All Articles
17859 articles total
Revisiting the Platonic Representation Hypothesis: An Aristotelian View
This research challenges the Platonic Representation Hypothesis and proposes the Aristotelian Representation Hypothesis, suggesting NN representations converge to shared local neighborhood relationships.
ReportLogic: Evaluating Logical Quality in Deep Research Reports
ReportLogic is a benchmark and open-source LogicJudge designed to evaluate the logical quality and auditability of deep research reports generated by LLMs.
Learning to Select Maximum Clique Algorithms: From Traditional Machine Learning to a Dual-Channel Hybrid Neural Architecture
Researchers developed GAT-MLP, a dual-channel hybrid neural architecture that combines Graph Attention Networks and MLPs to select the most effective algorithm for solving the NP-hard Maximum Clique Problem.
Through the Looking Glass: A Dual Perspective on Weakly-Supervised Few-Shot Segmentation
The TLG model introduces a homologous but heterogeneous network design to improve weakly-supervised few-shot semantic segmentation, significantly outperforming fully supervised models with fewer parameters.
Reconstruction Alignment Improves Unified Multimodal Models
Reconstruction Alignment (RECA) is a resource-efficient post-training method that improves image generation and editing fidelity in unified multimodal models by leveraging visual understanding embeddings.
Limited Reference, Reliable Generation: A Two-Component Framework for Tabular Data Generation in Low-Data Regimes
ReFine is a two-component framework for generating synthetic tabular data in low-data regimes by extracting symbolic rules from interpretable models and applying dual-granularity filtering.
Rotary Position Encodings for Graphs
The WIRE approach applies rotary position encodings (RoPE) to graph-structured data by rotating tokens based on the graph Laplacian spectrum to inject structural information.
The Journal of Prompt-Engineered (Moral) Philosophy Or: Why AI-Assisted Ethics Research Requires Process Transparency
This paper argues for process transparency in AI-assisted ethics research, proposing a documentation-adequacy framework to ensure agent-integrity and meaningful human control.
Patent Representation Learning via Self-supervision
Researchers propose a self-supervised patent representation learning strategy using mixed dropout-section positives to better align different sections of patent documents.
Pianist Transformer: Towards Expressive Piano Performance Rendering via Scalable Self-Supervised Pre-Training
Pianist Transformer uses large-scale self-supervised pre-training on MIDI data and an efficient asymmetric Transformer to generate expressive, human-like piano performances.
The Best of the Two Worlds: Harmonizing Semantic and Hash IDs for Sequential Recommendation
H2Rec is a framework that harmonizes semantic IDs and hash IDs in sequential recommender systems to balance performance between head and tail items.
Improved Bounds for Private and Robust Alignment
This research establishes theoretical upper bounds on the suboptimality gap for private and robust alignment of language models, covering both offline and online settings.
US allows Anthropic to release Mythos to 'trusted partners'
The US government has allowed Anthropic to release its Mythos model to a select group of trusted partners.
Why does kinetic energy increase quadratically, not linearly, with speed? (2011)
An exploration into the physics of kinetic energy and why it increases quadratically with speed.
A Tiny Compiler for Data-Parallel Kernels
A presentation of a tiny compiler designed specifically for data-parallel kernels.
AI in Mathematics Is Forcing Big Questions
A discussion on how advancements in AI are forcing fundamental questions and changes in the field of mathematics.
Wearable Device-Based Real-Time Monitoring of Physiological Signals: Evaluating Cognitive Load Across Different Tasks
A study on using wearable devices to monitor EEG and HRV signals for real-time cognitive load assessment in students.
Byzantine-Robust Aggregation for Securing Decentralized Federated Learning
Introduction of WFAgg, a Byzantine-robust aggregation algorithm designed to secure decentralized federated learning environments.
Tuning Language Models by Mixture-of-Depths Ensemble
Introduction of the Mixture-of-Depths Ensemble (MoDE) framework to improve LLM reasoning by treating late layers as an ensemble.
Mitigating Hallucinations via Inter-Layer Consistency Aggregation in Large Vision-Language Models
Proposed DCLA, a training-free decoding mechanism to mitigate hallucinations in Large Vision-Language Models through inter-layer consistency.