Gaia Molinaro (@gaiamolinaro.bsky.social)

@gaiamolinaro.bsky.social Oct 8, 2026, 8:01 PM

Our reward function compression work is now out in Nature Communications! 🎉 In it, we present and empirically test a computational framework that formalizes the trade-offs between goal flexibility and learning efficiency. @annecollins.bsky.social @natureportfolio.nature.com doi.org/10.1038/s414...

Conceptual model of goal-directed reinforcement learning with reward
function compression. In our proposed conceptual model, we break down goald-dependent learning into several parts. Combined with other learning components, an RL learning algorithm informs value estimates W on a given trial t, which are entered into a policy. The agent selects an action a through a joint policy, to which the environment responds by outputting outcome information o and a new state s. A working memory (WM) component organizes and maintains task-relevant information (here, broadly intended and dubbed the task information), used to support the reinforcement learning (RL) process. a Initially, value attribution to o is operated by WM,which maintains the current goal (g) and supports its matching to the experienced outcome to generate a reward r used by the RL agent. WM also holds a trace of recently experienced goals. b If goals share enough similar features,
a compressed reward function can be created, which replaces active goal maintenance and can, over time, be transferred to long-term memory storage. Eventually, the compressed reward function can be directly employed to assign value to experienced outcomes, sparing resources for other WM tasks.

@gaiamolinaro.bsky.social Oct 3, 2026, 3:14 PM

Well deserved, complimenti Francesco!

@gaiamolinaro.bsky.social Sep 30, 2026, 1:06 PM

Huge congrats!!!! I can't wait to see what fantastic research comes out of your newly minted lab!

@gaiamolinaro.bsky.social Sep 19, 2026, 3:29 PM

These innovations also leave us with several open questions about the characteristics of human goals, AI goals, and synergistic self-driven ("autotelic") systems. Answering them, we believe, will require reciprocal exchange between AI and cognitive science.
www.sciencedirect.com/science/arti...

@gaiamolinaro.bsky.social Sep 19, 2026, 3:29 PM

We explain how, by operating over formats such as natural language and code, foundation models allow agents to: ✨ Generate more abstract and compositional goals. ✨ Approximate human judgments of interestingness and learnability. ✨ Support richer loops of problem generation and problem solving.

Illustration of the evolution of goal-directed learning agents across three paradigms: standard RL, pre-foundation model autotelic loop, foundation model-based autotelic loop.

@gaiamolinaro.bsky.social Sep 19, 2026, 3:27 PM

The ability to pursue self-determined goals is key to human intelligence. Could foundation models unlock similar capabilities in artificial agents? We explore this possibility in a short review — fresh off the press! @ccolas.bsky.social @pyoudeyer.bsky.social @annecollins.bsky.social

@gaiamolinaro.bsky.social Aug 28, 2026, 2:58 PM

This was a fun collaboration led by Jérémy Perez and @ccolas.bsky.social, with @pyoudeyer.bsky.social, @maximederex.bsky.social, and @clemmoulinfrier.bsky.social

@gaiamolinaro.bsky.social Aug 28, 2026, 2:54 PM

Can you think of a goal that’s shaped by your culture? Or one that changed society as a result of being pursued? If not, check out the paper for some hints! 👀
onlinelibrary.wiley.com/doi/10.1111/...

@gaiamolinaro.bsky.social Aug 28, 2026, 2:52 PM

To tackle these ideas, we propose framing individuals as “cultural autotelic agents” and outline ways to study how culture shapes the goals people generate, anticipate, and select.

@gaiamolinaro.bsky.social Aug 28, 2026, 2:52 PM

Curiosity and intrinsic motivation help keep this cycle open-ended, pushing people toward diverse and increasingly complex pursuits even when there is no obvious external reward.

@gaiamolinaro.bsky.social Aug 28, 2026, 2:51 PM

Why do some goals spread across people and generations, while others emerge only transiently? We argue that human goals are both drivers and products of cultural evolution: our individual pursuits are shaped by our cultural environments, and the goals we pursue in turn reshape those environments.