Source-linked AI summary
From Learner Behavior to Reusable Skills for Effective and Efficient Learner Simulation
Zijian Chen, Zheng Zhang, Miao Jia, Xingchen Hu, Weibo Gao, Linan Yue
TL;DR
Existing learner simulators repeatedly process growing histories to reconstruct learners, making acquired simulation capability costly to reuse across executors. Learner2Skill stores learner state and response patterns in reusable Skills, updates them with real interactions, and uses lightweight executor calibration; experiments report more faithful fine-grained simulation, lower token cost, and reuse across LLM executors.
Problem
Method
Learner2Skill distills learner state and recurring response patterns into persistent Simulation Skills, calibrates executor use without changing learner information, and evolves Skills as new real interactions arrive.
Source: §1 · p. 2, §5 · p. 12Results
Learner2Skill reproduces fine-grained learner behavior more faithfully while reducing overall token cost, and constructed Skills can be reused across different LLM executors.
Source: §1 · p. 2, §5 · p. 12Takeaways & Limitations
Learner-specific simulation capability can be decoupled from a particular LLM executor and preserved for continued reuse.
Source: §1 · p. 3, §5 · p. 12Skill Evolution uses only real interactions as update evidence; simulated responses are not treated as learner observations.
Source: §3.3 · p. 8
Abstract
from arXiv · showhide
Learner simulation aims to reproduce how a particular learner behaves on new tasks. Although Large Language Models (LLMs) can generate increasingly fine-grained learning behaviors, existing approaches often need to repeatedly process a growing interaction history to reconstruct the learner. This introduces additional context and inference costs and makes the acquired learner-specific simulation capability difficult to reuse across different LLMs. We therefore propose Learner2Skill, which externalizes the simulation capability acquired from historical interactions into a persistent and reusable Simulation Skill. The Skill captures the learner's current learning state and recurring response patterns, evolves as new real interactions arrive, and can be adapted to a new LLM through lightweight executor calibration without reconstructing the learner from scratch. Experiments show that Learner2Skill more faithfully reproduces fine-grained learner behavior while reducing overall token cost, and that the same constructed Skills can be effectively reused across different LLM executors.
1. Introduction
Learner simulation seeks to reproduce detailed task behavior, yet history-based agents repeatedly rebuild learner context and tie acquired capability to an executor. Learner2Skill externalizes that capability in a reusable Skill, with lightweight executor calibration and updates from new real interactions.
1. Introduction
- Existing simulators repeatedly process growing interaction histories, increasing cost, while their acquired learner-specific capability is difficult to reuse across executors.When the executor changes, historical interactions may need to be processed again to reconstruct an effective learner representation.Source: §1 · pp. 1–2
- Learner2Skill distills current learning state and recurring response patterns into a persistent Skill, retrieving full historical evidence only when finer-grained information is needed.Patterns include typical strategies, characteristic errors, corrections, and stopping tendencies.Source: §1 · p. 2
- Skill Construction, Executor Calibration, and Skill Evolution support construction, cross-executor reuse, and continued updating of learner-specific information.Calibration uses a small amount of real feedback from multiple learners to adapt a shared execution protocol without changing the constructed learner information.Source: §1 · p. 2
- Learner2Skill performs best on five of six fine-grained simulation metrics while reducing overall token cost.Evaluation uses 52,566 mathematics and physics interactions from 2,000 learners; generated responses also improve several computerized adaptive testing methods as additional training data.Source: §1 · p. 2
2. Related Work
Prior work spans human simulation, learner-specific agents, and reusable skills. Learner2Skill applies reusable-skill ideas to preserving an individual learner’s response behavior across LLM executors.
LLM-Based Human and Social Simulation
- Human simulation has shifted from response distributions and persona-driven societies toward grounded, person-specific, and temporally evolving behavior simulation.Examples include SocioVerse’s alignment with real-user populations and PersonaTwin’s construction of person-specific agents from rich individual information (Zhang et al., 2025a; Park et al., 2024).Source: §2 · p. 3
LLM-Based Learner Simulation
- Learner simulation expanded from predefined personas to agents grounded in authentic learning data, including individual-response simulators Agent4Edu and Embracing Imperfection (Gao et al., 2025, Wu et al., 2025).Other work studies alignment between simulated behavior and learner ability (Scarlatos et al., 2025, Que et al., 2026).Source: §2 · p. 3
Skills for LLM Agents
- Reusable-agent work extracts guidance or executable procedures from trajectories, while Learner2Skill preserves how a particular learner responds for reuse across LLM executors.The distinction is between preserving how an agent solves a task and preserving a learner’s response tendencies (Wang et al., 2024, 2023b; Zhao et al., 2024; Wang et al., 2026d; Gong et al., 2026).Source: §2 · pp. 3–4
3. The Proposed Learner2Skill Method
Learner2Skill externalizes historical learner-specific simulation capability into reusable Simulation Skills, supported by construction, executor calibration, and evolution as real interactions arrive.
