All Articles
17626 articles total
ART for Diffusion Sampling: Continuous-Time Control and Actor-Critic Learning
Introduces ART for Diffusion Sampling, a continuous-time control formulation that learns adaptive timesteps to improve sample quality in diffusion models.
A Multi-Branch Hierarchy-Aware Framework for Heterogeneous Audio Classification
A new framework for heterogeneous audio classification using CLAP-based representations and hierarchy-aware classifiers to improve sound taxonomy accuracy.
MolSight: A Graph-Aware Vision-Language Model for Unified Chemical Image Understanding
Introduction of MolSight, a graph-aware vision-language model that integrates molecular topology to improve the understanding of chemical images for drug discovery.
Do Newer Lightweight CNNs Perform Better Under Resource Constraints? A Controlled Multigenerational Study of Architecture, Initialization, Training Budget, and Efficiency
A comparative study of lightweight CNNs revealing that newer architectures do not always outperform older ones like EfficientNet-B0 under resource constraints.
Mirror Illusion Art
AutoMIA is an automated pipeline for creating 3D mirror illusion art by jointly optimizing shape and color from 2D images.
Towards Load-Aware Prefill Deflection for Disaggregated LLM Serving
A load-aware prefill deflection scheduler for disaggregated LLM serving that reduces P95 TTFT by up to 81% by interleaving prefill and decode phases.
OpenSafeIntent: Evaluating Intent-Calibrated Safe Completion Across Dual-Use Prompt Sets
OpenSafeIntent introduces a benchmark to evaluate LLM safety by varying intent while keeping tasks fixed, revealing brittle safety behaviors.
SPLIT: Cross-Lingual Empathy and Cultural Grounding in English and Ukrainian LLM Responses
The SPLIT benchmark evaluates cross-lingual empathy and cultural grounding in LLMs, specifically comparing English and Ukrainian responses in crisis contexts.
Beyond the Performance Illusion: Structure-Aware Stratified Partitioning and Curriculum Distributionally Robust Optimization for Spatially Correlated Domains
Proposed a framework (SASP and CDRO) to prevent data leakage and hidden stratification in AI evaluation for spatiotemporally correlated domains.
Prompt Coverage Adequacy
Introduction of 'Prompt Coverage Adequacy', a testing metric for LLM-generated code that measures how well test suites satisfy prompt requirements.
SA-HGNN: Sample-Adaptive Hyperbolic Graph Neural Network for EEG-Based Depression Recognition
SA-HGNN is a hyperbolic graph neural network designed to recognize depression by capturing hierarchical functional connectivity in EEG brain networks.
Soatok's Informal Guide to Threat Models
A guide by Soatok on threat modeling, discussing how to identify assets, threats, and mitigations for a system.
Show HN: A statically typed, cross-platform, easily bootstrappable build system
A presentation of a new statically typed, cross-platform build system designed to be easily bootstrappable.
AIriskEval-edu: New Dataset for Risk Assessment in AI-mediated K-12 Educational Explanations
Introduction of AIriskEval-edu, a dataset for assessing pedagogical risks in AI-generated K-12 educational content.
CausalSteward: An Agentic Divide-Conquer-Combine Copilot for Causal Discovery
CausalSteward (CAST) is a human-in-the-loop multi-agent framework for assembling large causal models using a divide-and-conquer approach.
PhysMani: Physics-principled 3D World Model for Dynamic Object Manipulation
PhysMani is a framework that uses a physics-principled 3D Gaussian world model for better dynamic object manipulation in embodied AI.
Conditional Co-Ablation: Recovering Self-Repair Backups in Transformer Circuits
Conditional Co-Ablation (CoAx) is a new method for discovering 'self-repair' backup circuits in Transformers, improving mechanistic interpretability.
Robust for the Wrong Reasons: The Representational Geometry of LLM Robustness to Science Skepticism
Research on how LLMs respond to science skepticism, finding that they generally resist sycophancy but vary in their representational geometry.
NeoMap: Training-free Novel-View Synthesis from Single Images and Videos
NeoMap is a training-free framework for high-fidelity novel-view synthesis from single images or videos using pre-trained video models.
Object Aligner: A Configurable JSON Schema Similarity Score for Graphs, Applied to LLM Prompt Optimization
Object Aligner is an open-source Python library for deterministically scoring JSON schema similarity, useful for LLM prompt optimization.