🎉 New paper: Amortizing causal sensitivity analysis via prior data-fitted networks tl;dr: We turn causal sensitivity analysis into in-context learning arxiv.org/abs/2605.10590

🎉 New paper: 𝐒𝐤𝐢𝐥𝐥𝐆𝐞𝐧 — Verified inference-time agent skill synthesis 👉 Skills can improve LLM agents — but are still mostly hand-written. We introduce #SkillGen: a multi-agent framework that learns skills from trajectories using contrastive induction + a generate–verify–refine loop. 🧵1/3

🎉New paper: #ConfoundingSHAP 🔎Which variables actually drive confounding? 👉We introduce a Shapley-based method to attribute confounding strength to covariates 📄 arxiv.org/abs/2605.10533

🎉 New paper: OrthoBO: Orthogonal Bayesian Hyperparameter Optimization Link: arxiv.org/abs/2605.06454 We identify a hidden failure mode in Bayesian optimization: 👉 acquisition functions are noisy estimates (e.g. via MC) Small noise → flipped rankings → bad decisions /🧵

🧠Do AI coding assistants make developers better? Meta-analysis (23 papers): 1️⃣Productivity ✅modest gain (g≈0.33), strong in lab settings 2️⃣Real-world impact: ⚠️near zero in complex settings (open-source, enterprise) 3️⃣Learning: ❌no gain (g≈0.14) arxiv.org/pdf/2605.04779

[1/5] 🇲🇦 Today I’m presenting our paper “Orthogonal Representation Learning for Estimating Causal Quantities” at AISTATS 2026 in Tangier, Morocco; and I’m very happy that it was accepted as an oral.

🎉Check out our 7 papers at #ICML2026 👇@icmlconf.bsky.social

🚀New paper @ npj digital medicine: 𝐓𝐡𝐞 𝐞𝐟𝐟𝐞𝐜𝐭 𝐨𝐟 𝐦𝐞𝐝𝐢𝐜𝐚𝐥 𝐞𝐱𝐩𝐥𝐚𝐧𝐚𝐭𝐢𝐨𝐧𝐬 𝐟𝐫𝐨𝐦 𝐥𝐚𝐫𝐠𝐞 𝐥𝐚𝐧𝐠𝐮𝐚𝐠𝐞 𝐦𝐨𝐝𝐞𝐥𝐬 𝐨𝐧 𝐝𝐢𝐚𝐠𝐧𝐨𝐬𝐭𝐢𝐜 𝐝𝐞𝐜𝐢𝐬𝐢𝐨𝐧𝐬 𝐢𝐧 𝐫𝐚𝐝𝐢𝐨𝐥𝐨𝐠𝐲 🔍What is the right format to present explanations in medical #LLM? 👉We ran a large experiment with 101👩‍⚕️(N=2020 diagnostic decisions) www.nature.com/articles/s41...

[1/5] 🇧🇷 Looking forward to presenting our new paper, “Overlap-Adaptive Regularization for Conditional Average Treatment Effect Estimation,” at ICLR 2026 in Rio de Janeiro. The paper tackles a core challenge in causal ML: estimating treatment effects when overlap is low.

[1/4] 🇧🇷 Excited to share our new paper, “GDR-learners: Orthogonal Learning of Generative Models for Potential Outcomes,” which I’ll be presenting at ICLR 2026. The goal: better estimate potential outcome distributions from observational data — not just averages, but the full uncertainty.

If you are around #ICLR2026, check out our papers 👇

If you are around @chi.acm.org, check out our papers 👇 👉 Partnering with Generative AI: Experimental Evaluation of Model-Led and Human-Led Interaction in Human-AI Co-Creation arxiv.org/abs/2507.23492

🚀Introducing 𝐆𝐔𝐈𝐃𝐄-𝐋𝐋𝐌: A reporting checklist for using LLMs in behavioral & social science ✅GUIDE-LLM is a reporting checklist designed by 80+ experts to improve transparency, reproducibility & ethical accountability of LLM-based research 📄 llm-checklist.com