AI Credit Assignment via Cooperative Game Theory
A team of researchers suggests applying the least core solution concept from cooperative game theory to allocate credit for content generated by AI among creators whose intellectual property is included in the context window. They have created algorithms aimed at approximating the least core by utilizing innovative constraint seeding and separation techniques. In a task focused on web retrieval credit assignment, their methods achieve an approximation of the least core while requiring significantly fewer LLM calls compared to other approaches.
Key facts
- Proposes incentive-aligned mechanisms for in-context credit assignment
- Based on the least core solution concept from cooperative game theory
- Least core distributes value to ensure no subset of creators is significantly under-compensated
- Algorithms approximate the least core using constraint seeding and separation
- Tested on a web retrieval credit assignment task
- Approaches use orders of magnitude fewer LLM calls than alternatives
Entities
Institutions
- arXiv