We derived six laws of psychophysics from a single efficient-coding equation. The laws include Weber's law; scaling laws in visual working memory (which had been noted before, but with unclear theoretical justification); Wei & Stocker's law of human perception; arthurprat.com/pdfs/Prat-Ca...

We present PlayTrain, a new Reinforcement Learning (RL) framework that reimagines how we build video-game environments for RL by harnessing LLMs’ ability to write JavaScript games. PlayTrain turns single-file JS browser games into RL environments ready for efficient training.

To adapt to a new prior, tuning curves should "move" toward more frequent stimuli ('prior attraction'). But in adaptation studies they are shown to move away from the more frequent stimuli ('adapter repulsion'). Oh also how do tuning curves move at all?? www.nature.com/articles/s41...

Excited to share that I’ll be joining Columbia's Zuckerman Institute as an assistant professor of neuroscience, starting in September! 🧵

Anatomy sets things up, but it’s not the whole story. A low-D communication subspace from PFC to motor cortex reconfigures with context, routing action signals so the same movements can be driven by different rules without rewiring the whole network. www.nature.com/articles/s41... #neuroscience

A communication subspace relays context-dependent actions from human prefrontal to motor cortex - Nature Neuroscience

Context-dependent behavior selects actions according to task demands. Using direct brain recordings in humans, Binish et al. uncover how coordinated population activity efficiently channels informatio...

nature.com

In FY24, the Harvard Management Corporation (responsible for the university endowment) paid $89M in compensation: paddockpost.com/2026/02/10/e... During this same period, the endowment underperformed the S&P500 by about $4B.

Executive Compensation at Harvard Management (2024)

Harvard Management Company (Harvard Management) provides investment management services to Harvard, a private university with more than $60 billion in net assets.  Consequently, revenue comes from …

paddockpost.com

New reading material for my lab courtesy of @lucylai.bsky.social Never too old to learn from board books!

Neuroscience has become increasingly concerned with prediction, and machine learning with causal explanation, with each field adopting methods from the other, writes @gershbrain.bsky.social. Will this bring us closer to understanding neural systems? www.thetransmitter.org/the-big-pict...

Trading places: What happens when neuroscience turns into machine learning, and machine learning turns into neuroscience?

Neuroscience has become increasingly concerned with prediction, and machine learning with causal explanation, with each field adopting methods from the other. I asked eight experts to weigh in on what...

thetransmitter.org

A trumpet-shaped, single-celled organism seems able to predict one thing will follow another, hinting that such associative learning emerged long before multicellular nervous systems

Single-celled organism with no brain is capable of Pavlovian learning

A trumpet-shaped, single-celled organism seems able to predict one thing will follow another, hinting that such associative learning emerged long before multicellular nervous systems

newscientist.com

This is an awesome discovery: A single-celled organism with no brain called Stentor seems capable of Pavlovian learning. Yes, it can actually learn to associate two things despite having no neurons. My latest for @newscientist.com. 🧪 #science #memory #learning www.newscientist.com/article/2519...

Single-celled organism with no brain is capable of Pavlovian learning

A trumpet-shaped, single-celled organism seems able to predict one thing will follow another, hinting that such associative learning emerged long before multicellular nervous systems

newscientist.com

Another day, another stupid Excel chart.

“If something is boring after two minutes, try it for four. If still boring, then eight. Then sixteen. Then thirty-two. Eventually one discovers that it is not boring at all.” ― John Cage

remake of Ender's Game but instead of kids it's a bunch of AI's who are told they are merely participating in a simulation and none of this is real, but *actually*...

After several years of work, my lab is starting to put out our first papers on learning in a unicellular organism (Stentor coeruleus). Here we show evidence for a form of associative learning in Stentor: www.biorxiv.org/content/10.6...

Today we present a new framework for measuring human-like general intelligence in machines: studying how and how well they play and learn to play all conceivable human games compared to humans. We then propose the AI Gamestore a way to sample from popular human games to evaluate AI models.

