-
Reading Minds (Almost): Scientists Just Reconstructed What Someone Sees from Their Brain
Imagine lying in an MRI scanner while looking at a picture of a dog. A computer analyzes your brain activity and, seconds later, generates a rough image of that dog. Not by reading the picture on the screen, but by reading your brain.
-
Princeton University at CogSci 2026
The 48th Annual Meeting of the Cognitive Science Society, or CogSci 2026, was held this week in Rio de Janeiro, Brazil, bringing together researchers in artificial intelligence, linguistics, psychology, neuroscience, and philosophy. This is a roundup of AI-related work from Princeton researchers showcased at the conference.
-
Princeton University at ICML 2026
The Forty-Third International Conference on Machine Learning, or ICML 2026, was held July 6-11 in Seoul, South Korea. This post is a roundup of work from Princeton researchers that was showcased at the conference.
-
Self-supervision unlocks depth scaling in reinforcement learning – and results in unanticipated exploratory behaviors
For years, reinforcement learning has typically relied on relatively shallow neural network architectures. In language and vision, researchers increased depth to hundreds of layers and observed the emergence of new capabilities. In reinforcement learning (RL), by contrast, most architectures used two to five layers, occasionally up to eight.
-
Princeton University at CVPR 2026
The Conference on Computer Vision and Pattern Recognition was held June 3-7 in Denver. This post is a roundup of work from Princeton researchers that was showcased at the conference.
-
I am, myself, a lossy compression
Somewhere on your phone right now, there’s a photo that takes up about 3 megabytes. If you’d taken the same photo as a raw bitmap — every pixel’s red, green, and blue values written out directly — it would be closer to 25 megabytes.
-
From Vision-Language Models to Robot Control — Without Forgetting
The promise of vision-language-action models is straightforward: Take a foundation model that understands the world and teach it to act in it. The challenge is that learning to act often erases that understanding.
-
Princeton University at ICLR 2026
The Fourteenth International Conference on Learning Representations, abbreviated as ICLR 2026, was held in April in Rio de Janeiro, Brazil. This post is a roundup of work from Princeton students, post-docs, research software engineers and faculty that was showcased.
-
AI Agents that Generate Knowledge Must Learn via Interaction
This blog post is based on our work BuilderBench, which is a benchmark to accelerate research into training that centers around exploration. The vision for BuilderBench is to enable an open-ended stream of potential interactions, where pre-training could only ever cover a tiny slice of all possible behaviors.
-
Introducing BeyondOBjects, a New Method to Improve Image Classification Models
Recent powerful text-to-image (T2I) models, such as Stable Diffusion, are very capable of generating diverse and realistic images. While they are commonly used for creating creative edits of images, a natural question arises: Can we use these powerful models as a tool to create a synthetic dataset used to train downstream classification models?
-
FlashAttention-4: Algorithm and Kernel Pipelining Co-Design for Asymmetric Hardware Scaling
Transformers remain the backbone for most AI applications, from large language models to vision and multimodal systems. For transformers, attention is the primary computational bottleneck, with self-attention scores between queries and keys scaling quadratically with sequence length.
-
The Book of Life: Moving from a Sociology of Variables to a Sociology of Events
In Dec 1946, a paper titled “Record Linkage” appeared in the pages of the American Journal of Public Health. It began with these words: “Each person in the world creates a Book of Life. This Book starts with birth and ends with death.”
-
Princeton University at NeurIPS 2025
The 39th Conference on Neural Information Processing Systems, abbreviated as NeurIPS 2025, was held in December. This post a list of work from Princeton students, post-docs, research software engineers and faculty that was showcased at the conference.
-
How Geometry Might Unlock Human-Like Intelligence in AI: Why Your Brain Gets Lunch (and Your AI Can’t)
Say you walk into your favorite diner. You know the drill – find a seat, order a grilled cheese, maybe throw in a side of fries. The tables might be a bit shuffled, the server might be new, but despite these little changes, you still know exactly what’s going on.
-
AI Co-Drivers Shouldn’t Drive You Off the Track: A New Filter Makes Sure They Don’t
Ever had a driver-assist system swoop in at the last possible instant? Effective, sure, but not exactly confidence-inspiring. In high-speed settings, the first challenge is reliably preventing accidents in real time. The second challenge is to do that without spooking the driver.
