Agent4Edu: Generating Learner Response Data by Generative Agents
for Intelligent Education Systems
Agent4Edu: 通过生成式代理为智能教育系统生成学习者响应数据

The learner profile module represents some overall learning features of human learners, which are typically stable and derived from long-term learning experiences. We configure each agent $agent_{u}$’s profile based on its corresponding learner $u$’s response data¹¹1Note that if zero-shot simulations are performed and user data is unavailable, the profile needs to be randomly generated.
请注意，如果进行零样本模拟且用户数据不可用，则需要随机生成个人资料。
学习者个人资料模块代表人类学习者的某些总体学习特征，这些特征通常稳定且源自长期学习经验。我们根据每个代理 $agent_{u}$ 对应的学习者 $u$ 的响应数据 ¹ 配置其个人资料。. Each agent’s initial configuration is divided into two categories: explicit practice styles and implicit cognitive factors.
每个代理的初始配置分为两类：显式练习风格和隐性认知因素。

Practice styles are statistical features explicitly derived from the available practice record $l_{u}$ of each learner $u$, such as learning activity (Baker 2001; Gao et al. 2021), practice diversity (Bi et al. 2020), success rate, and preference. Activity indicates learners’ enthusiasm for learning and provides clues for simulating their practice behaviors. For example, learners with higher enthusiasm for learning usually perform better. Mathematically, the activity level of learner $u$ is defined as $P_{\text{act}}^{u}=\frac{|l_{u}|}{|E|}$. Practice diversity reflects the knowledge coverage practiced by learners, represented as $P_{\text{div}}^{u}=\frac{|K_{u}|}{|K|}$, where $|K_{u}|$ is the number of knowledge concepts practiced by learner $u$. Higher diversity indicates greater curiosity in learners. Success rate correlates with the probability of learners answering questions correctly, making it another essential characteristic. The success rate for learner $u$ is mathematically represented as $P_{\text{suc}}^{u}=\frac{\sum_{y_{u,e_{i}}\in l_{u}}y_{u,e_{i}}}{|l_{u}|}$. Preference refers to the knowledge concepts that learners practice most frequently.
练习风格是明确从每个学习者的可用练习记录 $l_{u}$ 中提取的统计特征，例如学习活动（Baker 2001；Gao 等人 2021）、练习多样性（Bi 等人 2020）、成功率以及偏好。活动表明学习者对学习的热情，并为模拟他们的练习行为提供线索。例如，对学习更有热情的学习者通常表现更好。从数学上讲，学习者 $u$ 的活动水平定义为 $P_{\text{act}}^{u}=\frac{|l_{u}|}{|E|}$ 。练习多样性反映了学习者练习的知识覆盖范围，表示为 $P_{\text{div}}^{u}=\frac{|K_{u}|}{|K|}$ ，其中 $|K_{u}|$ 是学习者 $u$ 练习的知识概念数量。更高的多样性表明学习者更有好奇心。成功率与学习者正确回答问题的概率相关，是另一个重要特征。学习者 $u$ 的成功率从数学上表示为 $P_{\text{suc}}^{u}=\frac{\sum_{y_{u,e_{i}}\in l_{u}}y_{u,e_{i}}}{|l_{u}|}$ 。偏好是指学习者练习最频繁的知识概念。

