InflEns
A Python research package that implements the influencer's game and routes queries across small language models using proportional allocation and trait-space geometry from Lovett & Fu.
InflEns (infl_ens) extends the influencer’s game to routing and learning for small language models.
The package builds a trait space from a calibration corpus, places model agents in that space, and allocates queries with the proportional rule (G_i(\mathbf{x}, b)) using a multivariate Gaussian influence kernel. It supports both fixed-position routing and closed-loop training where agents update positions from the traffic they receive.
Highlights
- The training mechanism inherits the game-theoretic properties of the influencer’s game.
- Router agents with proportional allocation and stability thresholds lead to increased interpretability of agent learning and specialization.
- Trait-space construction creates resource landscapes that agents can specialize over.