How fun, the sky sure is full of stars!

Time flies and I am looking for full time research positions! I am finishing my PhD at Cambridge early 2027. My research focuses on simulating human behavior with LLMs, LLM personalization, evaluating and shaping model behavior, and understanding human–AI interaction. Please get in touch if you are hiring in relavant areas!

User simulators are only as faithful as the user representations they are conditioned on. Missing details force simulators to fill gaps with assumptions, leading to generic behavior⚠️ Introducing UserRepGen-4B, which generates representations optimized for simulation fidelity.

User simulators are trained to talk like users. But users don’t just talk, they judge. We introduce simulators that model both: what users say and what they think of the assistant.

Please come to our poster today afternoon (session #6) at #COLM2026 on *Weightless Fine-tuning for Personalizing LLMs without actual Fine-tuning* Supervised fine-tuning (SFT) is one of the strongest ways to personalize a model to an individual user. (In one of our recent papers with @TuhinChakr and Jane Ginsburg, for example, we used per-author SFT to emulate the writing styles of award-winning authors.) But what if we want to personalize a model to thousands or millions of different authors/users? SFT gets expensive very quickly: - run a separate optimization for each user; - store separate model/LoRA weights for each user; - retrain as new user data arrive.

WildChat, LMSYS, ShareChat… each human-LLM chat dataset aims to capture "real-world" AI use. But do they paint the same picture? Not quite. From user messages alone, a classifier can name a chat's source dataset, and user models trained on them inherit these signatures 🧵👇

I'm at COLM this week, presenting: - LLMs Corrupt Your Documents When You Delegate. At today (Wednesday's) afternoon poster 4.30-6.30. Come check out our very (!) colorful poster. - Giving a fun talk on mostly ongoing work at the Workshop on Lifelong Agents early Friday morning. Old and new friends, let's chat!

I'm on the job market this year! I work on measuring the societal impacts (overreliance, homogenization etc.) of AI/LLMs and proposing ways to mitigate these. I'm at #COLM this week, and #NeurIPS later this year, and would love to chat more!

Why can some people manage multiple coding agents while others can only manage one at a time? Led by @taolongg we worked on understanding this and piloting solutions to bridge the gap arxiv.org/abs/2609.33113

Accepted #NeurIPS2026 The Narrative Gap: Can LLMs help us Navigate Diverse Narratives Across Languages? What information capabilities should #AI systems support to help us navigate diverse multicultural narratives in current and future information ecosystems? Preprint out soon

Excited to share that our paper was accepted to #NeurIPS2026 E&D🎉 and stay tuned for our follow-up work that will be released soon! We present an approach for producing less homogeneous simulations that better preserve the population-level diversity of real users.

We can't build cumulative knowledge about #AI's impact on humans if we can't reproduce studies and evals ‼️ 97% of recent conference papers had replication issues👇 gricea.com is built to solve this problem 📄 arxiv.org/pdf/2609.22039 #HCI #NLProc #OpenScience

We put a 35B ternary LLM on an iPhone—in ~4 GB RAM. Millie beats Bonsai 27B 1-bit across every benchmark we tested, with ~60% faster generation and lower memory use. Works offline. Open weights + runtime Try it on the App Store: apps.apple.com/us/app/millie…

there is no single "correct" behavior for user simulators. the goal is to simulate diverse behaviors, and we want to be as close to the distribution of real users as we can our earlier work shows that there is a large distributional gap, and scaling up assistants do not lead to better user simulators --> 8b trained simulators are very competitive with GPT-5.4 and Gemini 3.1 pro. really cool to see a large-scale effort for user simulation and excited to try out the model!

Excited to share that our paper 𝐃𝐢𝐠𝐢𝐭𝐚𝐥𝐂𝐨𝐚𝐜𝐡 was accepted to the #EMNLP2026 𝐌𝐚𝐢𝐧! 🎉 Agents are getting better at automating software tasks. But can they teach us how to use software ourselves? We studied real expert–novice coaching interactions: 🧵

Introducing WikiSkill📚: compiling agent experience into persistent knowledge for skill evolution! 🧠 Persistent wiki + evolving skills 📈 Evaluated on 5 benchmarks & 5 models 📊 WikiSkill consistently maintain superior performance

Honored to be chosen for the TIME 100 influential people in AI. A year ago, I was fighting tooth and nail to convince people to pay attention to AI detection. We had made huge technical breakthroughs, but it felt like nobody cared. Looking toward the future, it’s a completely different story. As AI becomes more powerful and autonomous, I believe that human touch will be the one of the last remaining scarce resources, and Pangram has an incredible opportunity to help the world place value in humanity. I am grateful to be able to work with such an incredible team, and can't wait to show you all what's next.