Google DeepMind's DiffusionGemma Technical Report They feel text diffusion models open up a radically different part of the latency–quality Pareto frontier and hope the report makes it easier for researchers and engineers to understand the model, build on it, and create things we haven’t thought of

Are policy gradient methods hopeless without other machinery? Deep learning works well when the Hessian is well conditioned. But the policy objective is under no obligation to give us that. And no amount of architecture engineering can rescue us from problems w/ the objective.

At the #CCN2026, I present our on-going simulation study: we assessed whether habit and/or range adaptation can explain previously observed and replicated choice frequency effects in the Reward Pairs Task - a recent instrumental learning task developed to elicit habits. More info at the poster C29!

Hello world! 👋 We’re Minds, Machines, and Brains (MMB) 👤🤖🧠 a new open access journal from @mitpress.bsky.social exploring the principles of intelligence and cognition across natural and artificial minds. Submissions open this Fall! 🔗 direct.mit.edu/mmb

How does the brain represent the value of knowledge? In our paper--out now at Nature Neuroscience--we identify a representation of intrinsic information value within neural population structures in the orbitofrontal cortex in mice. www.nature.com/articles/s41593-026-02377-y #neuroskyence 🧠🧪 (1/8)

Representations of the intrinsic value of information in mouse orbitofrontal cortex - Nature Neuroscience

Mice are motivated to seek information of no extrinsic value. Recording neural activity in orbitofrontal cortex reveals a representation of the value of information that is distinct from the value of ...

nature.com

Dendrites keep their own memory of the past and even predict the future independently of the cell body. #neuroskyence By @natmesanash.bsky.social www.thetransmitter.org/memory/dendr...

Dendrites compute independently of cell body

Depending on their distance from the cell body, dendrites play different roles in learning and memory.

thetransmitter.org

We have recently published work in Psych Science and JEP:LMC showing a surprising relationship between initial capture and learned suppression of salient distractors. Ultimately, this has led us to believe that attentional control is much more flexible than often assumed. Read more on my blog!

From Capture to Suppression: How Distraction Shapes Attentional Control

By Nick Gaspelin and Yue Zhang Zhang, Y. & Gaspelin, N. (2026a). From capture to control: Initial capture increases learned suppression. Psychological Science. 37(4), 2…

gaspelinblog.wordpress.com

🚨 New preprint 🚨 How does the brain learn new things without rebuilding what it already knows? www.biorxiv.org/content/10.6... Fantastic work led by @yuyanghuang.bsky.social and Leonie Mehrke, with Tobias Bernklau and @laurabusse.bsky.social #neuroskyence #neuroscience #learning #PFC 🧪🧠 🧵👇

So if I'm understanding this right, a bigger token budget for an employee may become a status symbol? Interesting.

The default network is often thought to interfere with the selective deployment of attention—but our new @pnas.org paper shows a different story… 👉 Find the paper here: doi.org/10.1073/pnas... 1/🧵

Probably should've done this ages ago, but I just setup a Ko-fi page. Extra support is deeply appreciated. I'm paying for hosting out of pocket, now a couple of months into being unemployed, and my birthday is in a couple of days. ko-fi.com/ens0dotme

Support Thorne 🌸

Support Thorne 🌸

ko-fi.com

ken@kenklabnik.bsky.social · last mo.

Is there a ko-fi or similar? Just wondering

An interesting research direction that changes the guts of #AI models and their performance / capacity. Even more so because it can be dropped in with little work (in theory)... research.ibm.com/blog/cofrgen...

CoFrGeNets replace the ‘bones’ of transformer-based models

This new approach, a practical and conceptual shift, points to lighter-weight generative AI models that perform competitively, and in many cases even better.

research.ibm.com

Spread the word if you are interested in a ML bsky community

ICML Conference@icmlconf.bsky.social · 4w ago

We appreciate the love ❤️ But limited engagement and few followers on Bsky makes some (not us!) question whether it's worth it Be sure to follow (and share with your ML friends!) if you want to see more updates from ICML!

We appreciate the love ❤️ But limited engagement and few followers on Bsky makes some (not us!) question whether it's worth it Be sure to follow (and share with your ML friends!) if you want to see more updates from ICML!

Marc Lanctot@sharky6000.bsky.social · 4w ago

Wow, awesome! Massive thank you to the comms & PR team @icmlconf.bsky.social 🤩 ✅️ Multiple update posts per day ✅️ Posts with photos!! 📷 ✅️ Great use of hash tags #icml2026 ✅️ Schedule and QR links at the local mall ✅️ Setting the gold standard for academic Bluesky 👍 🙏

Do people solve hard problems by first building a mental model, then searching for the best solution? With @gershbrain.bsky.social, our new manuscript suggests a different route: people often start by retrieving cached solutions, and scaffold representations around them. osf.io/preprints/ps...