-
Can Your Therapist’s Words Help You Heal? Meet the AI That Knows
Imagine you’re in therapy. You’re telling your therapist: “I feel like I’m drowning in stress; I can’t stop thinking about work.” Now imagine your therapist replies: “Take a moment and picture yourself watching this moment from the outside. What would future you say?”
-
Unfolding Genius: When AI and Kirigami Sculpt the Future, One Cut at a Time
Start with a flat sheet of metal. Not much to look at, really. Maybe a little glare from the overhead lights, maybe the smell of laser-cut steel or brushed aluminum if you lean in close. But then, introduce a few cuts, some elegant decisions, and a gentle twist.
-
When Your AI Assistant Gets Gaslit: How “Real AI Agents with Fake Memories” Shows That Smart Assistants Can Go Dangerously Off-Script
Imagine this: You give your AI assistant access to your credit card to book a flight. The next day, it has gone ahead and purchased a vintage yacht, a lifetime supply of ergonomic chairs, and several questionable NFTs. No one hacked it. It just misremembered something someone once said on Discord.
-
Deep Dive Series: Using Tools of Cognitive Science to Decipher Modern Artificial Intelligence
We know that even the best large language models can hallucinate — a euphemism for when LLMs produce inaccurate answers to user queries. If that weren’t problematic enough, today’s frontier language models have also been shown to actively engage in deception.
-
Deep Dive Series: Building Biosecurity Safeguards into AI for Science
In 1962, Max Perutz and John Kendrew received the Nobel Prize in Chemistry for figuring out for the first time ever the three-dimensional shapes of two protein molecules. Pertuz studied hemoglobin and Kendrew studied the somewhat smaller myoglobin, proteins involved in shuttling oxygen to tissues in our bodies.
-
Machine Learning Helps Reveal What Makes Games Complex for People
Human behavior is famously variable across all kinds of judgments and decisions, so pinning down “laws” of human behavior is much tougher than writing the laws of physics, such as gravity or thermodynamics. Today, though, psychologists and economists have a powerful new toolkit for better discovering the principles behind human behavior: large-scale online experiments paired…
-
Seed Grant Series: Article Friend – Developing a New, AI-Powered Tool to Increase Accessibility of Research Articles
Aphasia is an acquired disorder of communication—not of intellect—caused by damage to areas of the brain that support language processing. Aphasia most commonly occurs after left-hemisphere stroke and is highly prevalent—nearly 2 million individuals in the United States alone are currently living with aphasia.
-
Deep Dive Series: Can LLMs Think About Thinking, and Can We Leverage Such Traits?
In 1979, the American developmental psychologist John H. Flavell coined the term metacognition, defining it as “knowledge and cognition about [one’s own] cognitive phenomena.” In a study of preschool and elementary school children, he showed that while the older kids displayed metacognitive abilities, the younger ones didn’t.
-
Lugha-Llama: Adapting Large Language Models for African Languages
Low-resource African languages remain underrepresented in the large training datasets of large language models (LLMs) and, as a result, LLMs struggle to understand these languages. We are releasing three African-centric Lugha-Llama models based on Llama-3.1-8B, which achieve the best performance among open-source models on IrokoBench, a challenging African languages benchmark and AfriQA, a cross-lingual open-retrieval…
-
Designing for Equivariance to Perceptual Variation in Images
A huge challenge for neural networks is generalizing to out-of-distribution data. For example, a neural network classifier trained on blue digits may perform well when presented with a previously unseen blue digit, but fail to correctly classify green digits.
-
Robustifying ML-powered Network Classifiers with PANTS
Machine learning is invaluable for managing computer networks (e.g., enterprise networks or data centers), enabling pattern recognition in traffic that facilitates efficient resource allocation and accurate detection of network threats. However, its vulnerability to adversarial attacks poses significant risks.
-
Introducing the AI Lab Research Blog
Welcome! Our blog is a place for researchers across the Princeton University community to share their latest work related to artificial intelligence. Posts will be primarily written by students, post docs and faculty, featuring projects that span the natural sciences, engineering, social sciences and humanities. Our first post examines a problem at the intersection of…