Cognitive factors are implicit features studied in psychology (Baker 2001; Chen et al. 2024), which significantly impact learner $u$’s practice performance. We select problem-solving ability and knowledge proficiency (Cheng et al. 2024) for this study. Problem-solving ability is assumed to be stable during the learning process, while knowledge proficiency typically evolves with learning progress (Huang et al. 2020). Therefore, in the profile module, we only configure the ability factor $P_{\text{ab}}^{u}$, with knowledge mastery being considered in the subsequent memory module. To obtain implicit ability, we assign a psychological IRT model (Baker 2001) trained on the observed learner response records, as the tool for the agent, allowing it to infer each learner $u$’s ability factor from the response data $l_{u}$. The training and use of the IRT tool are detailed in Appendix B.
认知因素是心理学中研究的隐性特征（Baker 2001；Chen 等人 2024），它们显著影响学习者 $u$ 的实践表现。在本研究中，我们选择了问题解决能力和知识熟练度（Cheng 等人 2024）。问题解决能力在学习过程中被认为是稳定的，而知识熟练度通常随着学习进度而发展（Huang 等人 2020）。因此，在个人资料模块中，我们仅配置能力因素 $P_{\text{ab}}^{u}$ ，而知识掌握则在后续的记忆模块中考虑。为了获得隐性能力，我们为智能体分配了一个基于观察到的学习者响应记录进行训练的心理 IRT 模型（Baker 2001），作为其工具，使其能够从响应数据 $l_{u}$ 推断每个学习者 $u$ 的能力因素。IRT 工具的训练和使用在附录 B 中有详细说明。

Notably, we segment the values of each of the above features into several tiers in order to better prompt the generative agent inspired by (Wang et al. 2023d). For a detailed exposition, refer to Appendix A.1. Additionally, to ensure broad applicability and protect privacy, certain personal identifiers (such as name, gender, age, and occupation) are intentionally anonymized in this work (Zhang et al. 2023a; Li et al. 2023). While these attributes may help shape other types of agents, they are not primary factors affecting practice performance in education. Our approach, based on both behavioral practice styles and psychological cognitive settings, can support a comprehensive representation of real learners.
值得注意的是，我们将上述每个特征值划分为几个层级，以便更好地提示受（Wang et al. 2023d）启发的生成式代理。详细说明请参见附录 A.1。此外，为确保广泛适用性和保护隐私，本研究中有意对某些个人标识符（如姓名、性别、年龄和职业）进行了匿名化处理（Zhang et al. 2023a; Li et al. 2023）。虽然这些属性可能有助于塑造其他类型的代理，但它们并非影响教育实践表现的主要因素。我们的方法基于行为实践风格和心理认知环境，能够支持对真实学习者的全面表征。

The Memory module allows the LLM-based agent $agent_{u}$ to observe and summarize its corresponding learner $u$’s past practice experiences step by step. This module provides insightful clues to the agent for response simulation on unseen exercises. We follow the human learning mechanism (Atkinson 1968a; Cowan 2008; Huang et al. 2020; Wang et al. 2023d) to design three types of memories for each agent: factual memory, short-term memory, and long-term memory. Each memory is initially set to empty.
记忆模块允许基于 LLM 的智能体 $agent_{u}$ 逐步观察和总结其对应学习者 $u$ 的过往练习经验。该模块为智能体在未见过练习题上的响应模拟提供了有见地的线索。我们遵循人类学习机制（Atkinson 1968a; Cowan 2008; Huang 等人 2020; Wang 等人 2023d）为每个智能体设计三种类型的记忆：事实记忆、短期记忆和长期记忆。每种记忆最初均设置为空。

