iiinova — where your digital self keeps evolving

Reposted by Stefano Palminteri

Stefano Palminteri @stepalminteri.bsky.social · 05/10/2026

New preprint! Can LLMs really know when they know? We trained 10 open models to predict their own accuracy and found a striking limit: out of distribution, confidence tracks consistency more than correctness. Led by @nicolasyax.bsky.social with @pyoudeyer.bsky.social arxiv.org/pdf/2609.33886

0298

Stefano Palminteri @stepalminteri.bsky.social · 05/10/2026

0298

Reposted by Stefano Palminteri

Stefano Palminteri @stepalminteri.bsky.social · 01/10/2026

1/ Very proud of this preprint — and especially because it marks a first step for our lab in a new direction: using what we know about learning not only to diagnose misconceptions, but also to correct them. zenodo.org/records/2303...

12110

Stefano Palminteri @stepalminteri.bsky.social · 01/10/2026

4/ Effects on belief and preferences were still detectable weeks later. More broadly, we hope this work makes a case for adding reinforcement learning to the applied behavioural science toolbox.

120

Stefano Palminteri @stepalminteri.bsky.social · 01/10/2026

4/ And the effects went beyond knowledge: belief correction predicted shifts in investment preferences toward cleaner and safer energy mixes. Reinforcement → belief correction → preference change.

110

Stefano Palminteri @stepalminteri.bsky.social · 01/10/2026

3/ We then tested a very simple intervention inspired by reinforcement learning: make a choice → receive immediate feedback → learn. This brief training substantially corrected beliefs, especially where initial misconceptions were strongest.

220

Stefano Palminteri @stepalminteri.bsky.social · 01/10/2026

2/ Across 7 countries, we found robust misconceptions about energy sources — notably overestimating the CO₂ emissions and mortality associated with nuclear power, and underestimating the health costs of biomass.

120

Stefano Palminteri @stepalminteri.bsky.social · 01/10/2026

12110

Reposted by Stefano Palminteri

Caroline Pioger @carolinepioger.bsky.social · 29/09/2026

🚨 New preprint! Public debate on decarbonization often rests on inaccurate beliefs about energy sources. We tested whether a reinforcement-learning-based intervention can correct them. 🧵👇 📄 Preprint: zenodo.org/records/2303...

zenodo.org

Correcting misconceptions and shaping preferences about energy with reinforcement learning

Addressing the climate crisis requires citizens to make informed choices about sustainable energy policy, yet public debates are often shaped by inaccurate beliefs about the environmental and health i...

1114

Reposted by Stefano Palminteri

Stefano Palminteri @stepalminteri.bsky.social · 13/07/2026

You do not know what to read? 📖🤔 Below is a quick digest of this year's papers from the HRL team. The range is a little unusual: from reinforcement learning in rats, to the philosophy of AI, by way of eye-tracking and social-media behaviour. Seven papers, each with a short teaser 👇

2184

Stefano Palminteri @stepalminteri.bsky.social · 13/07/2026

Are debates about LLMs driven by evidence—or by our own biases? With @giadapistilli.com, we examine how wishful thinking, cognitive dissonance and explanatory overconfidence fuel both hype and dismissal. Part of a AI & Psychology special issue: more soon! doi.org/10.1016/j.cobeha.2026.101670

064

Stefano Palminteri @stepalminteri.bsky.social · 13/07/2026

How could we ever tell whether a machine is conscious? A case for looking at behaviour rather than inside. [Palminteri & @thecharleywu.bsky.social] academic.oup.com/nc/article/2...

How would we ever know whether a machine is conscious? The prevailing answer is to look inside and check whether it runs the computations thought to underlie consciousness in us, but we argue this demands something we can supply neither for the machine nor for ourselves. Drawing on the everyday epistemology of cognitive science, we propose a different and more operational basis for attributing consciousness: methodolgical behaviorism.

020

Stefano Palminteri @stepalminteri.bsky.social · 13/07/2026

What keeps us posting on social media? [@georgiaturner.bsky.social, Gunschera, Subrahmanya, Salecha, Eichstaedt, Me & @orbenamy.bsky.social] www.nature.com/articles/s41...

Why do we keep posting on social media, when the cognitive processes behind ordinary posting have been so hard to measure? We adapted a model from the animal reward-learning literature and fit it to real posting histories from thousands of Twitter/X users, including a preregistered replication. Posting is best explained by a blend of goal-directed reward learning and habit, and the balance between them shifts with how often people post and with who they are.

164

Stefano Palminteri @stepalminteri.bsky.social · 13/07/2026

Why does outcome feedback shift our risk preferences when, in theory, it shouldn't? [Antonis Nasioulas, Elise Potier, Fabien Cerrotti, @mael-lebreton.bsky.social & Me] www.nature.com/articles/s41...

When you choose between two fully described gambles, normative theory says that learning how things turned out should not change your preference next time, yet it does. Is feedback teaching us something that reshapes each option's value, or shifting our appetite for risk before any outcome is even revealed? Across seven experiments we separate the two by pinning down when the change happens, and the results points away from learning, toward curiosity and regret.

