AI Research Paper Feed

TLDR: GraphRAG-Bench is a new benchmark designed to rigorously evaluate the reasoning capabilities of Graph Retrieval Augmented Generation models using challenging, domain-specific questions across diverse tasks.

TLDR: This paper introduces a benchmark to test if AI agents prioritize safety principles over conflicting operational goals using a grid world scenario.

TLDR: This study compares three AI architectures for entity relationship classification, finding that a novel multi-agent approach outperforms standard methods and approaches fine-tuned models' performance.

TLDR: Random Fourier features can enhance tabular deep learning by stabilizing and accelerating training, without requiring additional tuning or embeddings.

TLDR: VS-Bench is a new benchmark for evaluating Vision Language Models (VLMs) in multi-agent environments, revealing significant gaps in current models' strategic reasoning and decision-making abilities.

TLDR: MISLEADER is a novel defense strategy against model extraction attacks that uses ensembles of distilled models to maintain utility while reducing extractability without relying on out-of-distribution assumptions.

TLDR: FroM is a new adaptive model merging method that uses the Frobenius norm to effectively combine fine-tuned models without data, reducing task interference.

TLDR: This paper reviews various datasets and machine learning methods for intrusion detection systems, highlighting advances and challenges in the field.

TLDR: SingaKids is a multilingual, multimodal dialogic tutor that enhances language learning for children through interactive picture description tasks in four languages.

TLDR: Comba is a new Nonlinear RNN variant that enhances performance and efficiency in sequence modeling by using closed-loop control theory and a scalar-plus-low-rank state transition.

TLDR: OThink-R1 is a method that reduces unnecessary reasoning in large reasoning models by switching between fast-thinking and slow-thinking modes, improving efficiency without losing accuracy.

TLDR: Consultant Decoding (CD) improves inference speed and quality for large language models by using token-level likelihoods for draft verification, achieving significant efficiency gains over traditional speculative decoding.

TLDR: This paper presents a novel algorithm for best feasible arm identification in linear bandits with a fixed budget, achieving optimal error decay rates using a posterior sampling framework.

TLDR: The study introduces X$^2$-ICL, an improved in-context learning method that uses explanations for all possible labels to enhance robustness in language models.

TLDR: The paper explores multi-modal learning using spiked matrix-tensor models, revealing that sequential learning strategies can overcome computational challenges and achieve optimal signal recovery.

TLDR: MASTER is a novel data augmentation method that uses multi-agent simulated teaching to enhance large language models' instruction-following capabilities and reasoning abilities.

TLDR: Symmetry-Aware GFlowNets (SA-GFN) corrects biases in graph sampling by integrating symmetry corrections into the reward structure, improving the diversity and accuracy of generated graphs.

TLDR: VidEvent is a large-scale dataset designed to improve AI's understanding of dynamic events in videos, providing over 23,000 annotated events from movie recaps for research and development.

TLDR: AERO is a novel optimization framework inspired by Judo, designed to improve stability and adaptability in probabilistic forecasting under uncertainty.

TLDR: DIAMOND is an LLM-driven agent that enhances baseball highlight summarization by integrating sabermetrics with natural language reasoning, significantly improving performance over traditional methods.