Researchers affiliated with the Vector Institute had 48 papers accepted at the 2026 International Conference on Learning Representations (ICLR), held April 23–27 in Rio de Janeiro, Brazil, according to the institute.
Vector describes ICLR as the leading venue for representation learning and deep learning research. The 48 accepted papers came from 60 Vector Faculty Members, Faculty Affiliates, and Distinguished Postdoctoral Fellows, and included 11 collaborative papers. Vector researchers also led 6 of the 40 accepted workshops, which the institute says represents 15 percent of the conference’s workshop program.
The accepted papers span several research areas, as described by Vector: reinforcement learning for advanced reasoning, generative AI and multimodal systems, autonomous AI agents, trustworthy and responsible AI, and scientific and healthcare applications.
Among the papers highlighted by Vector is work by Zongliang Ji, Yifei Sun, Andre Amaral, Anna Goldenberg, and Rahul G. Krishnan on generating portable clinical representations using large language models. The paper examines whether frozen LLMs can produce patient embeddings from ICU time-series data that allow a downstream predictor trained at one hospital to transfer to another with minimal retraining. The abstract states that the approach was tested across three cohorts — MIMIC-IV, HIRID, and PPICU — on multiple clinical forecasting and classification tasks, and was “surprisingly competitive with in-distribution” baselines while showing smaller performance drops when transferring to new hospitals. The authors also report that using LLM-generated representations does not increase demographic recoverability of age or sex relative to baselines, suggesting, in their assessment, “little additional privacy risk.”
A second highlighted paper, ChronoEdit, addresses temporal reasoning in image editing. The framework, from a team including Vector Faculty Member Sanja Fidler, reframes image editing as a video generation problem, treating input and edited images as the first and last frames of a video and using pretrained video generative models to constrain the editing to physically plausible transformations. The team introduces PBench-Edit, a new benchmark of image-prompt pairs requiring physical consistency.
A third paper, Computer Agent Arena, proposes an open-source platform for evaluating computer-use agents through head-to-head comparison, with Vector Faculty Affiliate Victor Zhong among the authors. The system uses cloud-hosted environments and collects human preference votes. The paper reports results from 2,201 high-quality votes across 12 agents, finding ranking reversals relative to static benchmarks and identifying agent-human interaction and self-correction as factors that boost user preference beyond task completion alone.
The conference proceedings include the complete list of 48 accepted papers from Vector-affiliated researchers. ICLR 2026 runs through April 27 in Rio de Janeiro.