Designing system algorithms remains challenging, where the discontinuous nature of the solution space often forces system engineers to rely on generic heuristics at the expense of performance. We study whether LLMs can practically drive algorithm generation, and find that they are biased towards well-known generic…
With the rapid advancement of AI-generated videos, there is an urgent need for effective detection tools to mitigate societal risks such as misinformation and reputational harm. In addition to accurate classification, it is essential that detection models provide interpretable explanations to ensure transparency for regulators…
Microsoft is proud to have over 150 accepted papers and be a sponsor of the 14th International Conference on Learning Representations (ICLR) (opens in new tab) occurring April 23 - 27, 2026 in Rio de Janeiro, Brazil. ICLR is the premier gathering of professionals dedicated…
Discrete diffusion models are a powerful class of generative models that demonstrate strong performance across many domains. However, for efficiency, discrete diffusion typically parameterizes the generative (reverse) process with factorized distributions, which makes it difficult for the model to learn a target process in a…
Masked diffusion models (MDMs) offer a compelling alternative to autoregressive models (ARMs) for discrete text generation because they enable parallel token sampling, rather than sequential, left-to-right generation. This means potentially much faster inference. However, effective parallel sampling faces two competing requirements: (i) simultaneously updated tokens…
Text prompt is the most common way for human-generative AI (GenAI) communication. Though convenient, it is challenging to convey fine-grained and referential intent. One promising solution is to combine text prompts with precise GUI interactions, like brushing and clicking. However, there lacks a formal model…
The rapid adoption of AI agents across domains has made systematic evaluation crucial for ensuring their usefulness and successful production deployment. Evaluation of AI agents typically involves using a fixed set of benchmarks and computing multiple evaluation metrics for the agent. While sufficient for simple…
While Cave Automatic Virtual Environment (CAVE) systems have long enabled room-scale virtual reality and various kinds of interactivity, their content has largely remained predetermined. We present Storycaster, a generative AI CAVE system that transforms physical rooms into responsive storytelling environments. Unlike headset-based VR, Storycaster preserves…
Ineffective meetings are pervasive. Thinking ahead explicitly about meeting goals may improve effectiveness, but current collaboration platforms lack integrated support. We tested a lightweight goal-reflection intervention in a preregistered field experiment in a global technology company (361 employees, 7196 meetings). Over two weeks, workers in…
Large Language Model agents are evolving from question-answering tools to genuine collaborators that run continuously, reason across turns, reflect on tool usage, and act on our behalf. Yet effective human-agent collaboration requires more than raw capability. It demands social intelligence: the ability to sense, understand,…
Generative AI has greatly transformed creative work in various domains, such as screenwriting. To understand this transformation, prior research often focused on capturing a snapshot of human-AI co-creation practice at a specific moment, with less attention to how humans mobilize, regulate, and reflect to form…
Although generative AI is being deployed into classrooms with promises of aiding teachers, educators caution that these tools can have unintended pedagogical repercussions, including cultural misrepresentation and bias. These concerns are heightened in low-resource language and Indigenous education settings, where AI systems frequently underperform. We…
