Valerio Pepe
@valeriopepe.bsky.social
Computer Science + Cognitive Science @harvard.edu, class of '26. Interested in language ∩ thought, language acquisition.
Visiting Student @MITCoCoSci @csail.mit.edu
Visiting Student @MITCoCoSci @csail.mit.edu
Pinned
Valerio Pepe
@valeriopepe.bsky.social
· Nov 16
Hello bsky! I'm Valerio, an undergrad at Harvard studying computer science and cognitive science.
I'm interested in the inductive biases that make language learning and reasoning so easy for us humans, and what their analogues are in machines.
If you're around Boston, I would love to grab coffee!
I'm interested in the inductive biases that make language learning and reasoning so easy for us humans, and what their analogues are in machines.
If you're around Boston, I would love to grab coffee!
Reposted by Valerio Pepe
Do AI agents ask good questions? We built “Collaborative Battleship” to find out—and discovered that weaker LMs + Bayesian inference can beat GPT-5 at 1% of the cost.
Paper, code & demos: gabegrand.github.io/battleship
Here's what we learned about building rational information-seeking agents... 🧵🔽
Paper, code & demos: gabegrand.github.io/battleship
Here's what we learned about building rational information-seeking agents... 🧵🔽
October 27, 2025 at 7:17 PM
Do AI agents ask good questions? We built “Collaborative Battleship” to find out—and discovered that weaker LMs + Bayesian inference can beat GPT-5 at 1% of the cost.
Paper, code & demos: gabegrand.github.io/battleship
Here's what we learned about building rational information-seeking agents... 🧵🔽
Paper, code & demos: gabegrand.github.io/battleship
Here's what we learned about building rational information-seeking agents... 🧵🔽
I'm really excited about this work (two years in the making!).
We look at how LLMs seek out and integrate information and find that even GPT-5-tier models are bad at this, meaning we can use Bayesian inference to uplift weak LMs and beat them... at 1% of the cost 👀
We look at how LLMs seek out and integrate information and find that even GPT-5-tier models are bad at this, meaning we can use Bayesian inference to uplift weak LMs and beat them... at 1% of the cost 👀
Do AI agents ask good questions? We built “Collaborative Battleship” to find out—and discovered that weaker LMs + Bayesian inference can beat GPT-5 at 1% of the cost.
Paper, code & demos: gabegrand.github.io/battleship
Here's what we learned about building rational information-seeking agents... 🧵🔽
Paper, code & demos: gabegrand.github.io/battleship
Here's what we learned about building rational information-seeking agents... 🧵🔽
October 28, 2025 at 7:39 PM
I'm really excited about this work (two years in the making!).
We look at how LLMs seek out and integrate information and find that even GPT-5-tier models are bad at this, meaning we can use Bayesian inference to uplift weak LMs and beat them... at 1% of the cost 👀
We look at how LLMs seek out and integrate information and find that even GPT-5-tier models are bad at this, meaning we can use Bayesian inference to uplift weak LMs and beat them... at 1% of the cost 👀
Reposted by Valerio Pepe
"Although I hate leafy vegetables, I prefer daxes to blickets." Can you tell if daxes are leafy vegetables? LM's can't seem to! 📷
We investigate if LMs capture these inferences from connectives when they cannot rely on world knowledge.
New paper w/ Daniel, Will, @jessyjli.bsky.social
We investigate if LMs capture these inferences from connectives when they cannot rely on world knowledge.
New paper w/ Daniel, Will, @jessyjli.bsky.social
October 16, 2025 at 3:27 PM
"Although I hate leafy vegetables, I prefer daxes to blickets." Can you tell if daxes are leafy vegetables? LM's can't seem to! 📷
We investigate if LMs capture these inferences from connectives when they cannot rely on world knowledge.
New paper w/ Daniel, Will, @jessyjli.bsky.social
We investigate if LMs capture these inferences from connectives when they cannot rely on world knowledge.
New paper w/ Daniel, Will, @jessyjli.bsky.social
Reposted by Valerio Pepe
It’s week 4 and probably time to start doing machine learning in machine learning class. We begin with the only nice thing we have: the perceptron.
Common Descent
Machine learning begins with the perceptron
www.argmin.net
September 23, 2025 at 2:31 PM
It’s week 4 and probably time to start doing machine learning in machine learning class. We begin with the only nice thing we have: the perceptron.
Reposted by Valerio Pepe
"for too long has my foot been allowed to carry my body" I say, as I load a shotgun and aim at it.
August 28, 2025 at 4:26 PM
"for too long has my foot been allowed to carry my body" I say, as I load a shotgun and aim at it.
Reposted by Valerio Pepe
Can LLMs learn social skills by playing games?
A blogpost on human-model interaction, games, training and testing LLMs
research.ibm.com/blog/LLM-soc...
