カーネギーメロン大学、ICLR 2026に194論文を発表

jQuery('.post-authors').width(0.33 * jQuery('.post-authors').parent().width());

jQuery('.affiliations').width(0.33 * jQuery('.affiliations').parent().width());

jQuery('.date').width(0.33 * jQuery('.date').parent().width());

jQuery('.post-authors').empty();

jQuery('.post-authors').append('Authors

');

jQuery('.post-authors').append('Naveen Raman, Kiriaki Fragkia');

jQuery('.affiliations').append('MLD, CMU

');

jQuery('.doi').remove();

CMU researchers are presenting 194 papers at the Fourteenth International Conference on Learning Representations (ICLR 2026), held from April 23rd-April 27th at the Riocentro Convention and Event Center in Rio de Janeiro, Brazil. Here is a quick overview of the areas our researchers are working on:

Here are our most frequent collaborator institutions:

body { font-family: Arial, sans-serif; margin: 20px; line-height: 1.6; }

h1 { text-align: center; }

h2 { color: #333; }

h3 { color: #555; margin-top: 20px; }

.paper { margin-bottom: 20px; }

.paper-title { font-weight: bold; }

.paper-authors { font-style: italic; margin-bottom: 10px; }

.oral-spotlight-space { margin-top: 10px; font-style: normal; }

.table-of-contents { margin-bottom: 30px; }

.table-of-contents ul { list-style-type: none; padding-left: 0; }

.table-of-contents ul ul { padding-left: 20px; }

Table of Contents

Oral Papers

Poster Papers

Applications

Computer Vision

Deep Learning

General Machine Learning

Optimization

Reinforcement Learning

Social Aspects

Theory

Uncategorized

Oral Papers

EditBench: Evaluating LLM Abilities to Perform Real-World Instructed Code Edits

Authors: Wayne Chi (CMU), Valerie Chen (Carnegie Mellon University), Ryan Shar (Apple), Aditya Mittal (CMU, Carnegie Mellon University), Jenny Liang (School of Computer Science, Carnegie Mellon University), Wei-Lin Chiang (UC Berkeley / LMSYS), Anastasios Angelopoulos (University of California Berkeley), Ion Stoica (), Graham Neubig (Carnegie Mellon University), Ameet Talwalkar (University of California-Los Angeles), Chris Donahue (CMU / Google DeepMind)

This work introduces EditBench, a new benchmark for testing how well AI models can edit existing code based on user instructions. Unlike prior benchmarks, it uses real-world coding tasks and contexts, including things like the surrounding code and cursor position. The benchmark includes 545 diverse problems, and results show that most models struggle—only a few achieve strong performance. The study also finds that having more realistic context significantly impacts how well models perform, highlighting the importance of evaluating code-editing in real-world settings.

UALM: Unified Audio Language Model for Understanding, Generation and Reasoning

Authors: Jinchuan Tian (CMU, Carnegie Mellon University), Sang-gil Lee (NVIDIA), Zhifeng Kong (NVIDIA), Sreyan Ghosh (Nvidia), Arushi Goel (NVIDIA), Chao-Han Huck Yang (NVIDIA Research), Wenliang Dai (NVIDIA), Zihan Liu (Nvidia), Hanrong Ye (NVIDIA), Shinji Watanabe (Carnegie Mellon University), Mohammad Shoeybi (NVIDIA), Bryan Catanzaro (NVIDIA), Rafael Valle (NVIDIA), Wei Ping (Nvidia)

This paper introduces the Unified Audio Language Model (UALM), a single model designed to handle audio understanding, text-to-audio generation, and multimodal reasoning together. Instead of treating these as separate tasks, UALM learns to both interpret and generate audio, achieving performance comparable to specialized state-of-the-art models. The authors also show that combining text and audio during the model’s reasoning process improves its ability to handle complex tasks. Overall, the work demonstrates a step toward more general AI systems that can reason across both language and sound.

Agent Data Protocol: Unifying Datasets for Diverse, Effective Fine-tuning of LLM Agents

Authors: Yueqi Song (CMU), Ketan Ramaneti (Amazon), Zaid Sheikh (Carnegie Mellon University), Ziru Chen (Ohio State University, Columbus), Boyu Gou (Ohio State University, Columbus), Tianbao Xie (the University of Hong Kong, University of Hong Kong), Yiheng Xu (University of Hong Kong), Danyang Zhang (Shanghai Jiao Tong University), Apurva Gandhi (Carnegie Mellon University), Fan Yang (Fujitsu), Joseph Liu (School of Computer Science, Carnegie Mellon University), Tianyue Ou (Carnegie Mellon University), Zhihao Yuan (Carnegie Mellon University), Frank F Xu (Carnegie Mellon University), Shuyan Zhou (Facebook), Xingyao Wang (All Hands AI), Xiang Yue (Carnegie Mellon University), Tao Yu (University of Hong Kong), Huan Sun (Ohio State University), Yu Su (Ohio State University), Graham Neubig (Carnegie Mellon University)

