This analytic-then-generative approach reveals that effective teaching is not an art beyond measurement, but a discoverable compositional semantic structure with predictable effects!
We demonstrate causality by synthesizing lessons that weight these dimensions optimally: synthetic lessons matched the behavioral benefits of the best human teachers, while control lessons weighted randomly produced no advantage.
By developing a novel computational pipeline (LLM-DISC) that decomposes natural teaching language into interpretable semantic dimensions, that independently predict both expert ratings and learning outcomes.
Teaching through language is uniquely human, yet we cannot predict which explanations will help learners succeed. This gap has profound implications: educational interventions remain trial-and-error, tutoring systems rely on heuristics rather than principles.
I haven't given any news a while, I've been nose deep into this novel preprint with my excellent collaborators @stepalminteri.bsky.social, @urihertz.bsky.social and Bahador Bahrami: "Uncovering the semantics of teaching in
experiential learning with Large Language Models".
doi.org/10.31234/osf...
Very happy to be a part of this! More coming soon!
I have the immense pleasure of of announcing that @psl-univ.bsky.social and @cognitionens.bsky.social have granted me an HDR diploma (i.e. the highest qualification in French academia). This wouldn't have been possible without @stepalminteri.bsky.social and the INCREDIBLE team at HRL/LNC2. Thank you
Oh heck yes! Wonder if I'll finally be able to drop Matlab for the remaining capricious RL models I have and go full R once and for all...
Amazing, dare I say mandatory, new resource
Like if they lead nation-wide committees that could end your career forever if they blacklisted you, so they get away with discrimination? Asking for a friend...
Most intriguing!!
I'm sure you saw that all data is available for our 11-country study www.nature.com/articles/s41...
Hmmmm. I get what's going on here, but indeed the only way to conclusively debunk this IMO is to replicate in rural areas and non-urban ethnic groups. As I see it now (pun intended) the effect is clearly low level in origin, but culture could defo determine its instantiation.
Hmmmm. I get what's going on here, but indeed the only way to conclusively debunk this IMO is to replicate in rural areas and non-urban ethnic groups. As I see it now (pun intended) the effect is clearly low level in origin, but culture could defo determine its instantiation.
End of an era? I also cited this study :o
We're getting there people!!
Fascinating work!
Epistemic biases in human reinforcement learning: behavioral evidence, computational characterization, normative status and possible applications.
A quite self-centered review, but with a broad introduction and conclusions and very cool figures.
Few main takes will follow
osf.io/preprints/ps...
Worth also pointing out that there are many "tests so easy no AI system can pass them".
Moravec's paradox remains.
E.g., arxiv.org/abs/2404.12390
Do include me please!
Lovely, thanks for sharing. Open for a chat indeed!
What a cautionary tale waiting to happen this is!!
This!!!!!! A plague in EEG and fMRI
It will do, it will do *faints into oblivion*
One of my personal stats heroes E J Wagenmakers has just liked one of my goofy comments. I can't begin to explain my excitement. Now I'm just missing @bbolker.bsky.social, and I can retire in peace.
Yes. I mean hell, half of the psychologists i know don't fully understand how linear regression works so...
I want in!
Feeling seen, graphical description
Here's the official link to our latest paper!
New paper!
We previously showed that people are more likely to follow advice than observed behavior.
@anllohernan.bsky.social asks whether this behavior makes sense - are people more likely to give advice when they are knowledgeable?
The answer is yes, but it depends on advice consequences👇
