The tutorial program for NeurIPS 2026 is set: the tutorial track chairs announced 19 accepted tutorials on Tuesday, September 29, 2026, chosen from 62 submitted proposals. The 19 tutorials will be split across the conference's three physical sites — Sydney, Paris, and Atlanta — for the first time NeurIPS has run as a multi-site event.

Each tutorial runs 150 minutes (two and a half hours) in person, per the official tutorial requirements — though where they run varies by design. Sydney, the main site, gets a full dedicated tutorial day on December 7 and hosts 12 of the 19 tutorials. Paris gets 4 tutorials, scheduled across December 9–11, and Atlanta gets 3 on December 9. The chairs — Quentin Berthet, Matthew Fahrbach, Yingzhen Li, Krikamol Muandet, and Yisen Wang — say the split was made in stages: tutorials with a hard single-site constraint were placed first, the rest were allocated by balancing authors' site preferences against topic diversity at each venue, and a few were reassigned late when new in-person attendance constraints surfaced.

What made the cut#

The lineup reads like a map of where the field is actually spending its energy. Agents are everywhere: Aadirupa Saha, Arun Verma, and Djallel Bouneffouf will teach a bandit-and-RL blueprint for reliable agentic AI, while Risto Miikkulainen brings a tutorial on the neuroevolution of intelligent agents to Atlanta. Safety and trust get their own slots — Gagandeep Singh's team covers formal guarantees for frontier AI systems, and Anshuman Chhabra, Ben Zhou, and Muhao Chen tackle interpretability, safety, and steering for agentic AI. Evaluation, the discipline that decides whether any of this works, gets Ana Gjorgjevikj's tutorial on robust evaluation of foundation models.

Editorial illustration: a reviewer's desk at night with pinned research paper abstracts, acceptance stamps, and a magnifying glass over a neural-network diagram
AI-generated illustration for this article.

Generative modeling is the other heavy thread. Chieh-Hsin Lai, Subham Sekhar Sahoo, and Stefano Ermon will teach the principles of diffusion models toward fast, scalable diffusion language models — a bet that the next generation of LLMs may not be autoregressive at all. Lars Holdijk, Sirui Lu, and Max Welling connect generative AI to stochastic thermodynamics in a physics-first tutorial in Paris. And uncertainty gets a pragmatic treatment from Ekaterina Fadeeva's team: detecting LLM hallucinations and strengthening reasoning and agents through uncertainty quantification.

Physical AI has its own corner. Adam White, Martha White, and Andrew Patterson will cover empirical design in reinforcement learning for physical AI — reinforcement learning for robots, not chatbots — and Yi Fang's Atlanta tutorial takes on the science of multi-agent communication. Causal representation learning gets a full tutorial from Julius von Kügelgen, Francesco Locatello, and Kun Zhang, and the economics of generative AI gets one from Ander Artola Velasco, Stratis Tsirtsis, and Manuel Gomez Rodriguez.

Why the list matters#

NeurIPS tutorials aren't ordinary talks. The track explicitly bans tutorials that narrowly pitch one lab's tools or results — each has to give a balanced overview of a mature area plus its open problems, aimed at PhD students from outside the subfield. The acceptance bar here (roughly one in three proposals) and the multi-stage review — at least two track chairs per proposal, nomination rounds, joint discussion of borderline cases — mean the 19 titles reflect a field-level consensus on what's worth teaching, not just what's trendy.

Editorial illustration: three wireframe city skylines — Sydney, Paris, Atlanta — connected by luminous data streams carrying brain, robot, and gear icons
AI-generated illustration for this article.

That consensus tilts toward agentic systems, diffusion-based generation, evaluation rigor, and safety guarantees — the same fault lines dominating the industry side of the week. It's also a logistical experiment: reviewing all 62 proposals centrally to keep quality consistent, then distributing 19 tutorials across three continents, tests whether the flagship ML conference can scale its educational program geographically without fragmenting it. The Paris and Atlanta satellites are new this year, so the tutorial track is effectively the dress rehearsal for a tri-site NeurIPS. If attendees in Paris get the same quality as Sydney, expect the 2027 program to look very different.

Sources