Carnegie Mellon at NeurIPS 2025

2026-02-11 · Source: Machine Learning Blog | ML@CMU | Carnegie Mellon University · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

Carnegie Mellon University (CMU) researchers are presenting 156 papers at NeurIPS 2025, covering diverse areas in machine learning. Key oral papers include the Encoder–Attender–Decoder (EAD) framework for tactile processing, MaxSup for overcoming representation collapse in label smoothing, and INFINITY-CHAT, a 26,000-query dataset revealing an "Artificial Hivemind" effect in language models. Other significant contributions feature MeanFlow for one-step generative modeling, OpenCUA for computer-use agents, and UMA, a family of universal models for atoms. Spotlight papers also introduce ARECHO for speech multi-metric estimation, SAILOR for robust imitation learning, KORGym for LLM reasoning evaluation, and CameraBench for camera motion understanding. The research spans applications in computer vision, data-centric AI, deep learning, reinforcement learning, and theoretical aspects, alongside tutorials on hyperparameter optimization, data privacy, imitation learning, and LLM test-time compute.

Key takeaway

For research scientists developing advanced AI systems, these NeurIPS 2025 papers highlight critical areas for innovation and caution. You should consider integrating novel techniques like MaxSup for robust model training or MeanFlow for efficient generative modeling. Be aware of the "Artificial Hivemind" effect in LLMs and the scalability challenges of long decision horizons in RL, guiding your design choices for more diverse and efficient AI.

Key insights

CMU's NeurIPS 2025 contributions advance AI across diverse domains, from fundamental theory to practical applications.

Principles

• Recurrent processing is crucial for cortical tactile computation.
• Label smoothing's error-amplification term drives representation collapse.
• Long decision horizons bottleneck offline RL scalability.

Method

MaxSup regularizes predictions by penalizing the top-1 logit to preserve intra-class diversity. MeanFlow uses average velocity for one-step generative modeling, requiring no pretraining. RLCF extracts checklists from instructions for flexible LLM alignment.

In practice

• Use MaxSup to improve generalization and prevent representation collapse.
• Employ SAILOR for robust imitation learning in robotics.
• Apply SuffixDecoding for up to 3.9x speedups in LLM agent tasks.

Topics

• Large Language Models
• Reinforcement Learning
• Generative Models
• Computer Vision
• Data Attribution & Privacy

Best for: Research Scientist, AI Researcher, AI Scientist, AI Student

Related on AIssential

• NeurIPS 2025 Papers with Code & Data — NeurIPS 2025 papers with public code/data span diverse ML areas, emphasizing open science and practical implementation.
• A Foundation Model for Generative Recommendation, A Unified Definition of Hallucination, and More! — Recent research advances generative recommendation, LLM integration, and efficiency in information retrieval systems.
• LLM Research Papers: The 2026 List (January to May) — LLM research in early 2026 emphasizes hybrid architectures and long-context efficiency beyond simple scaling.
• The Topological Trouble With Transformers — Transformers' feedforward design fundamentally limits dynamic state tracking, requiring a shift to recurrent architectures.
• Paper Digest: ICLR 2026 Papers & Highlights — The ICLR 2026 highlights reveal a strong focus on enhancing LLM reasoning, multi-modal capabilities, and robust AI agent development.
• AI Cartoon of The Year — and five (re)readings for 2026 — The future of AI may shift from pure LLMs to more robust, neurosymbolic approaches addressing current limitations.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning Blog | ML@CMU | Carnegie Mellon University.