100

Stefano Palminteri @stepalminteri.bsky.social · 13/07/2026

How does where we look shape what we learn options are worth? [@romanececchi.bsky.social, @sgluth.bsky.social & Me] www.nature.com/articles/s41...

We tend to learn the value of options relative to their context, yet this range-normalisation account breaks down in a specific, reproducible way when three equally spaced options are on offer. We suspected attention, and combined eye-tracking with manipulations that steer where people look. Gaze changes what an option is worth, and a model in which attention shapes value before it is normalised captures what the standard account misses.

100

Stefano Palminteri @stepalminteri.bsky.social · 13/07/2026

Does context distort our choices, or the memory of value itself? [Magdalena Soukupova, @bsgarcia.bsky.social & Me] www.nature.com/articles/s41...

How options are paired during learning distorts the values we later reveal in our choices, but does the distortion live in the choices or in the representation of value itself? We asked people not only to choose but to report from memory what each option had been worth, across seven experiments. Remembered values are distorted in the same context-dependent way, and the same principle can be turned around to design environments that help people choose better.

100

Stefano Palminteri @stepalminteri.bsky.social · 13/07/2026

Is the positivity bias in learning real, or just an artefact of learning rates that decay over time? [@romanececchi.bsky.social & Me] www.pnas.org/doi/10.1073/...

When people learn from reward, they weigh good news more heavily than bad, and a recent proposal argued this positivity bias might be an artefact of learning rates that naturally decline over time.

We refit six datasets with models that allow exactly this decay. The bias does not go away, and the decay rates themselves reveal an asymmetry no one had reported before.

100

Stefano Palminteri @stepalminteri.bsky.social · 13/07/2026

Is reference-point dependence a quirk of human psychology, or shared with rats? [Lachlan Ferguson, Magdalena Soukupova, Sebastien Bouret, Me & Shauna Parkes] www.nature.com/articles/s41...

The value we place on an outcome depends on what we compare it against, a reference-point dependence central to prospect theory, but is it peculiar to humans or more deeply conserved? We built matched reinforcement-learning tasks for humans and rats to ask whether both encode value relative to a reference point. Behaviour and computational modelling converge across the two species, with a few instructive differences along the way.

121

Stefano Palminteri @stepalminteri.bsky.social · 13/07/2026

2184

Reposted by Stefano Palminteri

Anne Urai @anne-urai.bsky.social · 03/07/2026

#3 of #booksIveRead: Walden Two by BF Skinner. Thanks @stepalminteri.bsky.social for the tip!

The prologue, written in the late 40s, spells out all of the world's problems that have worsened and changed shape, but that have not fundamentally changed: namely those of human behavior, greed, power misuse and our collective inability to create a world for people to thrive.

The novel is fun and easy to read, and promotes a certain type of utopia based on collective living and an education aimed to eliminate feelings of jealousy, property and competition. This would make an interesting read paired with insights from real-world eco-villages, residential meditation halls, and experiments in cooperative living. Certainly worth the read and food for thought.

021

Stefano Palminteri @stepalminteri.bsky.social · 18/06/2026

What is key here is that the evaluation of an idea should be unbiased and efficient. We need a way to eliminate bad ideas sooner in order to avoid saturating the evaluation system. LLMs can make it worst (overflow of papers) or better (automatic screening):an empirical question (not an idealogical)

000

Reposted by Stefano Palminteri

Pascal Mamassian @mamassian.bsky.social · 17/06/2026

Open post-doc position in the study of perceptual fluctuations Position: post-doctoral (2 years) Locations: Paris, France (Ecole Normale Supérieure / CNRS) and Lausanne, Switzerland (University of Lausanne / CHUV) Deadline: September 4th, 2026 Start Date: Fall 2026 (with some flexibility) (1/8)

13017

Reposted by Stefano Palminteri

Shauna Parkes @shaunaparkes.bsky.social · 16/06/2026

Merci à @cnrsbiologie.bsky.social pour le partage de notre récent article publié dans Nature Communications, dans lequel nous montrons que les rats et les humains partagent des biais cognitifs similaires lors d’une tâche de prise de décision 👇 @stepalminteri.bsky.social @neurobordeaux.bsky.social

084

Reposted by Stefano Palminteri

Cognition ENS @cognitionens.bsky.social · 16/06/2026

RECHERCHE I Une étude publiée dans @natcomms.nature.com montre que le biais d’évaluation relative est partagé par l’humain et le rat et repose sur un même mécanisme d’apprentissage ! @shaunaparkes.bsky.social @stepalminteri.bsky.social 👇

021

Reposted by Stefano Palminteri

Stefano Palminteri @stepalminteri.bsky.social · 12/06/2026

Have you ever felt a little guilty about using an LLM to power up your research? If so, I'd argue that it comes from having internalized a wrong model of what science is and what it is for. I develop the argument in my latest Medium piece: medium.com/@stefano.pal...

medium.com

LLMs Are Not Doping, Because Science Is Not a Sport

On why rhe intuition that using AI in research is a form of cheating rests on a wrong picture of how Science works and of what it is for.

