Really excited to have the opportunity to give a talk on this work @cogscisociety.bsky.social !!! last year was a blast can’t wait to go back to Rio in July 🇧🇷
HUGE thanks to my collaborators for the support @aydanhuang265.bsky.social @EricYe29011995 @natashajaques.bsky.social @maxkw.bsky.social 🙏
Our new short piece in TiCS on intuitive theories of truth: how people judge whether statements could be true, whether statements are true, and whether to assert them as true. A great collab with @keremoktar.bsky.social
@ihandleyminer.bsky.social @kevinzollman.com @lianeleeyoung.bsky.social
Amazing!
Can't wait to present this work @iclr-conf.bsky.social this year!!! Looking forward to hearing everyone's thoughts on the paper and learning more about peoples' research!
Thanks again to my collaborators for all of their help on this project!
🤔💭What even is reasoning? It's time to answer the hard questions!
We built the first unified taxonomy of 28 cognitive elements underlying reasoning
Spoiler—LLMs commonly employ sequential reasoning, rarely self-awareness, and often fail to use correct reasoning structures🧠
Forget modeling every belief and goal! What if we represented people as following simple scripts instead (i.e "cross the crosswalk")?
Our new paper shows AI which models others’ minds as Python code 💻 can quickly and accurately predict human behavior!
shorturl.at/siUYI%F0%9F%...
New paper challenges how we think about Theory of Mind. What if we model others as executing simple behavioral scripts rather than reasoning about complex mental states? Our algorithm, ROTE (Representing Others' Trajectories as Executables), treats behavior prediction as program synthesis.
Definitely, we should look closer at sample complexity for training but for things like webnav there are massive datasets so could be good fit.
In some sense, yes, in that you need diverse trajectories of the agent's behavior in different contexts, but you don't need to have access to those goals, or even the distribution, and the agent might be doing non-goal-directed behavior, such as exploration.
Great work led by @andyliu.bsky.social and collaborators:
@kghate.bsky.social, @monadiab77.bsky.social, @daniel-fried.bsky.social, @atoosakz.bsky.social
Preprint: www.arxiv.org/abs/2509.25369
When values collide, what do LLMs choose? In our new paper, "Generative Value Conflicts Reveal LLM Priorities," we generate scenarios where values are traded off against each other. We find models prioritize "protective" values in multiple-choice, but shift toward "personal" values when interacting.
Very cool! Thanks for sharing! Would be interesting to compare your exploration ideas on open ended tasks beyond little alchemy with EELMA
Work led by Jinyeop Song together with Jeff Gore. Check out the preprint here: arxiv.org/abs/2509.22504
Excited by our new work estimating the empowerment of LLM-based agents in text and code. Empowerment is the causal influence an agent has over its environment and measures an agent's capabilities without requiring knowledge of its goals or intentions.
Claire's new work showing that when an assistant aims to optimize another's empowerment, it can lead to others being disempowered (both as a side effect and as an intentional outcome)!
Person standing next to poster titled "When Empowerment Disempowers"
Still catching up on my notes after my first #cogsci2025, but I'm so grateful for all the conversations and new friends and connections! I presented my poster "When Empowerment Disempowers" -- if we didn't get the chance to chat or you would like to chat more, please reach out!
It’s forgivable =) We just do the best we can with what we have (i.e., resource rational) 🤣
Max giving a talk w the slide in OP
lol this may be the most cogsci cogsci slide I've ever seen, from @maxkw.bsky.social
"before I got married I had six theories about raising children, now I have six kids and no theories"......but here's another theory #cogsci2025
Quantifying the cooperative advantage shows why humans, the most sophisticated cooperators, also have the most sophisticated machinery for understanding the minds of others. It also offers principles for building more cooperative AI systems. Check out the full paper!
www.pnas.org/doi/10.1073/...
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.
Even in one-shot games with observability, the Bayesian Reciprocator learns from observing others' interactions and enables cooperation through indirect reciprocity
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.
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.
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.
We introduce the "Bayesian Reciprocator," an agent that cooperates with others proportional to its belief that others share its utility function.
Classic models of cooperation like tit-for-tat are simple but brittle. They only work in specific games, can't handle noise and stochasticity and don't understand others' intentions. But human cooperation is remarkably flexible and robust. How and why?
This project was first presented back in 2018 (!) and was born from a collaboration between Alejandro Vientos, Dave Rand @dgrand.bsky.social & Josh Tenenbaum @joshtenenbaum.bsky.social
Our new paper is out in PNAS: "Evolving general cooperation with a Bayesian theory of mind"!
Humans are the ultimate cooperators. We coordinate on a scale and scope no other species (nor AI) can match. What makes this possible? 🧵
www.pnas.org/doi/10.1073/...
As always, CogSci has a fantastic lineup of workshops this year. An embarrassment of riches!
Still deciding which to pick? If you are interested in building computational models of social cognition, I hope you consider joining @maxkw.bsky.social, @dae.bsky.social, and me for a crash course on memo!