iiinova — where your digital self keeps evolving

Finally, when we tested it against memory-1 strategies (such as TFT and WSLS) in the iterated prisoner's dilemma, the Bayesian Reciprocator: expanded the range where cooperation is possible and dominated prior algorithms using the *same* model across simultaneous & sequential games.

160

Even in one-shot games with observability, the Bayesian Reciprocator learns from observing others' interactions and enables cooperation through indirect reciprocity

160

In dyadic repeated interactions in the Game Generator, the Bayesian Reciprocator quickly learns to distinguish cooperators from cheaters, remains robust to errors, and achieves high population payoffs through sustained cooperation.

260

Instead of just testing on repeated prisoners' dilemma, we created a "Game Generator" which creates infinite cooperation challenges where no two interactions are alike. Many classic games, like the prisoner’s dilemma or resource allocation games, are just special cases.

190

It uses theory of mind to infer the latent utility functions of others through Bayesian inference and an abstract utility calculus to work across ANY game.

150

Max Kleiman-Weiner @maxkw.bsky.social · 29/06/2025

We introduce the "Bayesian Reciprocator," an agent that cooperates with others proportional to its belief that others share its utility function.

170

Settling in for my flight and apparently A.I. DOOM is now a movie genre between Harry Potter and Classics. Nothing better than an existential crisis with pretzels and a ginger ale.

060

Max Kleiman-Weiner @maxkw.bsky.social · 09/06/2025

LLMs learn beliefs and values from human data, influence our opinions, and then reabsorb those influenced beliefs, feeding them back to users again and again. We call this the "Lock-In Hypothesis" and develop theory, simulations, and empirics to test it in our latest ICML paper!

1306

Max Kleiman-Weiner @maxkw.bsky.social · 26/01/2025

Emergent transition from code to natural language for reasoning tasks when RL tuning a language model for math. Interesting to consider implications for "Language of Thought" style theories in cognition. hkust-nlp.notion.site/simplerl-rea...

1201

Max Kleiman-Weiner @maxkw.bsky.social · 15/12/2024

Josh Tenenbaum on scaling up vs growing up and the path to human-like reasoning #NeurIPS2024

1826

Max Kleiman-Weiner @maxkw.bsky.social · 14/12/2024

So much fun working with wonderful co-first authors: Zhijing Jin and Giorgio Piatti and collaborators Sydney Levine, Jiarui Liu, Fernando Gonzalez, Francesco Ortu, András Strausz, @mrinmaya.bsky.social, Rada Mihalcea, Yejin Choi, Bernhard Schölkopf

150

Max Kleiman-Weiner @maxkw.bsky.social · 05/12/2024

We benchmark 13 LLMs in GovSim. Overall, we find that this is quite challenging for LLMs! Even the best performing model (GPT-4o) survives the first 12 rounds less than 60% of the time. Most never survive.

100

Max Kleiman-Weiner @maxkw.bsky.social · 05/12/2024

Over the past 40 years, this question has received extensive empirical attention, culminating in the 2009 Nobel Prize to Elinor Ostrom for characterizing the many mechanisms people and communities use to cooperate sustainably from the bottom up.

110

Max Kleiman-Weiner @maxkw.bsky.social · 05/12/2024

Excited that our multi-agent LLM agent work, “Cooperate or Collapse: Emergence of Sustainable Cooperation in a Society of LLM Agents,” will be presented at #NeurIPS24 -- reach out if you want to meet up in Vancouver!

1287

Max Kleiman-Weiner @maxkw.bsky.social · 26/11/2024

Github Copilot output. A sad but fascinating alignment failure. Reveals hidden LLM biases by going out of their RLHF distribution

3152

Max Kleiman-Weiner @maxkw.bsky.social · 15/11/2024

I’m recruiting PhD students to join the Computational Minds and Machines Lab at the University of Washington in Seattle! Join us to work at the intersection of computational cognitive science and AI with a broad focus on social intelligence. (Please reshare!)

13313

Max Kleiman-Weiner @maxkw.bsky.social · 11/11/2024

These videos of a humanoid watching itself in a mirror are a cool test of a “self” model. The robot’s world model tries to predict the next visual inputs. It can predict its own movements from a first-person view but doesn’t map the mirrored body to its own, and thus, the

100

Max Kleiman-Weiner @maxkw.bsky.social · 11/11/2024

Would love to see these examples scaled up into a larger benchmark about models of self.

110

Max Kleiman-Weiner @maxkw.bsky.social · 11/11/2024

These videos of a humanoid watching itself in a mirror are a cool test of a “self” model. The robot’s world model tries to predict the next visual inputs. It can predict its movements from a first-person view but doesn’t map the mirrored body to its own, and thus, the mirrored view doesn't match.

150

Hope this is a step towards understanding computation in the brain.

020

Incredible to see that the fruit fly connectome has been completed. I remember as an undergraduate watching Sebastian Seung in 2007 give the presidential lecture at SFN mysteriously titled: “The Once and Future Science of Neural Networks.”

2293

The full scientific work is published across 8 nature papers in a single issue: www.nature.com/collections/hgcfafej…

000

Incredible to see that the fruit fly connectome has been completed. I remember as an undergraduate watching @SebastianSeung’s in 2007 give the presidential lecture at SFN mysteriously titled: “The Once and Future Science of Neural Networks.” He spent the first half talking about

100

Finally, we explore the effects of incomplete internalization and show "reward hacking," where the agent's behavior diverges from the intended goals.

100

Through simulations, we demonstrate that an ISR model prevents agents from unlearning socialized behaviors and enables them to generalize to new tasks. The ISR model continues to guide the agent even when the caregiver is no longer present.

100

During development, caregivers guide learners toward culturally aligned goals. But what happens when the learner must continue autonomously? Our model of value internalization shows how social feedback can train an internal social reward (ISR) system.

100