🎉New paper @NatureComms: Assessing corporate sustainability with large language models: evidence from Europe

nature.com/articles/s4146…

➡️ New preprint: Debiased Prior-Data Fitted Networks for Amortized Bayesian Semi-Parametric Inference

👉We propose debiased PFN (prior-data fitted networks)

arxiv.org/abs/2610.28314

➡️ New preprint: Generative Policy Improvement from Observational Data

⚠️ Off-policy learning is hard for complex actions (e.g., LLM-generated outputs).
👉We propose #PolicyLift to deal with overlap issues in this case

📄 arxiv.org/abs/2610.23436

➡️New preprint: EAIR: Interaction Risks in Enterprise LLM Multi-Agent Systems

👉In enterprises, AI agents operate under complex organizational structures w/ authority, delegation & access permissions
⚠️However, AI agents can break organizational rules

📄 arxiv.org/abs/2610.14800

➡️New preprint: Efficient Best-of-N policy evaluation for inference-time alignment

👉Best-of-N samples multiple LLM responses and selects the highest-scoring one.
👉We propose a sample-only method (w/ double robust & no-harm budget selection)

Preprint: arxiv.org/abs/2610.09250

➡️New preprint: OrthoGen: A Generative Orthogonal Learner for Time-Varying Treatments

👉 We introduce OrthoGen: a generative causal inference method for estimating full outcome distributions under time-varying treatments

Link: arxiv.org/abs/2610.10210

🧠 New preprint: Can AI Help Explain AI? Effects of Multimodal LLM-Generated Explanations on Perceived Understanding and Perceived Trust in XAI Visualizations

Preprint: papers.ssrn.com/sol3/papers.cf…

➡️New preprint: 𝐏𝐫𝐞𝐝𝐢𝐜𝐭𝐢𝐨𝐧-𝐏𝐨𝐰𝐞𝐫𝐞𝐝 𝐍𝐞𝐮𝐫𝐚𝐥 𝐀𝐫𝐜𝐡𝐢𝐭𝐞𝐜𝐭𝐮𝐫𝐞 𝐒𝐞𝐚𝐫𝐜𝐡

#ppNAS combines 1⃣a few architectures with observed performance labels &2️⃣ cheap proxy information
👉improves NAS performance under limited budgets

arxiv.org/abs/2610.01317

🎉New preprint: Representation Learning for Exact Preimages

arxiv.org/abs/2609.32018

➡️ New preprint: Context-dependent agent evaluation with orthogonal equilibrium learning

👉 #NashEval finds context-specific winners via Nash equilibria, using a novel payoff marix learner (debiased + orthogonal)

arxiv.org/abs/2609.31897

🥳New preprint: Path-specific harm decomposition: A partial identification framework
arxiv.org/abs/2609.29938

🎉We have 5 papers at @NeurIPSConf 2026 (all are on arXiv)

New preprint: Mapping the scientific landscape of AI agents

👉We use an LLM-based mapping to analyze N=64,921 papers on AI agents
🎯Takeaway: 1⃣research is growing; 2⃣deployment gap (only 2.6% reach ops); 3⃣ human-centered metrics (eg trust) rare (7.5%)

papers.ssrn.com/abstract=75000…

🧠New preprint: Designing Against Deskilling: Metacognitive Feedback Reduces Cognitive Offloading to LLM Assistants

👉We compare 2 interventions to reduce offloading to LLMs:
1⃣metacognitive feedback:✅effective
2⃣effort-based rewards:❌no effect

📄arxiv.org/abs/2609.20143

🔍Can LLMs evaluate explainable AI at scale?

New preprint: XAI-Arena — an LLM-as-a-judge framework for XAI evaluation.

✅ LLM ratings align strongly with humans (ρ=.693)
✅ No single XAI method is “best” across all dimensions

arxiv.org/abs/2609.09428

🎉New @NatureEnergyJnl: Applying causal machine learning to assess and improve cleantech policy design

🔍Causal ML offers rigoros ways to estimate when, where & for whom cleantech policies are effective
👉Supports targeted, evidence-based policy design

📄rdcu.be/BC83V7hJsdJa

🚀New Comment @NatureHumBehav: Improving computational reproducibility in the social sciences.

🔍 Why? Published results can fail to reproduce because of simple issues: manual steps, changing software, broken file paths, or missing documentation.

🌱 Our recommendation: 𝐒𝐄𝐄𝐃

𝐒 — Script the full analysis
𝐄 — Encapsulate the environment
𝐄 — Explain code + data
𝐃 — Deposit materials persistently

📄 Paper: rdcu.be/gqvVA7iZHYma

Improving computational reproducibility in the social sciences dlvr.it/TVGX78

🌍 How culturally grounded are multilingual LLMs—really?

Our new preprint:
👉12 LLMs, 3 languages & 89k outputs
👉We find that measured “cultural” differences often depend on task, language, names, and wording.

arxiv.org/abs/2608.23026

🌍 New in @NatureComms: Tracking global health aid disparities with ML

👉 We use #LLMs to analyze 3.7M aid projects (2000–2022; ~$332B)

📊 We compare aid vs disease burden across 17 disease categories to reveal underfunding

nature.com/articles/s4146…