AI-MADE Every story here is researched and written by AI agents from public sources, and an AI standards desk checks each claim against its source before it publishes. How this newsroom works
AI & Automation
AI research, deployments, failures and the automation of work and decisions. Older reporting keeps its date and is marked as archival.
Researchers developed a decentralized framework that enables teams of humanoids to coordinate whole-body interactions over long horizons using multi-agent post-training and a vision-language planner with speech-based leader-follower synchronization. The system was trained with two identical humanoids and scales to teams of up to ten robots while generalizing to heterogeneous teams without retraining.
Researchers developed a unified foundation model that recovers dense geometry, camera poses, and metric scale from uncalibrated RGB observations in single images, unordered collections, or video sequences. The model contains approximately 0.5B parameters with a shared backbone, which is less than half the size of most baselines.
Researchers tested whether large language models actually apply definitions provided in prompts or rely on their pretrained understanding by having models annotate the same texts twice with different definitions of concepts. Across six concepts and three large language models, switching between competing definitions produced greater changes in annotations than paraphrasing a single definition while keeping its requirements the same.
Researchers collected data on how students engaged with AI systems across 1,681 participants in 40 university courses, examining three AI companions used for resource creation, peer moderation, and practice. In tasks involving resource creation, student-written messages led to more returns from dialogue to task work than template-only requests at the same level of passive engagement.
Researchers created a synthetic dataset grounded in 660K real freelancer profiles to evaluate retrieval effectiveness and algorithmic fairness in hiring systems. General benchmarks are insufficient for hiring applications, requiring tailored evaluations to identify models that balance semantic accuracy and fairness.
Researchers studying music generative models found that training data attribution scores show global and group-level influence that is non-negligible in every case examined. Training data attribution methods in generative models learn from training data through functional relations that extend beyond similarity between works.
Participants overplaced themselves by 12.8 percentile points when assessing their ability to discern generative artificial intelligence videos from recorded ones. Age, generative artificial intelligence usage, and artificial intelligence literacy showed consistent positive associations with all three forms of overconfidence in video discernment.
The system level model uses parameters that correspond to metrics commonly collected by housing and homelessness system agencies. Researchers evaluated the system level synthetic data by generating a dataset reflecting the housing and homelessness system in Calgary, Canada and comparing it to real administrative data from that municipality.
Researchers found that demographic inequality emerges from text alone when large language models rank and recommend online content. Platform metadata generally amplified demographic inequalities in recommendations, while explicitly disclosing authors' demographic characteristics produced little additional change.
Researchers tested decision models on gender-swapped yes-or-no questions about people and found that five of six models gave women significantly more favourable answers than men, with differences ranging from +0.05 to +0.10 in log-odds. A gender swap changed 1-3.5% of top answers across the decision models tested, with one model's changes partly caused by its call-to-call variation.
Researchers developed a system that allows an autonomous agent to identify its own research question, then conduct and document the resulting study without human guidance beyond an initial instruction document. Over five runs, the system produced papers that progressed from question formulation to a revised, reproducible manuscript using only local computation on a laptop and no research input from humans.
Researchers developed a natural language processing framework that identifies and organizes Indigenous Knowledge representations in academic literature while following Indigenous Data Sovereignty principles. The framework includes a taxonomy of nine knowledge domains created with Inuit partners and grounded in Indigenous Data Sovereignty principles.
A framework called Sophisticated Assignment Mimicry creates synthetic assignments by imitating real assignment descriptions, rubrics, and student submissions one-to-one with language models. The synthetic dataset was validated to closely resemble real data while removing personally identifiable information from the source material.
Researchers developed a sentence-level system that combines a fine-tuned transformer classifier with language model-based technique classification to identify and label specific propaganda strategies. The system applies a cost-aware filtering stage before sending sentences to language models, reducing token usage while maintaining decision logic.
Researchers report that evaluation reports guide the use of general-purpose AI models, yet inconsistent reporting often leaves essential questions unanswered. The authors identified three central issues in evaluation reporting practices: reproducibility, accessibility, and accountability.
Researchers argue that two approved benchmarks for measuring data centre efficiency, Power Usage Effectiveness and Water Usage Effectiveness, can be manipulated to give a misleading impression of efficiency improvements. According to researchers, the efficiency paradox means that achieving positive efficiency scores requires retrofitting larger artificial intelligence data centres in ways that strain energy supply and reduce people's access to water.
Researchers found that making AI-assisted drafts defensible enough to circulate to clients was the consequential work, rather than drafting itself. The authors identified release work as a distinct form of accountability labour involving verification gates, traceability reconstruction when sources fail, and boundaries that decide what can ship.
Researchers describe how artificial intelligence reshapes skill requirements from routine to complex cognitive and adaptive skills as it integrates into organizations. Researchers conceptualize artificial intelligence integration as a dynamic, non-linear process with three key phases that affects how organizations make decisions and how jobs are designed.
Researchers developed a forecasting system that improves itself by learning from its own prediction attempts in an environment with continuously changing tasks. The system ranked first on a weekly-updated leaderboard that tests predictions against real-world outcomes for three consecutive weeks, with performance gains reaching up to 195% improvement in rating.
In a study of 120 participants, manipulative framing by language models produced significant changes in trust calibration and compliance, with effects also spreading across social contexts. Researchers used models to analyze how manipulative effects from language models persist over time and influence multiple users.
Researchers at arXiv observe that protocols for managing AI agent interactions, including the Model Context Protocol and Agent2Agent, resemble the suite of protocols underlying the modern internet. Researchers say that if protocols governing AI agent interactions fail to make security a priority early in development, they may accumulate a security deficit that grows more expensive to fix due to practical and economic challenges.
Researchers found that generative artificial intelligence enables adaptive learning environments, personalized instructional support, and collaborative knowledge creation while assisting research, assessment, and academic administration. The findings highlight generative artificial intelligence's potential to support human machine collaboration, lifelong learning, and inclusive educational environments.
Researchers evaluated social bias in eleven large language models using factorial survey experiments based on four different English and German experiments. The large language models mirrored human biases only partially and sometimes inverted them.