Jonathan Phillips

Happy to announce that my lab @ Yale Psychology (actcompthink.org) will be accepting PhD applications this year (for start in Fall '26)!

Come for the fun experiments on human learning, memory, & skilled behavior, stay for the best 🍕 in the US.

Please reach out if you have any questions!

Meta-reasoning @ CogSci Workshop Description People are general purpose problem solvers. We obtain food and shelter, manage companies, solve moral dilemmas, spend years toiling away at thorny math problems, and even adopt a...

If you’ll be at #CogSci2025, consider (or at least consider considering) attending our @cogscisociety.bsky.social workshop on meta reasoning
🤔🤨🧐
We’ll be discussing problem selection through various lenses represented by a great lineup of speakers!

Model-based algorithms shape automatic evaluative processing | PNAS Computational theories of reinforcement learning suggest that two families of algorithm—model-based and model-free—tightly map onto the classic dis...

A key takeaway from 20+ years of computational RL is: model-free=automatic, model-based=deliberate. My new paper w/ @benedek.bsky.social challenges this view, suggesting that MB algos are more ubiquitous, & automatic processing more sophisticated, than currently thought: www.pnas.org/doi/10.1073/...

In the second most important election happening today, I'm on the slate for potential new members of the governing board for the Cognitive Science Society! If you're a member, check your email for a link to vote and #DontRankCuomo

JOB!

3yr funded post-doc in Theory of Mind inspired by the knowledge first epistemology of Williamson, and the work of @jsphillips.bsky.social. Looking at knowledge and ignorance processing in adults with me and Richard O'Connor at the Uni of Hull. Please re-post.

www.jobs.ac.uk/job/DNE794/p...

bit of good news: approved technical staff position! link below. please be in touch if this matches your skills & interests! drive.google.com/file/d/16J2J... (hr listing posted harvard-internal now; external soon, per guidelines), happy for ?s & plan on quick turnaround! #CogSciSky #PsychSciSky 🐦🐦

Title: Representations of what’s possible reflect others’ epistemic states Authors: Lara Kirfel, Matthew Mandelkern, and Jonathan Scott Phillips Abstract: People’s judgments about what an agent can do are shaped by various constraints, including probability, morality, and normality. However, little is known about how these representations of possible actions—what we call modal space representations—are influenced by an agent’s knowledge of their environment. Across two studies, we investigated whether epistemic constraints systematically shift modal space representations and whether these shifts affect high-level force judgments. Study 1 replicated prior findings that the first actions that come to mind are perceived as the most probable, moral, and normal, and demonstrated that these constraints apply regardless of an agent’s epistemic state. Study 2 showed that limiting an agent’s knowledge changes which actions people perceive to be available for the agent, which in turn affects whether people judged an agent as being “forced” to take a particular action. These findings highlight the role of Theory of Mind in modal cognition, revealing how epistemic constraints shape perceptions of possibilities.

🏔️ Brad is lost in the wilderness—but doesn’t know there’s a town nearby. Was he forced to stay put?

In our #CogSci2025 paper, we show that judgments of what’s possible—and whether someone had to act—depend on what agents know.

📰 osf.io/preprints/ps...

w/ Matt Mandelkern & @jsphillips.bsky.social

Couldn't be more thrilled that Fred is coming to join us!! Dartmouth Cognitive Science is quickly growing into a group of amazing colleagues that I feel lucky to have around and think with!

Now out in JPSP ‼️

"Inference from social evaluation" with Zach Davis, Kelsey Allen, @maxkw.bsky.social, and @julianje.bsky.social

📃 (paper): psycnet.apa.org/record/2026-...
📜 (preprint): osf.io/preprints/ps...

Bonobo Nyota at Ape Initiative, a science and education nonprofit

Are humans the only species that communicates when a collaborator is missing information?

In @pnas.org, Luke Townrow and I show that our closest relatives, bonobos, can track when a partner is knowledgeable or ignorant, and tailor communication accordingly

www.pnas.org/doi/10.1073/...

Or this, hopefully now working, OSF link: osf.io/preprints/ps...

If the OSF link isn't working for you, the preprint can also be found here: dataverse.harvard.edu/file.xhtml?f...

We find that the visual system's representation of multiple possibilities is selectively disrupted by perceptual load, but not cognitive load, demonstrating that the key processes underlying the perception of possibilities occur before the information reaches high-level cognition!

The key idea (developed with Camden Parker and @violastoermer.bsky.social) was to use amodal completion as a case where the visual system can represent multiple possibilities (possible shapes) and then ask whether this representation is differentially disrupted by perceptual load or cognitive load.

In a new paper, we demonstrate the perception of possibilities but show that the processes underlying this phenomenon occur before the information reaches high-level cognition. The representation of these possibilities is distinctly perceptual(!) and separate from cognition. osf.io/preprints/ps...

