Researchers proposed a fine-tuned multi-agent framework that uses large language models to detect OCEAN personality traits from life narratives, with sub-agents adopting different trait perspectives and a judge LLM aggre…
Researchers developed OmniPM-Net, a neural network that fuses discrete station forecasts with gridded chemical transport model outputs to produce accurate PM10 predictions at both monitoring sites and continuous spatial …
Researchers introduced Semidirect Fourier Delta Attention (SFDA), a phase-controlled generalization of Kimi Delta Attention that replaces real diagonal decay with block-rotational Fourier control. The method achieves exa…
Researchers introduced viable path entropy (VPE) as a measure of an AI system's capacity to sustain coherent continuations under repeated reflection. Testing on GSM8K with Qwen2.5-Instruct models showed that increasing t…
Researchers proposed CARE-LoRA, a compressed activation reconstruction framework that reduces memory usage during LoRA fine-tuning of large pre-trained models by exploiting the projection structure of LoRA to replace ful…
Researchers propose BattVAE-GP, a hybrid physics-probabilistic framework that uses a Variational Autoencoder and Gaussian process to model long-horizon battery degradation trajectories with uncertainty quantification, en…
Researchers from an undisclosed institution found that the RG-Flow Transformer, a neural network with a renormalization-group inductive bias, matches the accuracy of a vanilla transformer on sleep staging from scarce EEG…
Researchers introduced a white-box instrument using hidden deterministic finite automata to separately measure reward success and latent-state learning in reinforcement learning agents. They found that high reward does n…
Researchers developed a self-evolving in-context learning framework for pilot-to-beamformer design in MU-MISO systems that adapts to multiple channel models without retraining. The scheme outperforms existing methods inc…
A new survey from arXiv examines in-context reinforcement learning (ICRL) under non-stationary environments, where accumulated context can become stale or misleading as task specifications, transition kernels, or observa…
Researchers argue that the alignment techniques used to improve large language models, such as GPT-2, represent an expression of 'optimization culture' that conflates measurable improvement with value. They trace this co…
Researchers introduced COJEPA, a self-supervised learning framework for volumetric brain MRI that combines joint-embedding predictive architecture with contrastive loss. Trained on 2,286 T1-weighted scans from two cohort…
Researchers introduced a framework for designing agent-ready e-commerce websites that improve AI web agents' performance. In tests with GPT-4.1, Gemini-2.5 Flash, and Grok-4 Fast, agent-ready sites achieved 89.3% strict …
Researchers at arXiv studied convention formation in open-weight language-model populations using a naming-game protocol, finding that graph feedback controls consensus and clique formation. Threshold-similarity routing …
Researchers introduced GenAI Evaluation, a governed pipeline for evaluating retail conversational agents at scale, processing 50,000 records daily and over two million interactions. The system achieved a macro F1 score o…
Researchers introduced Geospatial Foundation Models (GeoFMs), AI/ML models pre-trained on massive geospatial datasets that enable domain experts to fine-tune or prompt them for specific tasks. The paradigm separates comp…
Researchers introduced Cost-Governed RAG, an architecture integrating a codebook-oblivious vector index with a multi-tenant LLM governance gateway to achieve per-tenant cost attribution across retrieval and generation in…
Researchers introduced a Threshold Exceedance Criteria (TEC) framework to evaluate whether frontier language models materially increase non-experts' ability to plan CBRN attacks. In a large-scale empirical study, they fo…
A new study re-evaluating automatic harness evolution for LLM agents finds that the method does not consistently outperform simple test-time scaling and shows limited generalization, raising concerns about overfitting an…
Researchers proposed GenDiff, a diffusion-based framework for low-dose CT reconstruction that jointly models radiation dose and anatomical information, achieving superior image quality and robustness across different dos…