Research map / Reinforcement learning / Visual and representation RL
GenAL: Generative Agent for Adaptive Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GenAL, an LLM-based generative agent framework (Global Thinking Agent + Local Teaching Agent) that recommends learning paths from item text and educational tools, evaluated on Junyi, ASSIST09 and a new TextLog dataset.
Problem
Learning path recommendation in adaptive learning relies on item indexing, needs retraining for new data distributions, and is unstable on sparse learner interaction logs.
Approach
Uses frozen LLMs as teacher-like agents that read item text, maintain a learner profile and reflect on recommendations (with knowledge tracing tools and a knowledge graph), replacing RL policy training with prompting-based planning.
Key results
- At 20 learning steps, GenAL achieves Ep 0.5692 on Junyi vs 0.4206 for the prior SOTA GEHRL.
- On ASSIST09, GenAL reaches 0.3665 Ep vs 0.1971 for GEHRL.
- On the sparse TextLog dataset (13 records per learner), GenAL attains 0.3804 while RL baselines fail to run.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GenAL instead of GEHRL | Learning path recommendation · Junyi | Ep: 0.4206 -> 0.5692 | Helped |
| Use GenAL instead of GEHRL | Learning path recommendation · ASSIST09 | Ep: 0.1971 -> 0.3665 | Helped |
| Use GenAL rather than ID-only baseline methods | Learning path recommendation · TextLog | Ep | Helped |
| Remove textual item content and recommend using item IDs only (ablation) | GenAL · Junyi, ASSIST09, TextLog | Ep | Hurt |
| Remove Educational Tools and their prior knowledge (ablation) | GenAL · Junyi, ASSIST09, TextLog | Ep | Hurt |
| Remove Reflection and stop updating the recommendation strategy and learner profile (ablation) | GenAL · Junyi, ASSIST09, TextLog | Ep | Hurt |
| Use Llama3-8B instead of Llama2-7B | GenAL · TextLog, 5 Steps | Ep: 0.2235±0.046 -> 0.2720±0.047 | Helped |
| Use GPT-3.5-turbo instead of Llama2-7B | GenAL · TextLog, 20 Steps | Ep: 0.3591±0.007 -> 0.3804±0.011 | Helped |
Surprising findings
- Baseline RL methods such as CSEAL and GEHRL cannot complete learning paths on the sparse TextLog dataset, reporting 0.00 in the path-length plots.
Limitations
- High inference cost of LLM-based recommendation.
- Inefficient use of external knowledge and remaining hallucinations.
- Simulator-based evaluation with negative Ep values on Junyi and ASSIST09 due to unseen items in the simulator.
Experimental setup
Datasets: Junyi ASSIST09 TextLog
Benchmarks: Junyi ASSIST09 TextLog
Metrics: Ep learning effectiveness
Baselines: KNN GRU4Rec DQN Actor-Critic CB RLTutor CSEAL GEHRL
Builds on: DKT KES simulator GEHRL CSEAL
Closest papers by meaning
- Agent4Edu: Generating Learner Response Data by Generative Agents for Intelligent Education Systems (AAAI 2025 · similarity 0.87)
- Developing LLM-Powered Trustworthy Agents for Personalized Learning Support (AAAI 2025 · similarity 0.86)
- Personalized Exercise Recommendation with Semantically-Grounded Knowledge Tracing (NeurIPS 2025 · similarity 0.86)
- DataEnvGym: Data Generation Agents in Teacher Environments with Student Feedback (ICLR 2025 · similarity 0.85)
- Learn-by-interact: A Data-Centric Framework For Self-Adaptive Agents in Realistic Environments (ICLR 2025 · similarity 0.84)
- AgentRecBench: Benchmarking LLM Agent-based Personalized Recommender Systems (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.