🤖📈🧠
A blogpost on human-model interaction, games, training and testing LLMs
research.ibm.com/blog/LLM-soc...
🤖📈🧠
Can LLMs learn social skills by playing games?
A new open-source framework, TextArena, pits large language models against each other in competitive environments designed to test and improve their communication skills.
research.ibm.com
July 24, 2025 at 2:16 PM
Can LLMs learn social skills by playing games?
A blogpost on human-model interaction, games, training and testing LLMs
research.ibm.com/blog/LLM-soc...
🤖📈🧠
A blogpost on human-model interaction, games, training and testing LLMs
research.ibm.com/blog/LLM-soc...
🤖📈🧠
Reposted by Valerio Pepe
me: "I want my research to be featured in the NYT"
Genie: (grinning)
Genie: (grinning)
June 23, 2025 at 1:30 PM
me: "I want my research to be featured in the NYT"
Genie: (grinning)
Genie: (grinning)
New blog post! www.lesswrong.com/posts/qHudHZ...
Following Emergent Misalignment, we show that finetuning even a single layer via LoRA on insecure code can induce toxic outputs in Qwen2.5-Coder-32B-Instruct, and that you can extract steering vectors to make the base model similarly misaligned 🧵
Following Emergent Misalignment, we show that finetuning even a single layer via LoRA on insecure code can induce toxic outputs in Qwen2.5-Coder-32B-Instruct, and that you can extract steering vectors to make the base model similarly misaligned 🧵
Emergent Misalignment on a Budget — LessWrong
TL;DR We reproduce emergent misalignment (Betley et al. 2025) in Qwen2.5-Coder-32B-Instruct using single-layer LoRA finetuning, showing that tweaking…
www.lesswrong.com
June 8, 2025 at 8:39 PM
New blog post! www.lesswrong.com/posts/qHudHZ...
Following Emergent Misalignment, we show that finetuning even a single layer via LoRA on insecure code can induce toxic outputs in Qwen2.5-Coder-32B-Instruct, and that you can extract steering vectors to make the base model similarly misaligned 🧵
Following Emergent Misalignment, we show that finetuning even a single layer via LoRA on insecure code can induce toxic outputs in Qwen2.5-Coder-32B-Instruct, and that you can extract steering vectors to make the base model similarly misaligned 🧵
Reposted by Valerio Pepe
Sam is 100% correct on this. Indeed, human babies have essential cognitive priors such as permanence, continuity, and boundary of objects, 3D Euclidean understanding of space, etc.
We spent 2 years to systematically to examine and show the lack of such in MLLMs: arxiv.org/abs/2410.10855
We spent 2 years to systematically to examine and show the lack of such in MLLMs: arxiv.org/abs/2410.10855
May 24, 2025 at 5:55 AM
Sam is 100% correct on this. Indeed, human babies have essential cognitive priors such as permanence, continuity, and boundary of objects, 3D Euclidean understanding of space, etc.
We spent 2 years to systematically to examine and show the lack of such in MLLMs: arxiv.org/abs/2410.10855
We spent 2 years to systematically to examine and show the lack of such in MLLMs: arxiv.org/abs/2410.10855
Reposted by Valerio Pepe
I think the BabyLM Challenge is really interesting, but also feel that there is something fundamentally ill-posed about how it maps onto the challenge facing human children. It's true that babies only get a relatively limited amount of linguistic experience, but...
Alona Fyshe @alonaf.bsky.social on the BabyLM Challenge, a competition that trains language models (LMs) on smaller datasets, more akin to how a baby learns, in search of solutions to some of the major challenges of today’s LLMs.
#NeuroAI #neuroskyence
www.thetransmitter.org/neuroai/the-...
#NeuroAI #neuroskyence
www.thetransmitter.org/neuroai/the-...
Can babies inspire more efficient learning algorithms?
A competition that trains language models on smaller datasets, more akin to how a baby learns, seeks solutions to some of LLM’s major challenges.
www.thetransmitter.org
May 19, 2025 at 3:24 PM
I think the BabyLM Challenge is really interesting, but also feel that there is something fundamentally ill-posed about how it maps onto the challenge facing human children. It's true that babies only get a relatively limited amount of linguistic experience, but...
Reposted by Valerio Pepe
Word learning is usually about what a word does refer to. But can toddlers learn from what it doesn’t?
Our new Cognition paper shows 20-month-olds use negative evidence to infer novel word meanings, reshaping theories of language development.
www.sciencedirect.com/science/arti...
Our new Cognition paper shows 20-month-olds use negative evidence to infer novel word meanings, reshaping theories of language development.
www.sciencedirect.com/science/arti...
May 20, 2025 at 4:12 PM
Word learning is usually about what a word does refer to. But can toddlers learn from what it doesn’t?
Our new Cognition paper shows 20-month-olds use negative evidence to infer novel word meanings, reshaping theories of language development.
www.sciencedirect.com/science/arti...
