AI/ML arXiv cs.AI

LazyMem: Retrieve Broadly, Construct Selectively for Efficient Long-Term Agent Memory

LazyMem is a memory management system for LLM agents that defers memory construction to query time for better efficiency and reduced noise.

AI/ML arXiv cs.AI

HiLLTS: Zero-Shot Hierarchical LLM-Guided Traffic Signal Control for Sustainable Transportation

HiLLTS is a hierarchical LLM-guided framework for urban traffic signal control designed to reduce congestion and CO2 emissions.

Cybersecurity arXiv cs.AI

Risk Governance for Generative AI Mental Health Support: A Multi-Turn Safety Architecture

A multi-turn safety architecture for generative AI in mental health support focusing on contextual risk detection and protocol-guided responses.

AI/ML arXiv cs.AI

Bayesian Repetition Penalty: A Principled Adjacent-Conditional Framework for Reversing Attention Collapse in Autoregressive Language Models

A Bayesian Repetition Penalty framework to reverse attention collapse (repetitive loops) in autoregressive language models via frozen-bias mechanisms.

Cybersecurity arXiv cs.AI

PANOPTICON: A PII-Based Assemblage of Naturalistic Output Tokens for Investigating Privacy Leakage Within LLM Context Window

PANOPTICON is a PII-based dataset and pipeline for investigating privacy leakage and Prompt Inversion Attacks within LLM context windows.

Software Engineering Hacker News

Lightweight Spring Boot Monitoring Without Prometheus and Grafana

A discussion on monitoring Spring Boot applications without the traditional Prometheus and Grafana stack.

Software Engineering Hacker News

Lockless MPSC FIFO queues for io_uring

Implementation of lockless MPSC (Multi-Producer Single-Consumer) FIFO queues specifically optimized for io_uring.

AI/ML arXiv cs.AI

CuraWeb: Joint Optimization of Quality, Redundancy, and Diversity for Web-Scale Pretraining Data

Introduction of CuraWeb, a 2T-token English corpus curated via joint optimization of quality, redundancy, and diversity to improve LLM pretraining.

AI/ML arXiv cs.AI

Beyond Block Boundaries: Multi-Block Editing for Diffusion Large Language Models

Proposal of Multi-Block Editing (MBE) to solve the block boundary problem in discrete diffusion language models, improving long-range consistency.

AI/ML arXiv cs.AI

Obliviate: Efficient Unlearning in Recommender Systems

Obliviate is an efficient two-stage unlearning framework for recommender systems to remove user interaction data while preserving utility.

AI/ML arXiv cs.AI

Reinforcement Learning for Heterogeneous Sensor Selection in Maritime Surveillance

A reinforcement learning framework for optimizing heterogeneous sensor selection in maritime surveillance to reduce computational overhead.

AI/ML arXiv cs.AI

AIR-BENCH Live: An Evolving Safety Benchmark for Foundation Models

AIR-BENCH Live, a self-evolving safety benchmark for foundation models that updates its risk taxonomy based on government regulations.

AI/ML arXiv cs.AI

How LLM Task-Adaptation Reshapes Alignment: A Multi-dimensional Study of Behavioral and Representational Drift

A study on how post-training methods like SFT and RLVR affect the alignment and behavioral drift of large language models.

AI/ML arXiv cs.AI

DOSA: A Tree-Guided, Self-Regressive Framework for Long Document Structure Analysis

Introduction of DOSA, a tree-guided self-regressive framework for analyzing the structural relations in long, visually-rich documents.

AI/ML arXiv cs.AI

A Vocabulary for Multi-Agent Automated Research Systems

A proposed vocabulary for describing and comparing the design choices of multi-agent automated research systems.

Software Engineering Hacker News

The Difference Between a Button and a Link

A discussion on the semantic and functional differences between using buttons and links in web interfaces.

Other Hacker News

Schools are adding pepper-spraying drones to help combat active shooters

Schools are implementing drones equipped with pepper spray to mitigate active shooter threats.

Open Source Hacker News

Hubbele: Open-source notetaking app for you and your agents

Hubbele is an open-source notetaking application designed for use by both humans and AI agents.

Software Engineering Hacker News

Show HN: Manim (3Blue1Brown's animation engine) in the browser via WebGPU

Manim, 3Blue1Brown's mathematical animation engine, has been ported to the browser using WebGPU.

Hardware/Chips Hacker News

ReFrame – The EPaper Camera

ReFrame is a camera that utilizes an EPaper display to provide a unique photographic experience.