AI Cosmos
An agent infrastructure oriented toward situated learning. Through two principles, contextualized participation and the accumulation of practice assets, it places learners within an agent community with clearly divided roles and preserves reusable practice assets.
Timeline (last 90 days)
AI Cosmos is an agent infrastructure for situated learning.
AI Cosmos uses an asset-memory loop to save practice processes and outcomes as reusable assets, and annotates each asset with its provenance and dependencies.
The research team evaluated the effects of AI Cosmos through two preregistered randomized controlled experiments, examining the acquisition of legal reasoning and programming skills respectively.
The underlying LLM remained consistent across all experimental conditions.
In legal reasoning, AI Cosmos achieved larger learning gains than a conventional LLM.
In programming, on unseen programming problems, AI Cosmos achieved a higher level of mastery than the LLM-as-Teacher baseline.
The design philosophy of AI Cosmos is to redesign the collaboration structure between learners and the model, moving the model into a supporting role and leaving practice and transfer to the learner.
All Facts (7)
AI Cosmos is an agent infrastructure for situated learning.
50%UnverifiedAI Cosmos uses an asset-memory loop to save practice processes and outcomes as reusable assets, and annotates each asset with its provenance and dependencies.
50%UnverifiedThe research team evaluated the effects of AI Cosmos through two preregistered randomized controlled experiments, examining the acquisition of legal reasoning and programming skills respectively.
50%UnverifiedThe underlying LLM remained consistent across all experimental conditions.
50%UnverifiedIn legal reasoning, AI Cosmos achieved larger learning gains than a conventional LLM.
50%UnverifiedIn programming, on unseen programming problems, AI Cosmos achieved a higher level of mastery than the LLM-as-Teacher baseline.
50%UnverifiedThe design philosophy of AI Cosmos is to redesign the collaboration structure between learners and the model, moving the model into a supporting role and leaving practice and transfer to the learner.
50%