A new arXiv study (2607.20543v1) finds that reinforcement learning with verifiable rewards (RLVR) can improve one-sample accuracy while causing pass@k inversion, where the trained policy solves fewer distinct problems th…
Researchers have discovered that Mixture-of-Experts (MoE) routing in large language models functions as a form of Huffman coding, governed by a new Frequency-Diversity Law. The study, published on arXiv, shows that model…
Researchers have introduced RealVDeblur, an efficient generative framework for real-world video deblurring that uses a physically grounded blur synthesis pipeline from 3D Gaussian Splatting assets and high-frame-rate vid…
Researchers propose Agentic Designer, a multi-agent framework that generates interior furniture layouts by iteratively verifying geometric constraints, outperforming state-of-the-art methods in structural adherence and f…
Researchers introduced ViSTR-Bench, a benchmark to evaluate whether multimodal large language models (MLLMs) can reason from continuous visual cues in dynamic scenes, finding that current models remain far below human pe…
AsymVerify, a confidence-gated verification system for political evasion detection, scored 0.85 Macro F1 on the evaluation split (D_eval, n=237), placing 2nd out of 41 teams at SemEval-2026 Task 6. The system, developed …
Researchers present ODeform, a method extending Neural Ordinary Differential Equations to model continuous 4D dynamics of deformable objects in 3D space, eliminating discrete time steps while maintaining computational ef…
Researchers propose Graph Wavelet Compressed Sensing (GWCS), a learning-based framework for offline compression of graph signals that combines a nonparametric multilevel importance sampler with a scale-aware graph neural…
A new study from arXiv (2607.20425v1) finds that reasoning-enabled LLMs evaluate literary quality with 79.3% accuracy across six quality tiers, prioritizing craft, depth, and distinctive voice over correctness. In system…
Researchers propose EAACD, an expert-aware adaptive contrast decoding method for mixture-of-experts (MoE) models, to mitigate hallucinations in large language models (LLMs). The method leverages distinct expert activatio…
LLM-INSTRUCT, the winning system for the UZH Shared Task at ArgMining 2026, achieved 1st overall on the official leaderboard for paragraph-level argument mining in UN and UNESCO resolutions by using constraint-aware retr…
Researchers propose Moir, a method that estimates the preservation covariance directly from a language model's own decoding distribution to prevent capability erosion during knowledge editing. In tests on OLMo-2, Llama-3…
Researchers have developed Merge-Adversarial Training, the first method to make open-source LLM watermarks durable against model merging, achieving up to +51 percentage points true positive rate at 1% false positive rate…
A new family of graph theory-based defenses called TopoGuard, introduced by researchers in a preprint on arXiv, catches 21 times more split-knowledge attacks on Retrieval Augmented Generation (RAG) systems than LlamaGuar…
Researchers propose Skeleton-Guided Reasoning Editing (SGRE), an Answer-then-Edit framework that modifies reasoning traces post-hoc to prevent unauthorized knowledge distillation of proprietary large language models (LLM…
A new study using LLMs and prediction markets finds that English news context systematically biases territorial predictions in Ukraine-related markets, with errors occurring 64 to 72 percent of the time when predictions …
A new study from arXiv shows that large language models (LLMs) deliver deceptive responses with substantial verbalized confidence, and human annotators prefer the higher-confidence deceptive response 78% of the time in p…
Axolotl3D, a multi-modal and occlusion-aware 3D generation model introduced in arXiv:2607.20660v1, jointly conditions on images, visibility masks, camera parameters, and a partial point cloud to enable faithful 3D shape …
Researchers from the Data Science and Analytics Group at Georgia Tech (DS@GT ARC) developed a privacy-preserving framework for synthetic lung CT slice generation for the ImageCLEFmed GANs 2026 challenge, achieving a Priv…
Researchers propose Uncertainty-Guided Cascade Forward Refinement (U-CFR), a new inference-time framework for interactive image segmentation that enables models to autonomously self-correct after each user interaction. U…