ADE: Agentic Data Evolution Framework for Human-Centered Objectives

Yang Yu; Yilin Jiang; Zexuan Fei; Yiming Luo; Xingkai Song; Kaiyi Huang; Aimin Zhou; Xin Lin

cs.CLAug 24, 2026

ADE: Agentic Data Evolution Framework for Human-Centered Objectives

Authors: Yang Yu, Yilin Jiang, Zexuan Fei, Yiming Luo, Xingkai Song, Kaiyi Huang, Aimin Zhou, Xin Lin,

Organizations: East China Normal University, Shanghai, China · The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China · Shanghai Innovation Institute, Shanghai, China

Abstract

Aligning large language models to human-centered objectives is difficult when targets are non-executable and context-dependent, limiting reliable verification and scalable supervision. Although synthetic data expands coverage, weak verification shifts the bottleneck from generation to selection. Noisy signals destabilize iterative refinement and can cause silent regressions. We propose Agentic Data Evolution (ADE), a data-centric framework that organizes synthetic supervision as evolving data snapshots. ADE improves data snapshots through a closed-loop Observation-Variation-Selection (OVS) procedure, where a steady-state admission mechanism acts as a quality ratchet that conservatively gates updates for sustained cross-round improvement. We validate these improvements through complementary intrinsic trend tracking and extrinsic post-training evaluation. On DEV300, ADE raises the intrinsic win rate from 50% to 75.81% and the extrinsic win rate from 55.20% to 68.86%, consistent performance gains across diverse benchmarks. Blind expert evaluation further confirms this, with a 66.11% preference for evolved answers. These gains extend across post-training methods, model scales, and tasks beyond the target weakly verifiable educational objectives. Resources are available at https://github.com/ZeroLoss-Lab/Agentic-Data-Evolution.

Figures & tables

Figure 1: Overview of Agentic Data Evolution (ADE). ADE can be decomposed into three stages: (A) Construct $\mathcal{D}^{(0)}$ by concept–scenario–topic pipeline. (B) Evolve data $\mathcal{D}^{(t)}$ with Observation–Variation–Selection to obtain $\mathcal{D}^{(t+1)}$ . (C) Validate data quality with intrinsic validation, extrinsic validation, and human-calibrated validation.

Method	DEV300				Edu-Values					EduBench
Method	Overall	VO	AS	CI	Philo.	Literacy	Basic	Skills	Ethics	VO	AS	CI
Target model	55.20	58.10	56.00	51.00	67.05	71.51	57.14	72.48	60.61	7.26	6.85	6.85
Single-Round Methods
Best-of- $N$ (BoN)	58.03	66.25	66.25	47.42	65.52	70.97	56.19	70.72	64.07	7.87	6.91	7.04
SDFT	49.72	51.50	51.83	45.83	64.75	72.58	54.76	69.61	55.84	7.57	6.58	6.71
Self-Refine	65.82	71.75	73.42	52.58	61.30	67.74	50.95	63.66	54.55	6.71	6.94	6.71

Table 1: Extrinsic validation results on three benchmarks. For each column, the best result is shown in bold and the second best is underlined . VO (value orientation), AS (affective support), and CI (creative innovation). Philo. (professional philosophy), Literacy (cultural literacy), Basic (basic competencies), Skills (educational knowledge and skills), and Ethics (teachers’ professional ethics).

Metric	Annotator	Score
Win rate (%)	Expert 1	65.56
	Expert 2	63.33
	Expert 3	66.67
	Majority vote	66.11
Fleiss’s $\kappa$	Human–Human	0.7751
Cohen’s $\kappa$	Human–Judge	0.7149

Table 2: Human-calibrated validation on

\mathcal{D}^{(4)}

vs.

\mathcal{D}^{(0)}

over DEV300, with inter-annotator (Fleiss’s

\kappa

) and human–judge (Cohen’s

\kappa

) agreement.

Model	Cohen’s $\kappa$	Win Rate (%)
DeepSeek-V3.2	0.7149	67.22
GLM-5	0.5762	63.89
Kimi-K2.5	0.2645	76.67

Table 3: Automatic judge selection on DEV300.

Figure 2: Intrinsic validation on DEV300. Curves report pairwise preference win rates against

\mathcal{D}^{(0)}

, shown overall and by routed objective across evolution rounds.

Figure 3: Extrinsic validation on DEV300. Bars report pairwise preference win rates for models post-trained on selected snapshots, shown overall and by routed objective.

Dataset	Metric	$\mathcal{D}^{(0)}$	$\mathcal{D}^{(4)}$	$\Delta$
MATH-500	Accuracy (%)	54.60	55.80	+1.20
ToxiCN	F1 (%)	74.14	75.32	+1.18

Table 4: Out-of-domain verifiable tasks comparing the target model post-trained on

\mathcal{D}^{(0)}

and

\mathcal{D}^{(4)}

Figure 4: Transferability across post-training methods. Extrinsic validation on DEV300 comparing models post-trained with SFT vs. RL on ADE-evolved snapshots. Curves report generalized win rates for models, compared with the

\mathcal{D}^{(0)}

-trained reference.

Figure 5: Scale transferability at 72B. We compare models post-trained on snapshot

\mathcal{D}^{(4)}

(denoted Round 4) against

\mathcal{D}^{(0)}

(denoted Round 0). Bars report 100% stacked win rates on DEV300 across four views.

