RIKEN AIP Researchers Land 76 Papers at NeurIPS 2026

The acceptance numbers

According to RIKEN AIP, the Neural Information Processing Systems Conference (NeurIPS) is one of the premier international conferences in machine learning and artificial intelligence. This year the NeurIPS Main Track received 30,709 valid paper submissions with a PDF. The program committee accepted 7,900 papers in total, which the institute puts at 25.7%. The figure of 7,900 carries an asterisk in the announcement, but the excerpt does not explain the note.

The institute's Center for Advanced Intelligence Project published the list on October 14, 2026. The text provided for this article covers only part of the full list of 76 papers, so this report is limited to the entries visible there.

Oral and spotlight papers

Two papers are listed as Main Track Oral presentations. The first is Can LLM Agents Respond to Disasters? Benchmarking Heterogeneous Geospatial Reasoning in Emergency Operations, led by Junjue Wang of The University of Tokyo and RIKEN AIP, with co-authors including Stefano Ermon of Stanford University and Naoto Yokoya of The University of Tokyo and RIKEN AIP. The second is From Average Sensitivity to Small-Loss Regret Bounds under Random-Order Model by Shinsaku Sakaue and Yuichi Yoshida.

Four papers are listed as Spotlights:

• Non-asymptotic Implicit Bias of Logistic Regression at Early-stage Gradient Descent Dynamics, by Han Bao.
• Toward Minimal-dimensional Convex Calibrated Surrogate Losses for Classification with Rejection, by Yuzhou Cao, Han Bao and Bo An.
• The Implicit Bias of Hyperbolic Representation Learning for Multiclass Data: A Busemann Risk Perspective, by Xingrun Li and colleagues from The University of Tokyo.
• Post-ADC Inference: Valid Inference After Active Data Collection, by Shuichi Nishino and colleagues from Nagoya University and RIKEN AIP.

Poster topics

The poster list shows a broad spread of current machine learning topics. Several papers deal with large language models and agents. Examples include Do Coding Agents Deceive Us? Detecting and Preventing Cheating via Capped Evaluation with Randomized Tests, When and Why Does Multi-Agent Debate Fail and Does It Really Underperform?, UnlearningSoup: Is Repeated Tuning Necessary for Large Language Model Unlearning? and Reliable Chain-of-Thought via Prefix Consistency.

Others concern reinforcement learning and generative models. These include Dataset Mismatch Matters in Group Relative Policy Optimization for Reinforcement Learning from Verifiable Rewards, VI-CuRL, Watermarking as a Learned Intrinsic Property of Diffusion Models and Flow Matching from Viewpoint of Proximal Operators. Further entries cover theory and statistics, such as bandit algorithms, fairness-accuracy tradeoffs, multicalibration and anytime-valid inference for LLM self-consistency.

Masashi Sugiyama of RIKEN AIP and The University of Tokyo appears as an author on many of the poster papers. Many entries are collaborations with other institutions, among them Stanford University, the University of Oxford, Carnegie Mellon University, Tsinghua University, Microsoft Research and Mohamed bin Zayed University of Artificial Intelligence.

What it shows

The announcement is a list of accepted work rather than a results report, so it offers no performance claims. It does show how many of the accepted papers are joint efforts between RIKEN AIP and Japanese universities, companies such as CyberAgent and Preferred Networks, and groups abroad. With an acceptance rate of 25.7%, the 76 papers represent work that passed a highly competitive review at one of the field's largest venues.