This work introduces the Agent Data Protocol (ADP), a standardized format for representing training data for AI agents. The authors argue that the main challenge isn’t a lack of data, but that existing datasets are fragmented across different formats and tools. ADP acts as a common “interlingua,” making it easier to combine diverse data sources—like coding, browsing, and tool use—into a single training pipeline. By converting 13 datasets into this unified format, the authors show that models trained on the combined data achieve improved performance.

MotionStream: Real-Time Video Generation with Interactive Motion Controls

Authors: Joonghyuk Shin (Seoul National University), Zhengqi Li (Google), Richard Zhang (Adobe), Jun-Yan Zhu (Carnegie Mellon University), Jaesik Park (Seoul National University), Eli Shechtman (Adobe), Xun Huang (Adobe Research)

This paper introduces MotionStream, a system for generating videos in real time based on motion and text inputs. Unlike prior methods that take minutes to produce a video, MotionStream can stream results at up to 29 frames per second on a single GPU. The key idea is to train a fast, causal model that can generate video continuously, using techniques that prevent quality from degrading over long sequences. As a result, users can interactively control motion—like drawing paths or moving a camera—and see the video update instantly.

OpenThoughts: Data Recipes for Reasoning Models

Authors: Etash Guha (Stanford University, Anthropic), Ryan Marten (Harbor), Sedrick Keh (Toyota Research Institute), Negin Raoof (University of California, Berkeley), Georgios Smyrnis (University of Texas, Austin), Hritik Bansal (University of California, Los Angeles), Marianna Nezhurina (Juelich Supercomputing Center, LAION, Tuebingen University), Jean Mercat (Toyota Research Institute (TRI)), Trung Vu (Google), Zayne Sprague (New York University), Ashima Suvarna (UCLA), Benjamin Feuer (Stanford University), Leon Liangyu Chen (Stanford University), Zaid Khan (University of North Carolina at Chapel Hill), Eric Frankel (Department of Computer Science, University of Washington), Sachin Grover (Arizona State University), Caroline Choi (None), Niklas Muennighoff (Stanford University), Shiye Su (Stanford University), Wanjia Zhao (Stanford University), John Yang (Princeton University), Shreyas Pimpalgaonkar (New York University), Kartik sharma (Georgia Institute of Technology), Charlie Ji (University of California, Berkeley), Yichuan Deng (Department of Computer Science, University of Washington), Sarah Pratt (University of Washington), Vivek Ramanujan (Department of Computer Science, University of Washington), Jon Saad-Falcon (Computer Science Department, Stanford University), Stutee Acharya (University of South Florida), Jeffrey Li (Carnegie Mellon University), Achal Dave (Anthropic), Alon Albalak (SynthLabs), Kushal Arora (McGill University), Blake Wulfe (Toyota Research Institute), Chinmay Hegde (New York University), Greg Durrett (New York University), Sewoong Oh (University of Washington), Mohit Bansal (UNC Chapel Hill), Saadia Gabriel (University of Washington), Aditya Grover (UCLA), Kai-Wei Chang (University of Virginia Main Campus), Vaishaal Shankar (Apple), Aaron Gokaslan (Cornell University), Mike Merrill (None), Tatsunori Hashimoto (Stanford University), Yejin Choi (Stanford University / NVIDIA), Jenia Jitsev (LAION; Juelich Supercomputing Center, Research Center Juelich), Reinhard Heckel (Technical University Munich), Maheswaran Sathiamoorthy (University of Southern California), Alex Dimakis (Electrical Engineering & Computer Science Department, University of California, Berkeley), Ludwig Schmidt (University of Washington / Stanford / Anthropic)

This work introduces the OpenThoughts project, which aims to create high-quality, open-source datasets for training reasoning-focused AI models. The authors show that models trained on their public data can match or exceed the performance of strong existing systems that rely on private datasets. By carefully studying and improving their data generation process, they build larger and better datasets that significantly boost performance across math, coding, and science benchmarks. Overall, the project demonstrates that open data alone can be enough to train highly capable reasoning models.

