All Articles
18316 articles total
Enhancing Fatigue Detection through Heterogeneous Multi-Source Data Integration and Cross-Domain Modality Imputation
A framework for fatigue detection in human operators using heterogeneous multi-source data and cross-domain modality imputation to handle real-world noise.
When Cars Have Stereotypes: Auditing Demographic Bias in Objects from Text-to-Image Models
SODA framework introduced to audit demographic biases in objects generated by text-to-image models, revealing strong implicit stereotypes.
"Did you lie?" Evaluating Lie Detectors across Model Scale and Belief-Verified Model Organisms
Researchers introduce 'Did-You-Lie' (DYL), a new method and testbed to evaluate the effectiveness of lie detectors in language models across various scales.
Simple Domain Generalization Methods are Strong Baselines for Open Domain Generalization
A study demonstrates that simple domain generalization methods like CORAL and MMD serve as strong and computationally efficient baselines for open domain generalization.
A DeepLearning Framework for Dynamic Estimation of Origin-Destination Sequence
A new deep learning framework is proposed to dynamically estimate origin-destination sequences in transportation, improving on traditional numerical optimization.
Quality Perceptions and Intended Engagement in Response to AI-Generated and AI-Assisted News
A survey in Switzerland found that readers perceive AI-generated and human-written news similarly in quality, though disclosure of AI can increase initial curiosity.
Scalable Batch Bayesian Optimization Via Subspace Acquisition Functions
Researchers propose a scalable batch Bayesian optimization approach using subspace acquisition functions to improve convergence speed in parallel computing.
VidCRAFT3: Camera, Object, and Lighting Control for Image-to-Video Generation
VidCRAFT3 is introduced as a unified framework for image-to-video generation with precise, coupled control over camera motion, object motion, and lighting.
Efficient Zeroth-Order Federated Finetuning of Language Models on Resource-Constrained Devices
A new zeroth-order optimization method is proposed for efficient federated finetuning of LLMs on resource-constrained edge devices, reducing computation by up to 3x.
Depth-Width tradeoffs in Algorithmic Reasoning of Graph Tasks with Transformers
Research reveals a depth-width tradeoff in Transformers for graph tasks, showing that increasing width can allow for shallower, faster models.
Generalized Kullback-Leibler Divergence Loss
The Generalized Kullback-Leibler (GKL) Divergence loss is introduced to improve adversarial robustness and knowledge distillation in vision-language models.
Revealing Hidden Vulnerabilities in Autoencoders through Gradient Signal Restoration
The GRILL framework is introduced to expose hidden vulnerabilities in autoencoders by restoring gradient signals in ill-conditioned layers during adversarial attacks.
GLM-5.2: The Most Powerful Open Model yet and the Brutal Reality of Running It
Discussion on GLM-5.2, a powerful open-source model, and the high hardware requirements for running it locally.
To study how chips work, MIT researchers built their own operating system
MIT researchers created a custom operating system specifically designed to study the internal workings of computer chips.
Show HN: Talos – Open-source WASM interpreter for Lean
Talos is an open-source WebAssembly (WASM) interpreter specifically developed for the Lean theorem prover.
Building a robotics research setup that lives next to my desk
A developer shares their experience building a personalized robotics research setup for home use.
Notation Matters: A Benchmark Study of Token-Optimized Formats in Agentic AI Systems
A benchmark study comparing token-optimized formats like TOON and TRON against JSON for structured data exchange in agentic AI systems.
AI Sovereignty as National Learning Capacity: A Human-Centered Learning Mechanics Viewpoint on France, the United States, and China
A viewpoint paper framing national AI sovereignty through the lens of Human-Centered Learning Mechanics (HCLM), specifically analyzing France.
SkillRevise: Improving LLM-Authored Agent Skills via Trace-Conditioned Skill Revision
SkillRevise is an execution-grounded framework that iteratively refines LLM-authored agent skills to improve task success rates.
DN-Hypo-Pipeline: An AI-Driven Workflow for Hypothesis Generation via Large Language Models and Scientific Explanations
DN-Hypo-Pipeline is an AI-driven workflow that uses LLMs and scientific explanations to generate novel research hypotheses.