AI Research

A new LSTM Universal Transformer (LSTM-UT) with bounded gated memory outperformed both the Block Universal Transformer (BUT) and CoTFormer on Rule 30 cellular automata and a delayed-recall task, according to an arXiv pap…

Researchers proposed Compressed Active Subspaces (CAS), a method that maps model parameters into a compressed space via a structured isometric embedding before constructing the active subspace, cutting the memory needed …

A systematic mapping of 14,767 arXiv papers introducing or updating LLM evaluation resources between January 2022 and August 2026 finds growing emphasis on action, interaction, and professional applications, with LLM-bas…

A new arXiv study (2609.19244v1) analyzed Web search across four conversational platforms — ChatGPT, Claude, Grok, and DeepSeek — combining real-world user interactions with controlled API experiments. The study found th…

A position paper posted to arXiv (2609.19203v1) argues that compound agentic AI systems need a Foundation Model Operating System (FMOS), a system layer that virtualizes foundation model interactions the way virtual machi…

Researchers introduced MAGS, a unified multi-agent framework that generates executable programs with formal safety guarantees by using Dafny as a verification-aware intermediate representation, according to an arXiv pape…

A new arXiv paper (2609.19387v1) introduces AutoTuring, a benchmark that gives the same AI agent the same 15-dimensional accelerator design space twice — once as named architectural knobs with simulator counters and once…

A new arXiv paper (2609.19515v1) introduces LLM-as-an-Improver and the Verify--Repair--Reselect (VRR) method, which uses verification feedback to generate and reselect improved candidates rather than only ranking a fixed…

A new arXiv paper (2609.19465v1) from unnamed researchers proposes a dependency-graph framework that formalizes compositional reasoning in language models into three levels of increasing complexity, and reports a consist…

Researchers introduced QVAC Genesis III, a 191.43B-token STEM-focused synthetic corpus spanning 19 domains, built via a dual generation strategy that uses a weak edge-scale student model's failures as corrective explanat…

A paper on arXiv (2609.19519v1) proposes a three-part hierarchical architecture for long-horizon language-model agents — time-scale-indexed levels, a clocked tick as the unit of autonomous action, and cascaded intelligen…

A new arXiv paper (2609.19526v1) introduces Recursive Self Improvement via Fast Tree-search (SIFT), a sample-efficient framework that lets coding agents recursively modify their own implementations under strict budget co…

A new arXiv paper (2609.19524v1) proposes a unified evaluation framework for trustworthy AI that assesses large language models, agentic systems, and multimodal models across eight trustworthiness dimensions: capability,…

A new arXiv paper (2609.19538v1) proposes a hierarchical hybrid large language model (LLM) and multi-agent reinforcement learning (MARL) architecture for low-altitude wireless networks (LAWNs) supporting heterogeneous un…

The Modality Discrepancy Transformer (MDT), a conflict-aware multimodal fusion model built on the framework of Bekhouche et al., achieved 0.7408 Macro F1 on the labelled test split and 0.7368 on the private leaderboard o…

Researchers introduced FakeSpotter, a content- and strategy-agnostic tool that estimates viral misinformation risk by measuring structural fingerprints of misinformation rather than adjudicating truthfulness, using repea…

A training-free method that repeatedly samples completions of a single prompt at elevated temperature and applies Principal Component Analysis to pooled hidden activations can discover prompt-conditional stylistic axes i…

Researchers constructed UC-Bench, a human-annotated benchmark for detecting user-side implicit conflicts in Human-LLM dialogue, and found existing LLMs struggle with the task, especially when conflicts stem from implicit…

A new benchmark called Neo-Classic, introduced in arXiv paper 2609.19154v1, shows that state-of-the-art large language models lose 20 to 50 percent accuracy when moving from historical Classical Chinese Poetry to strictl…

VisKG-LM, a new method detailed in arXiv paper 2609.19158v1, compiles retrieved knowledge graphs once offline into cached images that a language model reads as read-only visual memory, decoupling graph encoding from lang…