Mamba-3: Improved Sequence Modeling using State Space Principles

Authors: Aakash Sunil Lahoti (CMU, Carnegie Mellon University), Kevin Li (Carnegie Mellon University), Berlin Chen (Princeton University), Caitlin Wang (Princeton University), Aviv Bick (Carnegie Mellon University), Zico Kolter (Carnegie Mellon University), Tri Dao (Princeton University), Albert Gu (Cartesia AI CMU)

This paper introduces Mamba-3, a new model designed to make AI inference faster and more efficient without sacrificing performance. While many efficient alternatives to Transformers reduce computation, they often struggle with tasks like tracking long-term information; Mamba-3 addresses this with improved state modeling and a more expressive update mechanism. The model also uses a multi-input, multi-output design to boost accuracy without slowing down generation. Overall, Mamba-3 shows that it’s possible to improve both efficiency and capability at the same time, pushing forward the tradeoff between speed and performance.

Overcoming Joint Intractability with Lossless Hierarchical Speculative Decoding

Authors: Yuxuan Zhou (Independent Researcher), Fei Huang (Alibaba Group), Heng Li (Carnegie Mellon University), Fengyi Wu (University of Washington), Tianyu Wang (University of Washington), Jianwei Zhang (Alibaba Group), Junyang Lin (Alibaba Group), Zhi-Qi Cheng (University of Washington)

This paper introduces Hierarchical Speculative Decoding (HSD), a new method to speed up large language model inference by improving the verification step in speculative decoding while preserving exact output distributions. It addresses the challenge of “joint intractability” in sequence-level verification by organizing resampling into a hierarchy that redistributes probability mass across branches, enabling more tokens to be accepted at once. The approach is theoretically proven to be lossless and empirically shows consistent speed improvements across models and benchmarks, outperforming prior tokenwise and blockwise verification methods. Overall, HSD offers a practical and general way to accelerate decoding without sacrificing fidelity, achieving state-of-the-art efficiency when integrated into existing frameworks.

Distributional Equivalence in Linear Non-Gaussian Latent-Variable Cyclic Causal Models: Characterization and Learning

Authors: Haoyue Dai (Carnegie Mellon University), Immanuel Albrecht (FernUniversität in Hagen), Peter Spirtes (Carnegie Mellon University), Kun Zhang (Carnegie Mellon University & MBZUAI)

This paper studies causal discovery in linear non-Gaussian models with latent variables and cycles, focusing on when different causal graphs are observationally indistinguishable. It provides the first general characterization of distributional equivalence in this setting, introducing new tools—especially edge rank constraints—to describe when two models generate the same observed data. Building on this theory, the authors derive practical graphical criteria and transformations to enumerate all equivalent models and propose an algorithm to recover the entire equivalence class from data. Overall, the work removes the need for strong structural assumptions and offers a general, principled framework for latent-variable causal discovery.

Revela: Dense Retriever Learning via Language Modeling

Authors: Fengyu Cai (Technische Universität Darmstadt), Tong Chen (University of Washington), Xinran Zhao (Carnegie Mellon University), Sihao Chen (Microsoft), Hongming Zhang (Tencent AI Lab Seattle), Sherry Wu (Carnegie Mellon University), Iryna Gurevych (Technical University of Darmstadt / Mohamed bin Zayed University of Artificial Intelligence), Heinz Koeppl (TU Darmstadt)

This paper introduces Revela, a self-supervised framework for training dense retrievers by leveraging language modeling objectives instead of relying on annotated query-document pairs. It augments next-token prediction with an in-batch attention mechanism that allows documents to attend to each other, enabling the retriever to learn cross-document relationships jointly with a language model. Experiments across domain-specific, reasoning-intensive, and general benchmarks show that Revela matches or surpasses supervised and API-based retrievers while using significantly less data and compute. Overall, the work demonstrates a scalable and efficient alternative for retriever learning directly from raw text with strong generalization across domains.

Latent Particle World Models: Self-supervised Object-centric Stochastic Dynamics Modeling

Authors: Tal Daniel (Carnegie Mellon University), Carl Qi (University of Texas at Austin), Dan Haramati (Brown University), Amir Zadeh (Lambda), Chuan Li (Lambda Labs), Aviv Tamar (Technion), Deepak Pathak (Carnegie Mellon University), David Held (Carnegie Mellon University)

This paper introduces the Latent Particle World Model (LPWM), a self-supervised, object-centric world model that learns to decompose scenes into latent particles (e.g., keypoints, masks, and object attributes) directly from raw video without supervision. It proposes a novel per-particle latent action mechanism that models stochastic dynamics, enabling the system to capture complex multi-object interactions and generate diverse future predictions. The model is trained end-to-end and supports flexible conditioning on actions, language, and goal images, achieving state-of-the-art performance on both real-world and synthetic video prediction tasks. Beyond video modeling, LPWM also demonstrates strong potential for decision-making applications such as imitation learning by leveraging its learned latent dynamics.

LoongRL: Reinforcement Learning for Advanced Reasoning over Long Contexts

Authors: Siyuan Wang (Shanghai Jiao Tong University), Gaokai Zhang (Carnegie Mellon University), Li Lyna Zhang (Microsoft Research Asia), Ning Shang (Microsoft), Fan Yang (Microsoft Research), Dongyao Chen (Shanghai Jiaotong University), Mao Yang (Peking University)

The authors introduce LoongRL, a reinforcement learning framework designed to improve long-context reasoning in large language models by training them on challenging, synthesized tasks. They propose KeyChain, a data construction method that embeds hidden question chains within long documents, forcing models to perform multi-step planning, retrieval, and reasoning rather than relying on shortcuts. Through RL training, models develop an emergent “plan–retrieve–reason–recheck” reasoning pattern that generalizes from shorter (16K) to much longer (128K) contexts. Experiments show that LoongRL significantly boosts long-context reasoning performance while maintaining strong short-context abilities, achieving results comparable to much larger models.

Exchangeability of GNN Representations with Applications to Graph Retrieval

Authors: Kartik Nair (Carnegie Mellon University), Indradyumna Roy (IIT Bombay, Aalto University), Soumen Chakrabarti (IIT Bombay), Anirban Dasgupta (IIT Gandhinagar), Abir De (Indian Institute of Technology Bombay)

This paper introduces the concept of exchangeability in graph neural networks (GNNs), showing that the dimensions of learned node embeddings are statistically interchangeable due to random initialization and permutation-invariant training. This property implies that embedding components share identical distributions, enabling simplifications in how graph similarities are computed. Leveraging this insight, the authors approximate complex transportation-based graph distances using simpler Euclidean operations on sorted embedding values. They further propose GRAPHHASH, a locality-sensitive hashing framework that enables efficient and scalable graph retrieval, achieving strong performance compared to existing methods.

Poster Papers

Applications

TusoAI: Agentic Optimization for Scientific Methods

Authors: Alistair Turcan (School of Computer Science, Carnegie Mellon University), Kexin Huang (Stanford University), Lei Li (School of Computer Science, Carnegie Mellon University), Martin J. Zhang (Carnegie Mellon University)

AutoLibra: Agent Metric Induction from Open-Ended Human Feedback

Authors: Hao Zhu (Carnegie Mellon University), Phil Cuvin (Stanford University), Xinkai Yu (University of Pennsylvania, University of Pennsylvania), Charlotte Yan (Stanford University), Jason Zhang (Stanford University), Diyi Yang (Stanford University)

DistMLIP: A Distributed Inference Platform for Machine Learning Interatomic Potentials

Authors: Kevin Han (Carnegie Mellon University), Bowen Deng (UC Berkeley), Amir Barati Farimani (CMU, Carnegie Mellon University), Gerbrand Ceder (University of California, Berkeley)

Vlaser: Vision-Language-Action Model with Synergistic Embodied Reasoning

Authors: Ganlin Yang (University of Science and Technology of China), Tianyi Zhang (Zhejiang University; Shanghai Artificial Intelligence Laboratory), Haoran Hao (Carnegie Mellon University), Weiyun Wang (Fudan University), Yibin Liu (Northeastern University), Dehui Wang (Shanghai Jiaotong University), Guanzhou Chen (Shanghai AI Laboratory, Shanghai Jiaotong University), Zijian Cai (Shenzhen University), Junting Chen (national university of singaore, National University of Singapore), Weijie Su (University of Science and Technology of China), Wengang Zhou (University of Science and Technology of China), Yu Qiao (Shanghai Aritifcal Intelligence Laboratory), Jifeng Dai (Tsinghua University, Tsinghua University), Jiangmiao Pang (Shanghai AI Laboratory), Gen Luo (Shanghai AI Laboratory), Wenhai Wang (Shanghai AI Laboratory), Yao Mu (Shanghai Jiao Tong University), Zhi Hou (Shanghai Artificial Intelligence Laboratory)

DataMIL: Selecting Data for Robot Imitation Learning with Datamodels

Authors: Shivin Dass (University of Texas at Austin), Alaa Khaddaj (OpenAI), Logan Engstrom (Massachusetts Institute of Technology), Aleksander Madry (OpenAI), Andrew Ilyas (Carnegie Mellon University), Roberto Martín-Martín (University of Texas at Austin)

TimeSeriesExamAgent: Creating Time Series Reasoning Benchmarks at Scale

Authors: Malgorzata Gwiazda (Technical University of Munich), Yifu Cai (Millennium Management LLC), Mononito Goswami (Carnegie Mellon University), Arjun Choudhry (Georgia Institute of Technology), Artur Dubrawski (Carnegie-Mellon University)