AI Research Paper Feed

TLDR: RRAttention is a novel attention mechanism that reduces computational complexity while maintaining performance in processing long contexts by using a dynamic block sparse attention method with a round-robin sampling strategy.

TLDR: The Share method enables efficient continual learning by using a single, adaptable low-rank subspace, greatly reducing parameters and memory compared to traditional methods while maintaining performance.

TLDR: The paper introduces Dynamic Sliding Block (DSB), a training-free scheduling method for diffusion large language models that adapts to semantic difficulty, improving both quality and efficiency of text generation.

TLDR: KV-CoRE is a method to evaluate the data-dependent compressibility of kv-caches in large language models, revealing patterns linked to model architecture and training data across multiple languages.

TLDR: CASTLE is a benchmark designed to evaluate the personalized safety of large language models in educational settings, revealing significant deficiencies in current models' abilities to tailor responses to individual student needs and risks.

TLDR: ADANA, an optimizer with time-varying schedules for hyperparameters, improves large-scale language model training efficiency by up to 40% compared to AdamW.

TLDR: The paper introduces exPreCast, a deterministic model for accurate and efficient nowcasting of both normal and extreme rainfall using a balanced dataset from the Korea Meteorological Administration.

TLDR: The SAGE benchmark reveals that traditional BM25 outperforms LLM-based retrievers for scientific literature retrieval, with enhancements possible through document augmentation using LLMs.

TLDR: The paper introduces a flow matching-based framework for disentangled representation learning that improves semantic alignment and disentanglement scores by using a non-overlap regularizer to reduce factor interference.

TLDR: This study examines how multilingual large language models (LLMs) respond to value-laden multiple-choice questions across different languages, revealing variability in consistency and language-specific behaviors.

TLDR: The study finds no systematic bias against non-native Czech speakers in LLM-based text detectors and shows that modern detectors do not rely on perplexity to identify generated text.

TLDR: LongR is a framework that improves long-context reasoning in reinforcement learning by using a dynamic 'Think-and-Read' mechanism and dense utility rewards, achieving significant gains on benchmarks like LongBench v2.

TLDR: The paper introduces a novel method using causal front-door adjustment to effectively bypass safety mechanisms in large language models for jailbreak attacks.

TLDR: MentorCollab improves small model reasoning by selectively using guidance from a large model, enhancing performance with minimal additional cost.

TLDR: This paper introduces relaxed unitary convolutions for graph neural networks to improve performance in dynamics modeling by balancing smoothness preservation with natural physical system requirements.

TLDR: BudgetMem is a runtime memory framework for LLMs that optimizes performance-cost trade-offs using query-aware budget-tier routing across memory modules.

TLDR: PatchFlow improves defect detection in die casting using local patch features and a flow-based model, reducing error rates significantly on multiple datasets.

TLDR: The paper introduces a framework using piecewise deterministic Markov processes (PDMP) for efficient Bayesian inference in complex inverse problems, demonstrating improved accuracy and efficiency over traditional methods.

TLDR: The paper introduces new Radon--Wasserstein gradient flows for efficient high-dimensional sampling using interacting particles with linear scaling costs.

TLDR: CORAL is a method that improves the accuracy and calibration of large language models during inference without retraining.