Kempner Institute opens Sept 2026 learning dynamics workshop

Harvard's Kempner Institute has opened submissions for a three-day workshop on learning dynamics in natural and artificial intelligence, running September 2 through 4, 2026 at the David Rubenstein Treehouse in Allston. The framing is deliberately cross-disciplinary. The Institute says the workshop will examine the principles governing learning and training dynamics across natural and artificial systems, and how biological and artificial systems each achieve, or fail to achieve, efficiency, flexibility, robustness, and generalization.

The program is organized around five threads: comparative learning in biological and artificial systems, training dynamics and inductive bias, evolution and inheritance and selection, multi-agent learning and adaptation, and foundations for next-generation learning systems. The listed speaker affiliations pull from both sides of that boundary, spanning the Kempner, Google Brain, Cold Spring Harbor, Meta FAIR, MIND Institute, Harvard, Stanford, NYU, Carnegie Mellon, and Hebrew University. Seeing FAIR and Cold Spring Harbor on the same agenda is not a common thing.

Why this matters if you are not going to Allston in September. Training dynamics has moved from a niche interpretability corner into something practitioners actually care about, because the specifics of how a system learns end up shaping which capabilities emerge from a given recipe. A room that puts neuroscience researchers next to industry-lab researchers on the same set of questions is a reasonable weathervane for where those questions are being reformulated.

Students, fellows, and early-career researchers can submit abstracts for talks or posters. The deadline is July 22, 2026 at 8:00 p.m. EST, with notifications going out by August 1. The honest caveat is that this is a call for participation, not a program. The specific talks and named speakers are not on the page yet, so anyone deciding whether to travel should treat the listed affiliations as an invitation set rather than a settled lineup. If the workshop delivers on its cross-disciplinary premise, the useful output will not be any single talk but the shared vocabulary that emerges between the neuroscience and machine learning halves of the room.