🎉 Updated preprint: Automated reproducibility assessments in the social & behavioral sciences using LLMs. We test whether LLMs can help automate reproducibility checks across N=180 studies Preprint: arxiv.org/abs/2606.13670

🎉 New preprint: OncoSynth 👉We introduce a causally-aware generative ML framework for synthetic oncology cohorts ✅ Preserves structure (covariate→treatment→survival) to improve treatment effect estimation Preprint: arxiv.org/abs/2606.25762

Some new work on AI & science: A reporting checklist for LLMs in behavioral science Thanks to @sfeuerriegel.bsky.social for spearheading. As LLMs become more common in research, transparency is essential. We developed a reporting checklist for reporting how LLMs were used - please share! 1/4

⚽🤖 Can AI predict football? Researchers from @unicologne.bsky.social, Paderborn & @lmu.de benchmark GPT, Claude & Gemini on every 2026 World Cup match — predictions locked, results tracked live. 👉 www.llm-soccerarena.com How this helps management: uni.koeln/6LNX5 #WM #AI @sfeuerriegel.bsky.social

🎉 New preprint: Can #LLMs automate reproducibility checks? 👉We test this on 76 social & behavioral studies ✅Our LLM pipeline arrived at original qualitative conclusions in 96% of cases 🎯We recovered the effect sizes in 41% (human reanalysts: 34%) arxiv.org/abs/2606.13670

⚽️Introducing 𝐋𝐋𝐌 𝐒𝐨𝐜𝐜𝐞𝐫𝐀𝐫𝐞𝐧𝐚: a benchmarking platform for real-world LLM predictions ✅Predictions for 104 games @FIFAWorldCup with frontier LLMs 🏆Opus 4.8: Spain / GPT-5.5: Spain / Grok 4.3: Argentina / Mistral-large: France ➡️All predictions & leaderboard llmsoccerarena.up.railway.app

LLMs are used widely in the behavioral sciences. But we have no good standards for how to do so. We introduce a consensus-based reporting checklist to improve transparency, reproducibility and ethical accountability of LLM-based research in the behavioural sciences. www.nature.com/articles/s41...

📣 Reporting checklist for LLMs in behavioral and social science New article presenting a consensus-based reporting checklist (GUIDE-LLM) for the use of LLMs in the behavioral and social sciences to foster transparency, reproducibility, and ethical use. 🔗 www.nature.com/articles/s41...

Happy to have provided my tiny contribution to this!

Very pleased to have been on the leadership team for this paper! LLMs are already being used all over behavioural science. But it is often pretty hard to work out exactly what has been done, and therefore how much confidence to place in the results. www.nature.com/articles/s41...

📣 New paper @nathumbehav.nature.com: A reporting checklist for large language models in behavioural science www.nature.com/articles/s41...

🎉New preprint: #CausalML to analyze effects of behavioral interventions • We use #CausalML to learn when loss vs. gain framing works • 2 field experiments (N=41,207) from retirement saving • Personalization⬆️participation +51% & savings +18% papers.ssrn.com/sol3/papers....

🏥💊New paper on #CausalML in #oncology: Treatment effect heterogeneity of radiotherapy in localized Ewing sarcoma: A secondary analysis of the EURO-E.W.I.N.G. 99 and Ewing 2008 trial 📄 doi.org/10.1016/j.ej...

📢Checkout our new overview on #CausalML in Wiley #StatsRef 👉We give a concise overview of ML for causal inference — incl. IPTW, AIPTW, TMLE, meta-learners, Neyman orthogonality, ... 📄 doi.org/10.1002/9781... (or PM me) with @larsvanderlaan3.bsky.social @valik-melnychuk.bsky.social

Our papers at #KDD👇 1️⃣Synthetic Interaction Data for Scalable Personalization in Large Language Models arxiv.org/abs/2602.12394 2️⃣Slowing Down Forgetting in Continual Learning (in Manufacturing) arxiv.org/abs/2411.06916 3️⃣Causal Methods for LLM Development and Evaluation arxiv.org/abs/2605.25998

🧠 #AI can improve diagnostic accuracy in radiology – but only if its reasoning is transparent. A new #LMU study found that step-by-step explanations helped physicians better assess and critically evaluate AI recommendations, leading to significantly higher accuracy. 📈 #LMUMunich #Radiology

📣New: 𝐀𝐝𝐚𝐩𝐭𝐢𝐯𝐞 𝐄𝐱𝐩𝐞𝐫𝐢𝐦𝐞𝐧𝐭𝐚𝐭𝐢𝐨𝐧 𝐟𝐨𝐫 𝐂𝐞𝐧𝐬𝐨𝐫𝐞𝐝 𝐒𝐮𝐫𝐯𝐢𝐯𝐚𝐥 𝐎𝐮𝐭𝐜𝐨𝐦𝐞𝐬 👉We develop ASE: an adaptive experimentation framework for survival outcomes with censoring ☑️achieves A-optimal semiparametric efficiency bound 📄 arxiv.org/abs/2605.18459

🎉 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