News

Researchers propose RIMS, a three-stage preference optimization framework for small-scale language models in retrieval-augmented generation, which uses synthetic chain-of-thought preference data and a differentiable soft…

Researchers introduced PPO-HSC (Proximal Policy Optimization with High-order Sampling Coverage), an exploratory reinforcement learning framework that addresses mode collapse in Large Language Model fine-tuning by incenti…

Researchers propose JUMP (Joint Uncertainty-Guided Mask Probing), a single-pass membership inference attack on fine-tuned discrete diffusion language models that improves mean ROC-AUC from 0.82 to 0.90 over the prior SAM…

Researchers propose 3D FaceShell, a framework that adds subtle, learnable perturbations to 3D face avatars to mislead vision-language models (VLMs) from inferring sensitive attributes, while preserving geometric fidelity…

Researchers introduced ARGO, a fully-sensorized smart eyewear platform that uses the STM32N6 microcontroller's integrated Neural Processing Unit (NPU) to run on-device machine learning, achieving 10 FPS and ~113 minutes …

A new study on arXiv (2607.16231v1) evaluates OrthoGrad, a geometric intervention that removes the radial component of weight gradients during optimizer updates, in noisy-label image classification. On MNIST with small-d…

The NOWJ team presents adaptive pipelines for all five tasks of the COLIEE 2026 competition, including legal case retrieval, entailment, statute law retrieval, textual entailment, and judgment prediction. Their approach …

A new arXiv preprint introduces ColGraphRAG, a late-interaction evidence retrieval method for multimodal GraphRAG that replaces single-vector bi-encoder similarity with MaxSim-style multi-vector scoring for graph-linked …

Researchers from arXiv identify a structured confound in Reinforcement Learning from Human Feedback (RLHF) where pairwise preference labels may reflect the rater's state during annotation rather than just output quality,…

A new survey from arXiv examines LLM unlearning as a cyber defense mechanism, finding that gradient-based methods dominate the field but a central question remains unresolved: whether current methods genuinely remove kno…

Researchers have introduced Shapley Context Pruning (SCP), a novel framework for context reranking in Retrieval-Augmented Generation (RAG) systems that applies cooperative game theory to attribute importance at the sente…

Researchers introduced GenSyn10, a CIFAR-10-aligned synthetic image dataset of 60,000 images generated using three architecturally diverse models (FLUX.2-dev, HunyuanImage-3.0, and Qwen-Image-2512) to benchmark AI-genera…

A new calibration method for mixed-precision tolerance in tensor kernels, tested on the 26-entry gpuemu corpus with 8,076 result rows, achieves up to 2,184× tighter absolute tolerances than hand-picked values. For seven …

Researchers have developed a token-level cross-modal transformer with contrastive multi-task learning that jointly performs breast cancer subtype classification and survival prediction by enabling fine-grained interactio…

A new survey from arXiv provides a comprehensive review of Graph Neural Network (GNN)-based link prediction, proposing a taxonomy that categorizes advances by technique (GCN, GAE, GAT, GFormer) and application (knowledge…

A new study from arXiv reveals that multi-agent LLM systems are vulnerable to planning-phase prompt injection attacks, with GPT-5 achieving the highest attack success rate (ASR = 0.68) across 3,479 episodes, contradictin…

A new framework called Generative Ontology Induction (GOI) achieves 95-100% structural coverage across four diverse ontologies, according to a preprint on arXiv. The domain-agnostic method induces a typed graph with six …

A new study from arXiv evaluates nine open-weight small language models between 135 million and 3 billion parameters on a 1,085-example, 16-topic benchmark, finding that Qwen Coder 3B achieves the highest base accuracy a…

Researchers introduce masked diffusion language models (MDLMs) as steerable text-based world models for reinforcement learning, achieving up to 47% absolute gains over baselines in zero-shot transfer across three out-of-…

Researchers propose BACON, a four-stage pipeline that combines budgeted human calibration with multiple AI-judge outputs to produce more accurate annotations, treating AI judges as auxiliary measurements rather than grou…