Inspired by human learning mechanisms, if an agent repeatedly practices similar questions or knowledge, their memory is strengthened (Huang et al. 2020). Therefore, we introduce an additional counter $f_{u,i}$ (initially set to 1) for each record $l_{u,i}$ in factual memory to track the number of times it has been reinforced, a simple yet effective method that has been successfully used in user preference simulation (Wang et al. 2023d). Formally, for each agent $agent_{u}$, assume it has observed $n$ factual memories is $M_{u}=\{l_{u,1},l_{u,2},\ldots,l_{u,n}\}$, then it is allowed to receive a new response record $l_{u,n+1}$. We first calculate the similarity between $l_{u,n+1}$ and each existing factual memory $l_{u,i}$ in the current memory $M_{u}$. The similarity between records can be defined as a metric that can be evaluated by LLMs, cosine similarity between text vectors, and other similar measures. In this case, we use the similarity relationships between the knowledge concepts involved in the records for the calculation. Specifically, we employ the statistical tool²²2https://github.com/bigdata-ustc/RCD
受人类学习机制启发，如果一个智能体反复练习相似的问题或知识，其记忆会得到加强（Huang 等人 2020）。因此，我们为事实记忆中的每条记录 $l_{u,i}$ 引入一个额外的计数器 $f_{u,i}$ （初始值设为 1），以跟踪其被强化的次数，这是一种简单而有效的方法，已成功应用于用户偏好模拟（Wang 等人 2023d）。形式上，对于每个智能体 $agent_{u}$ ，假设它已观察到 $n$ 条事实记忆，记为 $M_{u}=\{l_{u,1},l_{u,2},\ldots,l_{u,n}\}$ ，那么它允许接收一条新的响应记录 $l_{u,n+1}$ 。我们首先计算 $l_{u,n+1}$ 与当前记忆 $M_{u}$ 中的每条现有事实记忆 $l_{u,i}$ 之间的相似度。记录之间的相似度可以定义为一个 LLMs 可以评估的度量，如文本向量的余弦相似度以及其他类似度量。在这种情况下，我们使用记录中涉及的知识概念之间的相似关系进行计算。具体而言，我们采用统计工具 ² released by RCD (Gao et al. 2021) to determine whether two knowledge concepts are similar. If there is a similarity between the knowledge concepts involved in two records, the two records are considered similar. For similar records, we increment the counter for $l_{u,i}$ by 1 (i.e., $f_{u,i}\leftarrow f_{u,i}+1$), indicating that it has been reinforced by $l_{u,n+1}$, and then add $l_{u,n+1}$ to factual memory; otherwise, $l_{u,n+1}$ is directly recorded without any reinforcement. After processing and saving a new response record, factual memory triggers updating short-term and long-term memories.
由 RCD（高等人，2021）发布的工具用于判断两个知识概念是否相似。如果两条记录中涉及的知识概念存在相似性，则认为这两条记录相似。对于相似记录，我们将 $l_{u,i}$ 的计数器加 1（即 $f_{u,i}\leftarrow f_{u,i}+1$ ），表示它已被 $l_{u,n+1}$ 强化，然后向事实记忆中添加 $l_{u,n+1}$ ；否则，直接记录 $l_{u,n+1}$ ，无需任何强化。在处理并保存一条新的响应记录后，事实记忆会触发短期记忆和长期记忆的更新。

We emphasize that the agent can only save response records into factual memory but cannot directly retrieve it, thereby allowing the retention of all exercise textual information and responses without being constrained by the LLM’s context length limitations.
我们强调，该智能体只能将响应记录保存到事实记忆中，但不能直接检索它，从而允许保存所有练习文本信息和响应，而不受 LLM 的上下文长度限制。

Additionally, each factual record in long-term memory may be forgotten following the human forgetting curve theory (Averell and Heathcote 2011; Huang et al. 2020) that human memory decay starts rapidly and then gradually slows over time. We define a forgetting function associated timestamp $i$ and current observed step $n$, i.e., $g(l_{u,i})=\frac{1}{1+e^{-(n-i)}}$, to simulate human learners’ forgetting. For each factual record in the long-term memory $M_{u}$, it is forgotten if $g(l_{u,i})$ exceeds a predetermined threshold $\lambda$ and its reinforcement frequency $f_{u,i}$ in factual memory is then reset as $1$.

Overall, the factual response records are specific, while learning memory summaries are more general. By combining them, the agent can accurately perceive the learner’s practice process. Please note that traditional simulators (Piech et al. 2015; Zhao et al. 2023) can be regarded as owning the short-term memory but no long-term memory.

To help agents interact with the personalized learning environment, we introduce three memory operations:

To equip the agent with learner profiles and memory modules, enabling it to exhibit human-like problem-solving behaviors and responses based on current observations, we design a specialized action module for each agent within Agent4Edu tailored for personalized learning. This module encompasses three main categories of actions: