CMU researchers are presenting 194 papers on the Fourteenth Worldwide Convention on Studying Representations (ICLR 2026), held from April Twenty third-April twenty seventh on the Riocentro Conference and Occasion Heart in Rio de Janeiro, Brazil. Here’s a fast overview of the areas our researchers are engaged on:<\/p>\n Listed below are our most frequent collaborator establishments:<\/p>\n EditBench: Evaluating LLM Talents to Carry out Actual-World Instructed Code Edits<\/a><\/p>\n This work introduces EditBench, a brand new benchmark for testing how properly AI fashions can edit present code based mostly on consumer directions. In contrast to prior benchmarks, it makes use of real-world coding duties and contexts, together with issues like the encircling code and cursor place. The benchmark consists of 545 numerous issues, and outcomes present that almost all fashions battle\u2014just a few obtain robust efficiency. The research additionally finds that having extra real looking context considerably impacts how properly fashions carry out, highlighting the significance of evaluating code-editing in real-world settings.\n<\/p>\n<\/p><\/div>\n UALM: Unified Audio Language Mannequin for Understanding, Era and Reasoning<\/a><\/p>\n This paper introduces the Unified Audio Language Mannequin (UALM), a single mannequin designed to deal with audio understanding, text-to-audio era, and multimodal reasoning collectively. As a substitute of treating these as separate duties, UALM learns to each interpret and generate audio, reaching efficiency akin to specialised state-of-the-art fashions. The authors additionally present that combining textual content and audio in the course of the mannequin\u2019s reasoning course of improves its potential to deal with advanced duties. Total, the work demonstrates a step towards extra normal AI techniques that may purpose throughout each language and sound.\n<\/p>\n<\/p><\/div>\n Agent Knowledge Protocol: Unifying Datasets for Numerous, Efficient Positive-tuning of LLM Brokers<\/a><\/p>\n This work introduces the Agent Knowledge Protocol (ADP), a standardized format for representing coaching information for AI brokers. The authors argue that the principle problem isn\u2019t an absence of information, however that present datasets are fragmented throughout completely different codecs and instruments. ADP acts as a standard \u201cinterlingua,\u201d making it simpler to mix numerous information sources\u2014like coding, searching, and power use\u2014right into a single coaching pipeline. By changing 13 datasets into this unified format, the authors present that fashions skilled on the mixed information obtain improved efficiency.\n<\/p>\n<\/p><\/div>\n MotionStream: Actual-Time Video Era with Interactive Movement Controls<\/a><\/p>\n This paper introduces MotionStream, a system for producing movies in actual time based mostly on movement and textual content inputs. In contrast to prior strategies that take minutes to supply a video, MotionStream can stream outcomes at as much as 29 frames per second on a single GPU. The important thing concept is to coach a quick, causal mannequin that may generate video constantly, utilizing methods that forestall high quality from degrading over lengthy sequences. Because of this, customers can interactively management movement\u2014like drawing paths or transferring a digital camera\u2014and see the video replace immediately.\n<\/p>\n<\/p><\/div>\n OpenThoughts: Knowledge Recipes for Reasoning Fashions<\/a><\/p>\n This work introduces the OpenThoughts challenge, which goals to create high-quality, open-source datasets for coaching reasoning-focused AI fashions. The authors present that fashions skilled on their public information can match or exceed the efficiency of robust present techniques that depend on non-public datasets. By rigorously finding out and bettering their information era course of, they construct bigger and higher datasets that considerably increase efficiency throughout math, coding, and science benchmarks. Total, the challenge demonstrates that open information alone may be sufficient to coach extremely succesful reasoning fashions.\n<\/p>\n<\/p><\/div>\n Mamba-3: Improved Sequence Modeling utilizing State House Rules<\/a><\/p>\n This paper introduces Mamba-3, a brand new mannequin designed to make AI inference quicker and extra environment friendly with out sacrificing efficiency. Whereas many environment friendly options to Transformers cut back computation, they typically battle with duties like monitoring long-term info; Mamba-3 addresses this with improved state modeling and a extra expressive replace mechanism. The mannequin additionally makes use of a multi-input, multi-output design to spice up accuracy with out slowing down era. Total, Mamba-3 reveals that it\u2019s attainable to enhance each effectivity and functionality on the similar time, pushing ahead the tradeoff between velocity and efficiency.\n<\/p>\n<\/p><\/div>\n Overcoming Joint Intractability with Lossless Hierarchical Speculative Decoding<\/a><\/p>\n This paper introduces Hierarchical Speculative Decoding (HSD), a brand new methodology to hurry up massive language mannequin inference by bettering the verification step in speculative decoding whereas preserving actual output distributions. It addresses the problem of \u201cjoint intractability\u201d in sequence-level verification by organizing resampling right into a hierarchy that redistributes chance mass throughout branches, enabling extra tokens to be accepted directly. The strategy is theoretically confirmed to be lossless and empirically reveals constant velocity enhancements throughout fashions and benchmarks, outperforming prior tokenwise and blockwise verification strategies. Total, HSD presents a sensible and normal strategy to speed up decoding with out sacrificing constancy, reaching state-of-the-art effectivity when built-in into present frameworks.\n<\/p>\n<\/p><\/div>\n Distributional Equivalence in Linear Non-Gaussian Latent-Variable Cyclic Causal Fashions: Characterization and Studying<\/a><\/p>\n \nThis paper research causal discovery in linear non-Gaussian fashions with latent variables and cycles, specializing in when completely different causal graphs are observationally indistinguishable. It supplies the primary normal characterization of distributional equivalence on this setting, introducing new instruments\u2014particularly edge rank constraints\u2014to explain when two fashions generate the identical noticed information. Constructing on this idea, the authors derive sensible graphical standards and transformations to enumerate all equal fashions and suggest an algorithm to get better your entire equivalence class from information. Total, the work removes the necessity for robust structural assumptions and presents a normal, principled framework for latent-variable causal discovery.\n<\/p>\n<\/p><\/div>\n Revela: Dense Retriever Studying through Language Modeling<\/a><\/p>\n \nThis paper introduces Revela, a self-supervised framework for coaching dense retrievers by leveraging language modeling aims as a substitute of counting on annotated query-document pairs. It augments next-token prediction with an in-batch consideration mechanism that enables paperwork to attend to one another, enabling the retriever to study cross-document relationships collectively with a language mannequin. Experiments throughout domain-specific, reasoning-intensive, and normal benchmarks present that Revela matches or surpasses supervised and API-based retrievers whereas utilizing considerably much less information and compute. Total, the work demonstrates a scalable and environment friendly different for retriever studying instantly from uncooked textual content with robust generalization throughout domains.\n<\/p>\n<\/p><\/div>\n![\"\"](\"https:\/\/blog.ml.cmu.edu\/wp-content\/uploads\/2026\/04\/image-7.png\")
<\/figure>\n![\"\"](\"https:\/\/blog.ml.cmu.edu\/wp-content\/uploads\/2026\/04\/image-8.png\")
<\/figure>\n
\n <\/p>\nOral Papers<\/h2>\n