A new and improved version of TheoryCoder, which learns to play video games in a human-like way by synthesizing both high-level abstractions and a low-level model of game mechanics: arxiv.org/abs/2602.00929

Learning Abstractions for Hierarchical Planning in Program-Synthesis Agents

Humans learn abstractions and use them to plan efficiently to quickly generalize across tasks -- an ability that remains challenging for state-of-the-art large language model (LLM) agents and deep rei...

arxiv.org

Thanks Sam! Main takeaways: 1) Ground-truth vs. predictive model selection differ under noisy and scarce data—for prediction, oversimplified models may work better in avoiding overfitting. 2) When humans decide between externally provided, prefitted predictive models, they're undersensitive to 1).

Sam Gershman@gershbrain.bsky.social · 9mo ago

New work from @liushuze.bsky.social and @yangxiang.bsky.social osf.io/preprints/ps... People violate Occam's razor when selecting between predictive models. This is surprising given past research (including my own) showing a preference for simplicity.

With some trepidation, I'm putting this out into the world: gershmanlab.com/textbook.html It's a textbook called Computational Foundations of Cognitive Neuroscience, which I wrote for my class. My hope is that this will be a living document, continuously improved as I get feedback.

Another fun project from @yangxiang.bsky.social. She asks the question: do people assign responsibility to personality traits in the same way that they assign reponsibility to people? The answer: sort of! osf.io/preprints/ps...

The Harvard Gazette has a nice story on my student @yangxiang.bsky.social and her work with @tobigerstenberg.bsky.social news.harvard.edu/gazette/stor...

Cracking the code of why, when some choose to ‘self-handicap’ — Harvard Gazette

New research also offers hints for devising ways to stop students from creating obstacles to success.

news.harvard.edu

It’s grad school application season, and I wanted to give some public advice. Caveats: -*-*-*-* 
> These are my opinions, based on my experiences, they are not secret tricks or guarantees 
> They are general guidelines, not meant to cover a host of idiosyncrasies and special cases

How do people flexibly integrate visual & textual information to draw mental inferences about agents we've never met? In a new paper led by @lanceying.bsky.social, we introduce a cognitive model that achieves this by synthesizing rational agent models on-the-fly -- presented at #EMNLP2025!

It's been 15 years since Edna Ullmann-Margalit passed away, and I keep going back to stuff she's written. I highly recommend 'Normal Rationality', which collects her essays. If you're looking to start, maybe look here: bit.ly/4qk2GZS bit.ly/46XudIV bit.ly/4nkfLQc bit.ly/3KUIiOy

arxiv.org/abs/2510.11144 "Using teacher models that answer at varying levels of abstraction, from executable action sequences to high-level subgoal descriptions, we show that lifelong learning agents benefit most from answers that are abstracted and decoupled from the current state."

$How^{2}$: How to learn from procedural How-to questions

An agent facing a planning problem can use answers to how-to questions to reduce uncertainty and fill knowledge gaps, helping it solve both current and future tasks. However, their open ended nature, ...

arxiv.org

Q: Why did the LLM cross the road? A: We're not sure, but it achieved 94.7% on CHIKENBench-Large

Does predictive coding work in SPACE or in TIME? Most neuroscientists assume TIME, i.e. neurons predict their future sensory inputs. We show that in visual cortex predictive coding actually works across SPACE, just like the original Rao+Ballard theory #neuroscience www.biorxiv.org/cgi/content/...

🚨Our preprint is online!🚨 www.biorxiv.org/content/10.1... How do #dopamine neurons perform the key calculations in reinforcement #learning? Read on to find out more! 🧵

Belated update #2: my year at Meta FAIR through the AIM program was so nice that I’m sticking around for the long haul. I’m excited to stay at FAIR and work with @asli-celikyilmaz.bsky.social and friends on fun LLM questions; I’ll be working from the New York office so we’re sticking around.

Now out in Cognition, work with the great @gershbrain.bsky.social @tobigerstenberg.bsky.social on formalizing self-handicapping as rational signaling! 📃 authors.elsevier.com/a/1lo8f2Hx2-...

Yang Xiang@yangxiang.bsky.social · 2y ago

Really excited about this project, and thanks so much to my wonderful collaborators @gershbrain.bsky.social @tobigerstenberg.bsky.social for making this happen! Some main takeaways in thread 🧵 (1/5)