.All three slices upweighted in R3 improve after training; Figure 11 reports the gap-to-shift relationship.Source: §B.8 · p. 28, Figure 11 · p. 28

B.9 STRUCTURAL SUPPORT EXPANSION ACROSS SYNTHESIS ROUNDS

  • At matched final-round data volume, EVOSYN covers 1.32× more CAR-Bench sequences and 2.19× more Airline sequences than Static.Across rounds, CAR-Bench distinct sequences increase from 270 in R1 to 952 in R3 under EVOSYN, versus 415 for Static at R3.Source: §B.9 · p. 29, Figure 12 · p. 29

B.10 DO PROMOTED FACTORS CAPTURE PREVIOUSLY UNMODELED LEARNER WEAKNESS?; B.11 DO PROMOTED FACTORS CAPTURE FAILURE STRUCTURE BEYOND GENERIC COMPLEXITY?

  • Promoted factors retain failure information beyond existing synthesis coordinates and generic tool-count complexity, including 0.241 nats for Airline num write after conditioning on complexity.For CAR-Bench, num read retains 0.075 nats beyond generic complexity; the promoted factors are not merely proxies for trajectory length.Source: §B.11 · p. 30, §B.10 · p. 30

B.12 OBSERVED MARGINAL LEARNING GAIN ACROSS SYNTHESIS ROUNDS; B.13.1 CAR-BENCH

  • CAR-Bench EVOSYN success rises from 0.607 in R1 to 0.726 in R3, while Static Multi-Round reaches 0.613; the R2→R3 observed gain is +0.153 per 1K examples.The marginal-gain analysis is descriptive, not a causal estimate of per-example utility; the CAR-Bench trace links R1→R2 write-factor promotion and R2→R3 read-factor promotion to synthesis changes.Source: §B.12 · p. 31, §B.13.1 · p. 32

B.13.2 TAU-BENCH AIRLINE

On τ²-Bench Airline, EVOSYN’s feedback loop promotes new synthesis dimensions as different capability gaps emerge. Its task success improves across rounds while Static Multi-Round performance is unstable.

B.13.2 TAU-BENCH AIRLINE

  • After R1 reveals under-represented, difficult high-write interactions, EVOSYN promotes num write; later, it promotes computation for numerical or derived tool arguments.The Airline trace describes a shift toward more balanced computation-aware coverage in R3.Source: §B.13.2 · pp. 32–33, Table 15 · p. 34
  • EVOSYN Airline success rises from 0.648 in R1 to 0.713 in R3, while Static Multi-Round changes from 0.648 to 0.660.By R3, EVOSYN’s trajectory-pattern diversity is 0.769 versus 0.538 for Static Multi-Round.Source: §B.13.2 · p. 33, Table 14 · p. 34

C.1 SYNTHETIC DATA FOR POST-TRAINING AND SELF-EVOLUTION

The discussion traces a shift from offline synthetic corpora and verifiable reasoning toward learner-feedback loops, while identifying that agentic synthesis still lacks a revisable control space. The representative prompts and audit package operationalize role-separated task planning and evidence-based synthesis-space revision.

C.1 SYNTHETIC DATA FOR POST-TRAINING AND SELF-EVOLUTION

  • Synthetic post-training spans instruction-following data, verifiable reasoning, and learner-aware generation, but is often weakly coupled to the learner’s post-training behavior.Existing feedback commonly guides harder problems or concept variants within a relatively self-contained space.Source: §C.1 · p. 35

C.2 FEEDBACK-DRIVEN AGENTIC DATA SYNTHESIS

  • Agentic synthesis must produce feasible, executable, policy-compliant tasks with verifiable outcomes, yet existing feedback often targets validity or coverage rather than revising synthesis controls.The remaining gap is deciding which interaction factors deserve more budget and whether the current factors represent residual learner weaknesses.Source: §C.2 · pp. 35–36

D REPRESENTATIVE AGENT PROMPTS

  • The Designer allocates directions and budgets, while the Planner grounds them into task blueprints that downstream generation can realize as executable tasks.Directions specify targets, objectives, constraints, exclusions, budget, and rationale; blueprints select compatible dimensions and describe task types rather than fixed instances.Source: D REPRESENTATIVE AGENT PROMPTS · pp. 36–39

RUNTIME INPUT: SYNTHESIS AUDIT PACKAGE

  • The Reflector uses learner feedback to revise the synthesis space, considering failure-linked or under-covered factors while limiting each round to at most one new factor.Candidate factors should be measurable, interpretable, low-cardinality, non-redundant, and realizable; experimental factors can be promoted, retained, or deprecated.Source: RUNTIME INPUT: SYNTHESIS AUDIT PACKAGE · pp. 40–41

Related papers

Loading 2610.12593v1…