AI Research

Researchers introduced Mesh-RL, a spatial domain-decomposition framework that partitions environments into overlapping subgrids to accelerate temporal-difference learning in reinforcement learning. The method improved co…

Researchers introduced Neural Voxel Dynamics, a self-supervised framework that learns implicit 3D physics from video by lifting 2D features into a volumetric latent space. The method achieves long-term structural stabili…

Researchers introduced the Frame Forgetting Network (FFN) for test-time training on long videos, using only three frames per sliding window to reduce computational cost. The model defines a surprise metric to adaptively …

Researchers at arXiv propose a multi-dimensional evaluation framework for AI agents beyond accuracy saturation, using CORE-Bench Hard as a case study. They introduce CORE-Bench v1.1 and an out-of-distribution suite, find…

Researchers introduced CRISP, a new evaluation paradigm that diagnoses visual spatial intelligence in vision-language models by measuring consistency between implicit perception and explicit reasoning. The study found th…

Researchers introduced AlgoEvolve, an LLM-driven evolutionary framework that generates and iteratively improves algorithmic trading strategies expressed as Python code. The system exhibits emergent regime-adaptive strate…

Researchers introduced PhyEditBench, a benchmark for evaluating physics-aware image editing in generative models, revealing significant limitations in current state-of-the-art methods. The benchmark includes 238 real-wor…

Researchers developed a hybrid approach combining image processing and deep learning to assess fruit freshness, achieving over 90% accuracy in real-time classification of apples and oranges. The method uses a CNN for bin…

Researchers introduced OpenFinGym, a unified gym environment for evaluating large language model agents across multiple quantitative-finance tasks including forecasting, trading, and fraud detection. The platform address…

Researchers introduced a Geometry-Aware Monte Carlo Tree Search (MCTS) framework to solve extremal combinatorial geometry problems, achieving new best-known results on five out of six problems, including configurations o…

New research from arXiv reveals that larger language models consistently outperform smaller ones on reasoning benchmarks by an average of 6-7%, with the key advantage being constraint-guided reasoning—the ability to iden…

Researchers proposed Dynamic-dLLM, a training-free framework to accelerate Diffusion Large Language Models (dLLMs) by dynamically allocating cache budgets and calibrating decoding thresholds. The method achieves over 3x …

Researchers at arXiv found that large language models (LLMs) perform well on text-only statics problems but their accuracy drops when diagrams are added and multi-step reasoning is required. The performance decline is at…

A study comparing LLM-generated research methods against 1,000 arXiv computer-science papers found that LLMs suggest a much narrower range of methods, with provider choice showing the strongest imbalance. Academic models…

Researchers propose GeMoE, a dynamic routing method for Mixture-of-Experts-based large vision-language models that uses gating entropy to adaptively select the number of experts per token, achieving 99.5% performance ret…

A new study from Japan's AI Security Institute finds that LLM-as-judge safety evaluations are not reproducible even at temperature 0, with per-item disagreement up to 50% across runs. The researchers tested 690 API calls…

Researchers introduced HierBias, a hierarchical context-conditioned media bias detector that models document context to improve sentence-level bias classification. The system achieved 0.853 F1 and 0.723 MCC on BABE and B…

Researchers introduced Know2Guess, a contamination-aware multi-zone benchmark with 1,200 items across five domains to evaluate large language models' ability to distinguish answerable knowledge from abstention-expected u…

A new study on Llama 3.1 8B finds that helpfulness post-training (SFT and GRPO) significantly degrades animal compassion values compared to coding-domain post-training, with a 35.7% vs. 65.2% gap on the Animal Harm Bench…

Researchers released ProfileFoundry, a synthetic dataset of 100,000 person objects with 709,228 events, 40,338 households, and 52,491 employers, designed for evaluating LLM agents on memory, privacy, and tool-use tasks. …