2026-05-22visionalignment
Not Too Generative, Not Too Discriminative: The Human Alignment Sweet Spot
Jorge Chang Ortega, Bastien Le Lan, Thomas Serre, Victor Boutin
Key claim
Hybrid models maximize human alignment in visual tasks.
This study investigates how human-like visual representations can be better understood through a balance of discriminative and generative learning. The key finding is that human alignment is maximized at intermediate points of this continuum, suggesting that a hybrid approach yields better results in vision tasks.
Novelty
8.0/10
The paper introduces Joint Energy-Based Models to explore the balance between discriminative and generative learning, providing new insights into human-aligned vision.
Reliability
7.0/10
The methodology is solid with evaluations across multiple benchmarks, though the architecture remains fixed.