Realtime AI News
A new arXiv paper introduces the TRUSTMEM framework to address error accumulation and hallucination persistence in LLM agent long-term memory caused by generated write, revise, and delete operations.
Researchers propose the Agentic BKT pipeline, a multi-agent LLM architecture that stealthily assesses financial competencies from open-ended gameplay events without disrupting the learning experience.
Researchers propose a supervised reinforcement learning approach for coordinating distributed energy resources (DERs), achieving more efficient energy management under the uncertainty and complexity that challenge traditional optimization methods.
Researchers propose a novel approach that performs lightweight neural architecture search directly on deployment devices, allowing sensor edge devices to redesign tiny neural networks optimized for real-time data.
Researchers release MacroLens, a multi-task benchmark designed for contextual financial reasoning under macroeconomic scenarios, addressing key challenges like data leakage and reporting lags in time-series evaluation.
A new study shows that dense per-loop cross-entropy loss in looped language models only controls variables exposed by the readout, not all hidden-state variables active in the recurrent transition, creating a systematic supervision blind spot.
A new paper documents how human-AI co-discovery transformed a vague research intuition into concrete sign-embedding quantum algorithms for matrix equations and matrix functions, showing a new paradigm for AI-assisted mathematics.
AgentOdyssey procedurally generates open-ended text games to benchmark agents on exploration, knowledge acquisition, memory retention, and long-horizon planning.
OpenAI publishes a new research paper examining how AI agents are transforming work by handling longer, more complex tasks and expanding productivity across roles.
A new arXiv paper introduces DiARC, a method that improves large language models' performance on the Abstraction and Reasoning Corpus (ARC) by distinguishing positive and negative samples.
AI chipmaker Cerebras saw its stock plummet after its first earnings report since going public, with a narrower gross margin forecast spooking investors.
TechCrunch reports that companies are rushing to stop employees from exhausting AI budgets on low-value small tasks, marking a shift from the 'tokenmaxxing' era to an era of 'token rationing'.
Google Research explores how the reasoning process activates and retrieves parametric knowledge stored within large language models.