All Articles
17516 articles total
NonTextual Target Attack
Researchers propose NonTextual Target Attack (NTA), a gradient-based jailbreak method for LLMs that maximizes unsafety probability without requiring fixed target responses.
Rapidly Learning Soft Robot Control via Implicit Time-Stepping
A new approach using implicit time-stepping and the DisMech simulator enables rapid policy learning for soft robots, significantly outperforming previous frameworks.
Refine Thought: A Test-Time Inference Method for Embedding Model Reasoning
Refine Thought (RT) is introduced as a test-time inference method to improve the semantic reasoning capabilities of text embedding models through multiple forward passes.
T2T-VICL: Cross-Task Visual In-Context Learning via Implicit Text-Driven VLMs
T2T-VICL is a prompt-transfer framework that enables Visual In-Context Learning (VICL) across different tasks by distilling teacher VLM knowledge into a lightweight student model.
Thinking Ahead: Foresight Intelligence in MLLMs and World Model
Researchers introduce FSU-QA, a VQA dataset designed to evaluate and improve 'Foresight Intelligence' in multimodal LLMs and world models.
Show HN: I built a web tool to see and edit what an AI thinks before it answers
A developer has shared a web tool that allows users to see and edit the internal thought process of an AI before it provides a final answer.
How GitHub gave every repository a durable owner
GitHub explains the technical implementation of providing every repository with a durable owner to improve repository management.
Patch for Windows Defender 0-day could allow attackers to fill hard disk
A patch was released for a Windows Defender 0-day vulnerability that could allow attackers to fill the hard disk.
Allstate accuses Broadcom of auditing it because it quit VMware, CA
Allstate and Broadcom are engaged in a legal dispute regarding VMware audits following Allstate's decision to quit VMware.
Humanoid robots controlled by surgeons did world-first operation on live pigs
Humanoid robots controlled by surgeons successfully performed a world-first operation on live pigs in a preclinical trial.
Understanding Two-Layer Neural Networks with Smooth Activation Functions
Research exploring the training solutions of two-layer neural networks with smooth activation functions, aiming to reveal the 'black box' of the solution space.
L-GTA: Latent Generative Modeling for Time Series Augmentation
Introduction of L-GTA, a generative model based on VAE and Bi-LSTM for time series data augmentation, outperforming existing SOTA methods.
A Study of Commonsense Reasoning over Visual Object Properties
A systematic evaluation framework and benchmarks (OPTICS-CNT and OPTICS-CMP) designed to test commonsense reasoning over visual object properties in VLMs.
LHM-Humanoid: Long-Horizon Human Motion Control for Continuous Object Transport in Cluttered Scenes
LHM-Humanoid introduces a new approach for continuous, reset-free long-horizon human motion control in humanoid robots within cluttered scenes.
Explain Before You Answer: A Survey on Compositional Visual Reasoning
A comprehensive survey paper reviewing over 260 papers on compositional visual reasoning in multimodal AI, mapping the paradigm shift from 2023 to 2025.
Show HN: RandoFont, a browser for Google Fonts
RandoFont is a browser tool designed to help users explore and select fonts from the Google Fonts library.
I Changed My Name
A personal post regarding a name change, providing no technical or industry value.
The ChatGPT browser is already dead
OpenAI is shutting down ChatGPT Atlas, its AI-powered browser, less than a year after launch to focus on productivity features.
The floatable, powerful Soundcore Boom 2 speaker is over half off
The Soundcore Boom 2 waterproof Bluetooth speaker is currently on sale via Woot.
Shared API keys expose AI agents at 69% of enterprises, new VentureBeat research finds
Research reveals that 69% of enterprises use shared API keys for AI agents, creating significant security vulnerabilities and driving a wave of security acquisitions.