3.1. Constructing a Reusable Simulation Skill
- Each Skill represents current concept mastery alongside recurring condition–response patterns, capturing what a learner can do and how they typically respond.Patterns are retained only when supported by multiple real interactions.Source: §3.1 · pp. 5–6
- Complete histories stay in an external database for retrieval, while maintaining only current state and supported patterns under a fixed pattern budget keeps Skills compact.The executor can retrieve recent or task-related interactions when the summarized Skill lacks needed detail.Source: §3.1 · p. 6
3.2. Calibrating a Skill to an Executor
- Calibration targets recurring Skill-use errors across learners and accepts protocol changes only when independent learner validation shows consistent improvement.The executor model parameters remain fixed, and only the shared execution protocol is updated.Source: §3.2 · pp. 7–8
- For a new executor, constructed Skills and histories are retained while a new shared protocol is calibrated; analysis establishes conditions for lower total token cost than existing agents.The paper describes reuse without reconstructing learner Skills from complete histories.Source: §3.2 · p. 8
3.3. Test-Time Skill Evolution with Real Interactions
- As real interactions arrive, Skill Evolution updates learning state and response patterns from new evidence while keeping the executor, calibrated protocol, and model parameters fixed.State adaptation fits a residual correction to recent real performance without updating cognitive-model parameters; new response patterns require repeated support.Source: §3.3 · pp. 8–9
4. Experiments
Experiments assess Learner2Skill’s simulation effectiveness and cost, cross-executor reuse, component contributions, and use of generated responses for CAT.
4. Experiments
- Experiments use LearnerTrace-2K’s 52,566 mathematics and physics interactions from 2,000 learners, compare five LLM simulators and four supervised predictors, and assess fidelity, prediction, and cost.Learner trajectories are chronologically split 60%/10%/30% for Skill Construction, Calibration, and Evolution.Source: §4 · p. 9
4.1. RQ1: Can We Simulate Learners More Effectively and Efficiently?
- Learner2Skill raised Final-Answer Fidelity from 29.00 to 40.13 and Prediction–Response Consistency from 60.48 to 71.54 versus LLM-FullHistory, and reached 63.43 in Outcome Prediction Accuracy versus DAISim’s 62.63.It achieved the best results in Attempt, Process, Final-Answer, Outcome Prediction, and Prediction–Response Consistency.Source: Table 1 · p. 10, §4.1 · p. 10
- 299.160M Total Cost was lowest among LLM-based methods, 15.2% below LLM-FullHistory; Consistency Efficiency was 0.00239, 39.8% above the second-best result.Learner2Skill also had the lowest Capability Acquisition Cost (182.590M) and Test-Time Cost (116.570M).Source: Table 1 · p. 10, §4.1 · p. 10
4.2. RQ2: Can Simulation Skills Be Reused Across Different LLM Executors?
- The same constructed Skills remained effective with Gemini, Qwen, and DeepSeek after executor-specific calibration, without repeating the 125.36M-token Skill Construction stage.Reuse total costs were 192.68M, 156.33M, and 140.77M tokens for Gemini, Qwen, and DeepSeek, respectively, versus 299.16M for the initial pipeline.Source: Table 2 · p. 11, §4.2 · p. 11
4.3. RQ3: What Makes Learner2Skill Effective?
- Response Patterns particularly improved Process Fidelity, while Learning State mainly improved outcome-related metrics; Executor Calibration and Skill Evolution provided further gains.The authors attribute these roles to patterns characterizing problem-solving behavior and state reflecting current capability.Source: §4.3 · p. 11, Table 3 · p. 11
4.4. RQ4: Can Learner2Skill Benefit Real Educational Applications?
- RQ4 evaluates generated learner responses in CAT using FSI, KLI, and MAAT with 5- and 10-exercise tests; Figure 3 labels post-augmentation F1 improvements.The supplied passages describe the evaluation setup and figure annotation but do not report the plotted outcomes.Source: §4.4 · p. 12, Figure 3 · p. 12
5. Conclusion
Learner2Skill preserves learner-specific simulation capability from historical interactions in reusable Simulation Skills. The conclusion reports improved fine-grained simulation at lower token cost and reuse across LLM executors.
5. Conclusion
- Learner2Skill organizes simulation capability acquired from historical interactions into persistent, reusable Simulation Skills, avoiding repeated learner reconstruction.The paper frames this as preserving learner-specific capability for continued reuse.Source: §5 · p. 12
- Skill Construction, Executor Calibration, and Skill Evolution support construction, cross-executor reuse, and continual updating of learner-specific information.Source: §5 · p. 12
- Experiments show more faithful fine-grained learner behavior with reduced overall token cost, and direct reuse of constructed Skills across different LLM executors.Source: §5 · p. 12
- Generated learner responses provide useful supervision for practical educational applications.Source: §5 · p. 12