🙋 Looking for emergency reviewers for ARR May / #EMNLP2026 Topics: Interpretability/Explainability, RAG, Answer Attribution, Knowledge Conflicts If you're willing to complete a review before July 6 AoE, please reply or send a DM!

🧵 New paper from lab, CNN shape bias is optimal when training focuses on the object (similar to toddlers) bias peaks. Too much background, or over-cropping that breaks the object → it falls. doi.org/10.1080/0954... 💾 led by @niklasmuller.bsky.social, w/ @cgmsnoek.bsky.social @irisgroen.bsky.social

During sleep, neurons in the hippocampus 'replay' past activity patterns. This is thought to underlie memory consolidation 🧠 After a busy day, how does the brain sort out what is worth replaying? 🤔 Check out this preprint 📃 from Tirole, Duvelle and Bendor for answers! doi.org/10.64898/202...

Time, but not reward, shapes replay-based episodic prioritization

Why are some experiences remembered better than others? Leading theories propose that hippocampal replay prioritizes memories for consolidation according to their expected future value. We recorded hi...

doi.org

Our new preprint is out!! The human brain runs on ~20 watts. Computers and modern AI systems require vastly higher energy. This gap raises a fundamental question: how does the brain compute so efficiently? In our new study, we take a critical step toward measuring and understanding this directly 1/3

bioRxiv Neuroscience@biorxiv-neursci.bsky.social · last mo.

Learning Shapes the Energy Cost of Neural Tasks https://www.biorxiv.org/content/10.64898/2026.07.01.735889v1

I'm excited to share our latest work “Representation Learning Enables Scalable Multitask Deep Reinforcement Learning”. In this work, we revisit a fundamental question in reinforcement learning: What if representation learning was the key ingredient behind scalable RL? 1/🧵

there are serious risks involved with AI. for example, what if the most capable models fall into the wrong hands, who wish to use them to concentrate power, lock down access, hand unchecked control to authoritarian governments, and do secretive fingerprinting to sort users based on national origin?

How can the brain learn the hidden hierarchical structure from high dimensional data? In our latest work, we use synthetic datasets to analyze two classes of bio-plausible learning rules: variants of Direct Feedback Alignment, and local self-supervised learning. We find only the latter succeeds.

Most of the statistical tests you learned in stats class — t-test, ANOVA, correlation — are actually special cases of a single thing: linear regression. Ch 7 of Experimentology argues that thinking in models, not tests, is more flexible and a better foundation for theory. 🧵 experimentology.io

A *super* exciting new paper from @lizsiefert.bsky.social!! Human hippocampal ripple rates track very closely with several indicators of low arousal, during both sleep and wake. This is probably the most compelling data I have ever been associated with.

Liz Siefert@lizsiefert.bsky.social · last mo.

New preprint! We show low arousal states promote hippocampal ripple genesis in sleep and wake. This bridges rodent work, where ripples predominate during sleep, with human studies reporting ripples during active behavior, identifying low arousal as a common mechanism www.biorxiv.org/content/10.6...

@micahgallen.com Really missing those neuro feeds 😅

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

New short series of posts for #ComplexityThoughts – based on my recent papers – concerning a question that has been “consuming” most of my attention over the last years: can “architecture” be a scientific concept for living systems? Read it here 👉 manlius.substack.com/p/decoding-t... 🧵 1/

Decoding the “Architecture” of Living Systems: Chapter 1

DALV-01: Life has no architect, yet it has architecture (shaped by constraints rather than designed like a machine)

manlius.substack.com

What if you could upload and curate GIFs in your AT Proto PDS? All of the memes you care about at your fingertips in a way you control your data. What if you could use those GIFs on non-AT Proto sites too? That's what we're working on at atmomo 🍑

Our latest story is finally in preprint! 🧠✨ How does the brain solve the (hierarchical) credit assignment problem efficiently? Cell-type-specific cortical feedback holds the key, closely approximating backprop. Paper: doi.org/10.64898/202...

Cell-type-specific cortical feedback coordinates hierarchical credit assignment

Learning is thought to arise from synaptic modifications embedded in brain-wide circuits, yet how such circuits coordinate plasticity to support complex behaviour is not known. Inspired by deep learni...

doi.org

We make flexible choices in new situations by knitting together information from separate relevant memories. But what governs which memories are retrieved and when? In a new preprint, we captured how people build decision variables from different memories by tracking their gaze on a blank screen.

Flexible decisions arise from resource-rational memory sampling

Flexible decision making depends on retrieving and recombining memories. Yet because this process unfolds covertly, its governing principles remain unknown. Here we use gaze reinstatement to uncover t...

biorxiv.org