Research & Papers

MAVRL: Learning Reward Functions from Multiple Feedback Types with Amortized Variational Inference

Combines demonstrations, comparisons, and ratings into a single reward model, eliminating manual tuning.

Deep Dive

Researchers from ETH Zurich and University of Konstanz developed MAVRL, a new AI training method that learns reward functions from multiple feedback types (demonstrations, comparisons, ratings, stops) using amortized variational inference. The approach formulates reward learning as Bayesian inference, uses a shared encoder with feedback-specific decoders, and optimizes a single evidence lower bound. It outperforms single-feedback baselines, creates more robust policies, and provides interpretable uncertainty signals without manual loss balancing.

Why It Matters

Enables more efficient AI training by combining diverse human feedback, accelerating development of reliable autonomous systems.