PersonalPlan: 8B model for personalized multi-agent programming education
A new planner beats frontier LLMs with 3,043 personalized learning plans
PersonalPlan tackles the challenge of programming education by grounding multi-agent system (MAS) plans in learner profiles. The authors first created MAP-PPL, a dataset containing 3,043 query-profile-plan instances drawn from 1,730 Stack Overflow question groups and 2,738 distinct learner profiles. Each plan specifies agents, subtasks, executable steps, and prerequisite dependencies, providing a rich training ground for personalized instruction.
PersonalPlan itself is a two-stage planner: it first performs hierarchical supervised fine-tuning with separate LoRA adapters for profile-aware task decomposition and step dependency planning, then applies a Reward-Adaptive GRPO to encourage executable, personalized, and pedagogically scaffolded plans. In extensive tests against frontier LLMs and generic MAS frameworks, PersonalPlan’s 8B and 32B variants achieved state-of-the-art results in executability, personalization, and pedagogical quality, demonstrating that smaller, specialized models can outperform general-purpose giants in agent-student interactions.
- MAP-PPL dataset includes 3,043 personalized plans from 1,730 Stack Overflow groups and 2,738 learner profiles
- Uses two-stage training: hierarchical SFT with LoRA adapters then Reward-Adaptive GRPO
- 8B and 32B variants outperform frontier LLMs in plan executability and pedagogical quality
Why It Matters
Small, specialized planners can now deliver adaptive programming education at scale, reducing reliance on massive models.