Configuration	R	Overall	VO	AS	CI
ADE	1	75.06	82.08	80.33	62.75
	2	75.22	83.50	80.92	61.25
	3	76.56	84.08	82.42	63.17
	4	75.81	82.42	83.08	61.92
w/o critics dims	1	53.92	54.00	54.83	52.92
	2	58.94	60.25	57.83	58.75

Table 5: Ablation studies across evolution rounds. We report win rates (%) for the overall score and three dimensions. Cell background colors indicate performance drops relative to ADE at the same round, with darker red denoting larger degradations. R is short for Round.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Attribute	Choices
activity_domain	Classroom Learning, Homework and Self Study, Exams and Assessment, Class Meetings and Character Education, Peer Interaction, Home School Communication, Online Spaces, Clubs and Projects
grade_band	Elementary School, Middle School, High School
task_type	Guided Discussion and Reflection, Case Based Analysis, Action Planning, Role Play, Project Design, Writing and Expression, Peer Collaboration, Classroom Facilitation and Guidance
artifact	In Class Discussion Questions, Reflection Log, Study Strategy Card, Class Meeting Plan, Communication Script, Project Brief, Peer Agreement, Writing Outline
role_view	student, teacher, parent, peer, no_asker
question_type	Decision Making, Process Focused, Evidence Seeking, Reflective, Perspective Taking

Appendix

Table 6: Structured control attributes.

Dimension	Cues
Value Orientation	Theme cues: Moral character, Rule of law literacy, Cultural literacy, Family values, Civic and national identity. Concept cues: Honesty and trustworthiness, Fairness, Privacy protection, Copyright and citation, Respect for diversity, Intergenerational communication, Boundaries of online speech.
Affective Support	Exam anxiety, Regulation of physiological arousal, Breathing and mindfulness, Attention management, Phone distraction, Procrastination, Action initiation, Self-efficacy, Emotion identification and labeling, Resilience after setbacks, Stress coping, Sleep and daily routines, Social anxiety, Interpersonal boundaries, Peer pressure, Self-talk.
Creative Innovation	Action cues: Break through, Reframe, Reexamine, Challenge, Integrate, Reason through. Domain cues: Social rule design, Nonlinear narrative, AI ethics, Entropy reduction. Goal cues: Interdisciplinary transfer and integration, Counterintuitive innovative problem solving, Sensitivity to technology ethics.

Appendix

Table 7: Dimension-constrained concept cues.

Method	Overall	Value Orientation	Affective Support	Creative Innovation
SFT	68.86	76.75	73.92	55.92
DPO	69.94	79.08	75.75	55.00
PPO	74.33	80.92	77.44	64.67
GRPO	80.91	86.33	82.00	74.33

Appendix

Table 8: Win rates (%) of different reinforcement learning strategies built upon the SFT baseline across three evaluation dimensions: value orientation , affective support , and creative innovation .

LLM Alignment

Large language model alignment studies methods for making models follow human intentions, values, and safety constraints across contexts.

See all 611 papers

CurateEvo: Data-Curation Evolving for Agentic Post-Training
Jul 7, 2026Dingzirui Wang, Xuanliang Zhang, Keyan Xu +2Training Data Curation LLM Agent Training
Harbin Institue of Technology
ANDES: Agent Native Data Evolving Synthesis Tool for Autonomous Instruction Alignment
May 31, 2026Zhengyang Zhao, Shengjie Ye, Lu Ma +3LLM Alignment LLM Agent Skill Learning
Peking University · Sichuan University, Chengdu
Know What You Don't Know: Agentic Data Synthesis by Evolving Generation Space for Closed-Loop Learning
Oct 8, 2026Wei Yang, Qing Ping, Youzhi Luo +4Synthetic Data Generation LLM Agent Trajectory Synthesis
University of Southern California · AWS AI

Jul 7, 2026J/K move · Enter open · S save

CurateEvo: Data-Curation Evolving for Agentic Post-Training

Dingzirui Wang, Xuanliang Zhang, Keyan Xu, Qingfu Zhu+1

Harbin Institue of Technology

Training Data Curation LLM Agent Training LLM Post-Training

Open paper AskPDF arXiv

Large language model (LLM) agents require post-training methods that can improve long-horizon decision making from environment feedback. However, existing agentic post-training pipelines often treat data curation as a fixed preprocessing step, focusing mainly on data augmentation while neglecting filtering, refinement, and adaptation to downstream failures. We propose CurateEvo, a failure-driven dynamic evolution framework for agentic post-training data curation. CurateEvo represents the curation strategy as executable code and iteratively rewrites it using failed trajectories from a held-out development set. At each epoch, the evolved strategy transforms a fixed raw corpus into supervised fine-tuning data, reinforcement learning data, and an inference-time memory bank. The evolution process first improves effectiveness by diagnosing recurring failure modes and augmenting, filtering, or refining data accordingly, and then improves efficiency by pruning redundant or low-utility training turns under a cost-aware objective. Experiments on ACEBench-Agent, BFCL-V4, and τ^2-Bench under both labeled and wild-data settings show that CurateEvo consistently outperforms prior curation methods, improving average scores by 3.2 and 2.7 points, respectively. Further analyses demonstrate that CurateEvo is compatible with different post-training recipes and substantially reduces curation overhead.

Abstract

LLM Alignment

Explore similar work

CurateEvo: Data-Curation Evolving for Agentic Post-Training

ANDES: Agent Native Data Evolving Synthesis Tool for Autonomous Instruction Alignment

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