Our new Cognition paper shows 20-month-olds use negative evidence to infer novel word meanings, reshaping theories of language development.
www.sciencedirect.com/science/arti...
Let's go Lio!!!
Francesco Poli – 2024 PhD Thesis “Developing models for learning and exploration” from the University of Cambridge
Lionel Wong – 2024 PhD Thesis “From Words to World: Bridging Language and Thought” from @stanford.edu
Visit cognitivesciencesociety.org/glushko-diss... to learn more!
Lionel Wong – 2024 PhD Thesis “From Words to World: Bridging Language and Thought” from @stanford.edu
Visit cognitivesciencesociety.org/glushko-diss... to learn more!
April 23, 2025 at 12:14 AM
Let's go Lio!!!
saw a new pika model was out on twitter & robot-gasoline-bench does not disappoint
January 28, 2025 at 3:52 PM
saw a new pika model was out on twitter & robot-gasoline-bench does not disappoint
Reposted by Valerio Pepe
All the ACL chapters are here now: @aaclmeeting.bsky.social @emnlpmeeting.bsky.social @eaclmeeting.bsky.social @naaclmeeting.bsky.social #NLProc
November 19, 2024 at 3:48 AM
All the ACL chapters are here now: @aaclmeeting.bsky.social @emnlpmeeting.bsky.social @eaclmeeting.bsky.social @naaclmeeting.bsky.social #NLProc
Reposted by Valerio Pepe
(1/5) Very excited to announce the publication of Bayesian Models of Cognition: Reverse Engineering the Mind. More than a decade in the making, it's a big (600+ pages) beautiful book covering both the basics and recent work: mitpress.mit.edu/978026204941...
November 19, 2024 at 9:33 PM
(1/5) Very excited to announce the publication of Bayesian Models of Cognition: Reverse Engineering the Mind. More than a decade in the making, it's a big (600+ pages) beautiful book covering both the basics and recent work: mitpress.mit.edu/978026204941...
Hello bsky! I'm Valerio, an undergrad at Harvard studying computer science and cognitive science.
I'm interested in the inductive biases that make language learning and reasoning so easy for us humans, and what their analogues are in machines.
If you're around Boston, I would love to grab coffee!
I'm interested in the inductive biases that make language learning and reasoning so easy for us humans, and what their analogues are in machines.
If you're around Boston, I would love to grab coffee!
November 16, 2024 at 10:53 PM
Hello bsky! I'm Valerio, an undergrad at Harvard studying computer science and cognitive science.
I'm interested in the inductive biases that make language learning and reasoning so easy for us humans, and what their analogues are in machines.
If you're around Boston, I would love to grab coffee!
I'm interested in the inductive biases that make language learning and reasoning so easy for us humans, and what their analogues are in machines.
If you're around Boston, I would love to grab coffee!
Reposted by Valerio Pepe
Okay the people requested one so here is an attempt at a Computational Cognitive Science starter pack -- with apologies to everyone I've missed! LMK if there's anyone I should add!
go.bsky.app/KDTg6pv
go.bsky.app/KDTg6pv
November 11, 2024 at 5:27 PM
Okay the people requested one so here is an attempt at a Computational Cognitive Science starter pack -- with apologies to everyone I've missed! LMK if there's anyone I should add!
go.bsky.app/KDTg6pv
go.bsky.app/KDTg6pv
Excited to share this article I authored with Merrick Pierson Smela and George Church at the Wyss Institute! We present SeqVerify, the first automated pipeline for quality-control of whole-genome sequencing data, for all your stem-cell needs!
SeqVerify: A quality-assurance pipeline for whole-genome sequencing data https://www.biorxiv.org/content/10.1101/2023.09.27.559766v1
SeqVerify: A quality-assurance pipeline for whole-genome sequencing data https://www.biorxiv.org/content/10.1101/2023.09.27.559766v1
Over the last decade, advances in genome editing and pluripotent stem cell (PSC) culture have let re
www.biorxiv.org
September 29, 2023 at 10:24 PM
Excited to share this article I authored with Merrick Pierson Smela and George Church at the Wyss Institute! We present SeqVerify, the first automated pipeline for quality-control of whole-genome sequencing data, for all your stem-cell needs!
Reposted by Valerio Pepe
SeqVerify: A quality-assurance pipeline for whole-genome sequencing data https://www.biorxiv.org/content/10.1101/2023.09.27.559766v1
SeqVerify: A quality-assurance pipeline for whole-genome sequencing data https://www.biorxiv.org/content/10.1101/2023.09.27.559766v1
Over the last decade, advances in genome editing and pluripotent stem cell (PSC) culture have let re
www.biorxiv.org
September 28, 2023 at 3:47 PM
SeqVerify: A quality-assurance pipeline for whole-genome sequencing data https://www.biorxiv.org/content/10.1101/2023.09.27.559766v1