Researchers have developed POLAR, a multimodal memory-augmented framework that enables embodied AI agents to personalize assistance over long-term user interactions by organizing prior experiences into a knowledge graph.…
Researchers introduced AgingBench, a longitudinal benchmark designed to measure how AI agents degrade over time after deployment. The study found that agent reliability declines through four distinct aging mechanisms—com…
Researchers introduced Anchor, a task-generation pipeline that prevents artifact drift in AI agent benchmarks by formalizing business workflow specifications into constraint optimization programs. The system jointly prod…
Researchers have introduced OmniToM, a benchmark that evaluates large language models' theory of mind by requiring explicit modeling of belief structures for all actors in a narrative. The benchmark, built from 895 stori…
Researchers introduced JobBench, a new benchmark that evaluates AI agents on 130 tasks across 35 occupations based on workflows experts prioritize for delegation. The benchmark assesses models using fact-anchored rubrics…
A new autonomous research system called ScientistOne, developed with a Chain-of-Evidence framework, achieved zero fabricated citations and perfect score verification across 75 papers, while baselines exhibited hallucinat…
Researchers have developed CARL (Contrastive Action-based Representations for Reusable Local Control), a new algorithm that exploits local dynamics regularity to learn reusable skills for hierarchical reinforcement learn…
Researchers introduced MM-CreativityBench, a new benchmark designed to test whether large multimodal models can solve open-ended, physically grounded problems by repurposing objects in non-obvious ways. The study found t…
Large language models (LLMs) show reduced accuracy on math problems when simple details like names or numbers are changed, and a new study finds that using code execution methods does not improve this robustness. Researc…
MiniMax released the M2 series of Mixture-of-Experts language models, featuring 229.9 billion total parameters with only 9.8 billion activated per token, designed for agentic deployment. The models use agent-driven data …
Researchers introduced PolyFusionAgent, an interactive framework that combines a multimodal polymer foundation model with a literature-grounded design agent to predict polymer properties and generate novel polymer struct…
Researchers have developed Self-Verified Distillation, a method allowing large language models to improve themselves using only unlabeled prompts without external feedback. The approach, tested on Qwen3 models across mat…
A new survey from arXiv researchers has unified the study of Pretraining Data Exposure (PDE) in large language models, revealing growing risks to evaluation integrity and data privacy as training datasets expand. The pap…
Researchers introduced SPEAR (Sandboxed Prompt Engineer with Active Roll-back), a code-augmented agentic optimizer that autonomously rewrites prompts using a Python sandbox for structural error analysis, outperforming ex…
A new study published on arXiv describes "The Daily Dose," an LLM-driven system that automates clinical summarization and clinical-trial identification for radiation oncology. After one month of deployment, a survey of 5…
Researchers have developed RICE-PO, a policy optimization framework that converts retrieval interactions into localized learning signals for training reasoning-based retrieval agents. The framework addresses the credit-a…
Researchers have demonstrated that retrieval-augmented generation (RAG) can be understood as an in-context optimization process, where a single linear self-attention layer performs a gradient-descent step on a unified RA…
A new study reveals that large language models hallucinate on structured knowledge tasks due to systematic internal failures, not random errors. Researchers found that attention mechanisms disproportionately focus on sho…
Researchers have developed a framework for evaluating and adjusting the cultural values of large language models by analyzing their responses to scenario-based dilemmas rather than direct questioning. The method, which u…
Researchers introduced EnterpriseMem-Bench, a multi-turn Text-to-SQL benchmark of 300 sessions and 1,400 queries built from three enterprise domains, to evaluate how five frontier AI models handle memory across conversat…