AI Research

Researchers introduced Physics-Audited Agentic Scientific Machine Learning (PA-SciML), a verification-first workflow that ensures surrogate models discovered by LLM agents satisfy physics constraints such as boundary con…

Researchers introduced STST-JEPA, a self-supervised transformer for EEG analysis, pretrained on 47,703 sessions from ages 5-81. The model achieved a mean absolute error of 3.06 years in age regression and ranked first on…

Researchers introduced SPEAR, a Python library that programmatically controls Unreal Engine applications for photorealistic embodied AI research. It offers over 14,000 programmable functions, renders 1920x1080 images at …

A new benchmark called ImagingBench reveals that leading vision-language models and agentic AI systems, including Gemini, GPT, and Qwen, perform poorly on computational imaging tasks involving physics and inverse problem…

Researchers extended PubHealthBench, a public health QA benchmark, into a retrieval-augmented setting and found that hybrid retrieval consistently improves recall and ranking quality. Providing retrieved context substant…

Researchers introduced a multi-factor scoring system to evaluate large language model responses, assessing accuracy, conciseness, factual consistency, readability, and coherence. Tests on the TruthfulQA dataset showed LL…

Researchers at an undisclosed institution found that multi-teacher on-policy distillation for agentic language models can cause over-calling of tools, where models call tools on examples that should be answered directly.…

Researchers propose an evaluation framework combining regularized regression on item text embeddings with repeated cross-validated R-squared and two performance upper bounds—a reliability ceiling and a design ceiling—to …

A new theoretical analysis of in-context search in large language models shows that when self-reflection reliably localizes early mistakes, sequential reasoning can yield exponential improvements over the base model, sol…

Researchers introduced reasoning consistency scanning, a framework for auditing chain-of-thought validity in AI safety evaluations. The method detects logical inconsistencies between a model's stated reasoning and its fi…

Researchers studying harmful online communication in Discord cybercrime communities found that interpretation difficulty stems from slang, coded terms, and community-specific expressions. Their experiments showed that lo…

Researchers evaluated the relevance of several RAG metrics by comparing scores from four libraries (Ragas, DeepEval, RAGChecker, Opik) against human evaluators and standard metrics using a question-answering dataset. The…

Researchers introduced TF-Engram, a train-free memory system that stores phrase-specific semantic knowledge offline across GPU, DRAM, and SSD tiers, and uses predictive prefetching to reduce latency. In tests with Qwen3-…

Researchers introduced HALE, a hybrid agent-based and language-driven epidemic modeling framework that uses large language models to predict human decision-making in simulations, demonstrated by modeling COVID-19 in Salt…

Researchers propose DeLS-Spec, a speculative decoding method that decouples long- and short-context experts to accelerate LLM inference. By adding a lightweight local head to the existing DFlash model, DeLS-Spec achieves…

Researchers introduced SynthAVE, a large-scale benchmark for synthetic labeling of e-commerce attribute extraction, covering 12,726 products across 229 product types, 792 attributes, and four languages. They validated th…

Researchers at an undisclosed institution found that in hierarchical search agents using large language models, scaling the delegation role improves exact match by ~11 points while scaling execution only improves by ~2.6…

Researchers introduced AgentLens, a benchmark for evaluating interactive code agents that assesses entire trajectories—including instruction following, tool use, error recovery, and communication—rather than just binary …

Researchers proposed a ReAct-style agentic setup combining LLM reasoning with SageMath for solving research-level mathematical problems, achieving an average performance gain of 9.7 percentage points across models. GPT-5…

Researchers are developing Agentic Data Environments, an execution substrate that amplifies autonomous agent capabilities while enforcing safety guarantees, reframing data systems from passive stores into active substrat…