PEAM Framework Enables Agents to Internalize Skills from Experience in Minecraft
A new framework called PEAM (Parametric Embodied Agent Memory) has been developed by researchers for Minecraft, enabling agents to learn skills through experience instead of just relying on inference during retrieval. This framework combines a deliberative large language model, which facilitates open-ended reasoning, with a rapid parametric module designed for reflexive actions. The fast module employs a multimodal Mixture-of-Experts LoRA architecture, featuring physically isolated adapters for each category, which supports ongoing learning without the risk of catastrophic forgetting. Failures are recognized as valuable training signals, with failure-correction trajectory pairs integrated through joint behavioral-cloning and contrastive objectives. Additionally, PEAM implements a parameterization-worthiness score to determine which experiences to internalize and includes a scale-free self-tri mechanism. The research was published on arXiv under ID 2605.27762.
Key facts
- PEAM stands for Parametric Embodied Agent Memory
- Framework transforms agent memory from inference-time retrieval to parameter-resident skills
- Uses a slow deliberative LLM and a fast parametric module
- Fast module is a multimodal Mixture-of-Experts LoRA architecture
- Per-category physically isolated adapters prevent catastrophic forgetting
- Failure-correction trajectory pairs are internalized via behavioral-cloning and contrastive learning
- Introduces parameterization-worthiness score for experience selection
- Paper published on arXiv: 2605.27762
Entities
Institutions
- arXiv