DIAYN Agents With Intrinsic Motivation Match Drosophila Behavior Across 20 Algorithms

A fly wandering through a maze with no goal is actually making decisions in advance. Whole-brain activity recordings of Drosophila melanogaster show that neural representations of spontaneous direction choice appear more than 10 seconds before the turn itself. Aran Nayebi's group at NeuroAgents Lab, together with Declan Kutscher, Reece Keller, Alyn Kirsch and Felix Pei, asked a question: which learning objective gives an AI agent the same behavior and the same internal dynamics? They ran 20 intrinsic motivation algorithms in a fly simulation. DIAYN won.

A fly that decides 10 seconds before it turns

Animal autonomy usually looks like a puzzle. Nobody gives the fly external signals or promises a reward, yet it still explores space, changes direction and behaves unpredictably, but not chaotically. A classic trained model does not explain this: in RL, the agent almost always needs a reward function written by a human.

The authors built the experiment in two parts. The first is real recordings of the fly's brain, which show how a future decision is encoded ahead of time. The second is a neuromechanical digital twin: a physically realistic model of the fly's body in which the agent controls movement. The policy is hierarchical. The upper level chooses from a discrete library of low-level motor primitives, which is consistent with ethological theories of "behavioral syllables", the units from which animal behavior is thought to be built.

Twenty ways to want something without a reward

The agents were trained with no external reward at all. The researchers used 20 intrinsic motivation algorithms from three families, and this variety is what makes the work substantive: it compared not two or three ideas but a whole spectrum of hypotheses about what drives autonomous behavior.

• Curiosity methods: the agent seeks out what is new and unpredictable.
• Entropy methods: the agent maximizes the diversity of its actions or of the states it visits.
• Competence methods: the agent learns a set of skills that can be reliably told apart from one another.
• Evaluation environment: a virtual test with a ball under the fly's legs (tethered-ball assay), modeled on a real experiment.
• Behavior metrics: statistics of spontaneous locomotion.
• Brain metrics: the agent's internal representations and the structure of its neural activity.
• Decoding: the direction of a future turn was predicted from kinematics and from network activity, on the same timescales as in real flies.

Why DIAYN won

DIAYN (Diversity Is All You Need) is a competence method. It learns a discrete set of policies that maximize behavioral entropy while making the observations produced by different skills easy to distinguish. On combined behavioral and neural predictive encoding, it came closest to the flies.

It also matched the flies in subtler ways. The learned policies reproduce individual variability in exploration strategies: some flies behave one way, others another. The authors interpret this as a mechanism for covering the state space at the population level. Variation between individuals is not noise but the way the species as a whole covers as many options as possible. In addition, DIAYN's latent activity reproduced the "sum-difference" rule for interhemispheric encoding of kinematics, which is observed in the brains of real flies.

What this teaches agent developers

For neuroscience, the result provides a working tool: a digital twin plus a set of intrinsic goals makes it possible to test hypotheses about autonomy at the level of neural representations, not just trajectories. The code and project page are published by NeuroAgents Lab, so the comparison can be reproduced.

For the agent industry, the takeaway is more practical. A hierarchy of discrete skills and the objective of "being diverse and distinguishable" map well onto what robots and autonomous systems need when they have no hard-coded reward: a library of stable primitives with planning on top. The work is not about a commercial model but about which learning principles are worth building into agents that will have to explore an environment on their own. The biological reference here is not a metaphor but a measurable baseline: there is now a quantitative way to say which intrinsic objective is more realistic.