Are you interested in research experience before applying to PhD programs? Or just want to learn more about cognitive science? Consider joining my lab as a lab manager (joint w/the Griffiths Lab). We will begin reviewing applications one week from today: cognition.princeton.edu/news/2025/op...

Our new paper with Max Taylor-Davies introduces a resource-rational model of Theory of Mind.

The model can explain many of the successes and failures of mindreading in human adults and children, and non-human primates. 🧵

For anybody interested in this sort of thing, I think this is a valuable resource.

A graph of articles in the SEP (standord encyclopaedia of philosophy), showing connections; it can help explore the field.

www.visualizingsep.com#/domain/epis...

#Philosophy #philsky #SEP #graph #catalogue

This is joint work with Bryan Gonzalez, Pauline Amary, James Dungan, Brent Strickland, @xphilosopher.bsky.social, and @fierycushman.bsky.social. A huge amount of credit goes out to them!

Totally agree with this, but I'm not hopeful bc it's hard to know what the bounds of that broader space are for the kind of generalizability we care about. @asbear.bsky.social and I tried to make this point here (in response to @talyarkoni.com 's article): drive.google.com/file/d/1LKo5...

In sum, these studies collectively demonstrate that we can attribute or deny knowledge states without evaluating belief states and suggest that knowledge representation is distinct from belief representation and offers a conceptually primitive way to represent others’ minds.

Percent signal change in RTPJ for each mental state verb: 'know' (left points and violin), other factive verbs, e.g., 'saw' or 'realized' (middle points and violin) and 'think' (right points and violin). Small points indicate trial-level responses, violins illustrate scenario- or item-level distributions, large dark points depict overall means, and error bars depict +/- 1 SEM.

And finally, we used a quite different methodology to show that people's neural patterns reveal a similar relationship: they have a lower BOLD response in the theory of mind network when evaluating knowledge than when evaluating belief (showing they didn't compute belief in computing knowledge):

Response times for correct evaluations of factive mental state ascriptions (left points and violins) and non-factive mental state ascriptions (right points and violins) as a function of Information Condition (separate panels). Small points indicate trial-level responses, violins illustrate scenario- or item-level distributions, large dark points depict overall means, and error bars depict +/- 1 SEM.

Next, we showed that this is actually part of a more general pattern, where people are generally faster to make accurate evaluations of factive mental states (e.g., aware, recognize, understand) than non-factive mental states (e.g., believe, guess, assume):

Response times for correct evaluations of knowledge ('Savoir') ascriptions (left points and violins) and belief ('Penser') ascriptions (right points and violins) as a function of Information Condition (separate panels). Error bars depict Small points indicate trial-level responses, violins illustrate scenario- or item-level distributions, large dark points depict overall means, and error bars depict +/- 1 SEM.

We then find that this pattern is not specific to English, and that it generalizes to French speakers as well. French is an especially hard test case because in a lexical decision task, recognition of 'savoir' (know) is actually *slower* than 'penser' (think), and yet:

Response times of neruotypical (A) and autistic (B) groups for correct evaluations of knowledge ascriptions (left points and violins) and belief ascriptions (right points and violins) as a function of Information Condition (separate panels). Small points indicate trial-level responses, violins illustrate scenario- or item-level distributions, large dark points depict overall means, and error bars depict +/- 1 SEM. **C**: Difference in participant mean response time between belief and knowledge attribution trials as a function of participants’ score on the Autism Quotient-10 scale across all participants.

We then replicated this finding and showed that it extended to participants with Autism. For both, know < think, and this relationship is unrelated to AQ 10 scores. The pattern that knowledge evaluations are simpler and independent from belief is preserved across differences in neurotypicality!

Response times for correct evaluations of knowledge ascriptions (left points and violins) and belief ascriptions (right points and violins) as a function of Information Condition (separate panels). Small points indicate trial-level responses, violins illustrate scenario- or item-level distributions, large dark points depict overall means, and error bars depict +/- 1 SEM.

We first find simply that people are faster to accurately evaluate whether or not someone knows something than whether or not they think that same thing, indicating that they seem to be evaluating others' knowledge without first evaluating what they believe:

In a new preprint (doi.org/10.31234/osf...) a huge range of data+methods shows that people can evaluate what others know without first evaluating what they think/believe. Representations of knowledge seem to be an independent and conceptually primitive way of representing others' minds. 🧵 below!

🚨 Now accepting commentary proposals!! 🚨Thrilled to share that our paper --- "Resource-rational contractualism: A triple theory of moral cognition" --- was accepted for publication at Behavioral and Brain Sciences and is open for commentary!

🔊 New paper just accepted in JPSP 🥳

In "Inference from social evaluation", we explore how people use social evaluations, such as judgments of blame or praise, to figure out what happened.

📜 osf.io/preprints/ps...

📎 github.com/cicl-stanfor...

1/6

Deadline extended! Submit your abstracts by Friday, January 31 at 11:59 PM Eastern.