Researchers from Southeast University introduced a training-free paradigm for attributing AI-generated videos to their source generative models, achieving a Rank-1 accuracy of 20.5% and mean Average Precision of 16.6% on…
A new arXiv preprint (2607.28636v1) introduces Chain-of-Models (CoM), an automated audit pipeline in which a second model inspects a first model's reasoning trace to reduce cognitive biases in LLM judges. Testing 9 model…
Researchers introduced LegoQ, a classical density-matrix representation learning framework for hyperspectral image classification that maps spectral bands to positive semi-definite Hermitian matrices and uses spectral, s…
A team from ZeroR achieved 2nd place on hate speech detection (F1: 0.797) and 4th place on sentiment analysis (F1: 0.518) in the CHiPSAL 2026 shared task on multimodal hate speech and sentiment detection in Nepali memes,…
Researchers propose Stateful Knowledge Learning (SKL), a method that trains large language model agents to maintain explicit, state-anchored predictive knowledge instead of relying on trajectory-level reflection, and rep…
Researchers from Southeast University propose RAID, a new AI-generated image detection method that uses bit-reversed images and a gradient-based patch selection with a convolutional classifier, achieving state-of-the-art…
Researchers introduced TokenSwap, a method that replaces textual concepts with semantically aligned images to create image-interleaved inputs, and found that across 42 multimodal large language models, performance droppe…
A new study from arXiv (2607.28641v1) introduces the Agentic Formalism Trap and the Evaluative Dissonance Index (D_E), showing that LLM-as-a-Judge systems conflate structural proceduralism with semantic truth under adver…
A new study from arXiv (arXiv:2607.28658v1) finds that downstream fine-tuning on GLUE does not reliably preserve the reference ranking of federated pre-trained models, whereas direct next-token prediction on GLUE text st…
A new inference-time framework, Adjudicated Captioning, achieves CIDEr 117.6 and SPICE 21.9 on the COCO Karpathy test split for zero-shot image captioning, up from 108.0 and 20.3 for the baseline IFCap, a +9.6 CIDEr gain…
A new benchmark called FinIndices, introduced by researchers in a preprint on arXiv (2607.28661v1), tests large language models on financial reasoning over long-horizon statements up to 32K tokens and finds severe vulner…
A new arXiv paper (arXiv:2607.28666v1) finds that only 32 of 72 AI configurations (56.1%) in regulated financial services meet a production bar requiring sustained accuracy, reproducibility, verifiable attribution, and i…
Researchers introduced TELLER, a dual-path iterative preference optimization method for table entity linking, which improves accuracy on the TableInstruct entity-linking subset from 94.35% to 94.50% via a direct-answer p…
Researchers introduced Self-Supervised Skill Optimization (SSO), a framework that optimizes agent skills for frozen large language models using only unlabeled task instances, eliminating the need for ground-truth labels,…
A new arXiv preprint (2607.28801v1) introduces a dataset-centric meta-evaluation framework that audits LLM benchmark datasets at the sample level across five dimensions, revealing internal heterogeneity in MMLU, ARC, Win…
A study from arXiv:2607.28814v1 found that preference-optimizing large language models (LLMs) for motivational interviewing by penalizing confrontation reliably lowers goal persistence across all tested models, while att…
A new arXiv paper (2607.28840v1) argues that large language models in financial applications should not be approved for production based on benchmark scores alone, requiring system-level validation across data, model des…
Researchers introduced TextCloak, an RL-driven framework that protects textual data from unauthorized Large Language Model (LLM) exploitation by generating unlearnable examples that degrade model performance when fine-tu…
A new arXiv paper (arXiv:2607.28629v1) proposes a layered architecture for Agentic AI and validates it using OpenClaw and Ollama as a full-stack system, finding that capabilities like persistent memory and adaptive decis…
A benchmarking study of autonomous AI research systems found that FARS benchmark papers significantly outperform four leading AI Scientist frameworks, achieving mean scores of 2.14–2.47 on a 1–5 scale compared to 1.00–1.…