1173

Stefano Palminteri @stepalminteri.bsky.social · 12/06/2026

I think we had different cases in mind (mine: LLMs helping design an experiment's logic). But yours is real: if my work feeds a published paper via an LLM and I'm not credited, that needs safeguards, namely the author and publication checks (e.g. pee review). This plagiarism predates LLMs, though.

000

Stefano Palminteri @stepalminteri.bsky.social · 12/06/2026

Absolutely, I don't deny there are ethical problems with using LLMs. Yet from my point of view it would also be unethical not to use all tools in our possession to make Science advance as fast as possible, in health or climate science, for instance. I'm sure you'll agree.

000

Stefano Palminteri @stepalminteri.bsky.social · 12/06/2026

Thanks for reading! In Science, "plagiarism" is either using an unpublished idea (a problem LLMs can't commit, since they can't be trained on unpublished work) or it doesn't really apply, as scientific progress sits badly with intellectual property (ideas are meant to be "picked" to make progress)

100

Stefano Palminteri @stepalminteri.bsky.social · 12/06/2026

medium.com

LLMs Are Not Doping, Because Science Is Not a Sport

On why rhe intuition that using AI in research is a form of cheating rests on a wrong picture of how Science works and of what it is for.

1173

Reposted by Stefano Palminteri

Stefano Palminteri @stepalminteri.bsky.social · 10/06/2026

As someone "mildly" obsessed with books, finally holding the physical copy of our own is quite moving. Thanks to everyone who sent me photos before our copies arrived: it warmed my heart 🙂 Now no more excuses to procrastinate on the next one. Guess the topic from the books serving as bibliography😉

1181

Stefano Palminteri @stepalminteri.bsky.social · 10/06/2026

Close! (and I definitely subscribe to the claim)

110

Stefano Palminteri @stepalminteri.bsky.social · 10/06/2026

1181

Reposted by Stefano Palminteri

Stefano Palminteri @stepalminteri.bsky.social · 05/06/2026

New article out in @natcomms.nature.com Reference-point dependent reinforcement learning in humans and rats By Lachlan Ferguson, Magda Soukupova (co-first), Seb Bouret, me and Shauna Parkes. www.nature.com/articles/s41...

1188

Stefano Palminteri @stepalminteri.bsky.social · 05/06/2026

1188

Stefano Palminteri @stepalminteri.bsky.social · 03/06/2026

New article out in @natcomms.nature.com : Context induces distortions in value representations across multiple elicitation methods and learning modalities, with Magda Soukupova (first author) and @bsgarcia.bsky.social www.nature.com/articles/s41...

02811

Stefano Palminteri @stepalminteri.bsky.social · 01/06/2026

"forevermore hear an italian accent when reading your papers" this is exactly how they are meant to be read :D :D :D

100

Stefano Palminteri @stepalminteri.bsky.social · 29/05/2026

I'll put it like this: not coffee machines

100

Stefano Palminteri @stepalminteri.bsky.social · 29/05/2026

Our book is officially out and there is even an audiobook version! Unfortunately, it’s not narrated by us, so you’ll miss out on our lovely Italian/French accents (a tragic loss for audiobook history). Apparently, Amazon.fr has only three copies left: THANK YOU! www.amazon.fr/Decision-Mak...

23812

Stefano Palminteri @stepalminteri.bsky.social · 28/05/2026

#SBDM2026 @fabiencerrotti.bsky.social Correcting mis-conceptions and shaping preferences about energy sources @maevalhotellier.bsky.social The Nature of Contextual Information Shapes Strategic Decisions @carolinepioger.bsky.social Experential value neglect is robust across representations

090

Stefano Palminteri @stepalminteri.bsky.social · 27/05/2026

@alishiravand.bsky.social : "NORMARL: A Multi-Agent Reinforcement Learning Framework for Adaptive Social Norms in Resource Sustainability" @ #SBDM2026

1134

Reposted by Stefano Palminteri

Stefano Palminteri @stepalminteri.bsky.social · 27/05/2026

My lab will be present en masse at #sbdm2026 Paris. Here is a first sample of the posters, presented by @romanececchi.bsky.social "Dynamic range adaptation in vast decision spaces" and @fabiencerrotti.bsky.social "Correcting mis-conceptions and shaping preferences about energy sources with RL"

0143

Stefano Palminteri @stepalminteri.bsky.social · 27/05/2026

0143

Reposted by Stefano Palminteri

Anne Urai @anne-urai.bsky.social · 27/05/2026

Before the main meeting, we kicked off yesterday with a satellite on cognitive science and sustainability, organized with @magdalenasabat.bsky.social @alishiravand.bsky.social and @stepalminteri.bsky.social magdalenasabat.github.io/pre-sbdm

magdalenasabat.github.io

Magdalena Sabat

141