AI/ML arXiv cs.AI

LiveStarPro: Proactive Streaming Video Understanding with Hierarchical Memory for Long-Horizon Streams

Introduction of LiveStarPro, a proactive streaming video understanding assistant featuring hierarchical memory and streaming verification decoding for long-horizon streams.

Software Engineering arXiv cs.AI

Position: Coding Benchmarks Are Misaligned with Agentic Software Engineering

A critical analysis arguing that current coding benchmarks are misaligned with agentic software engineering by conflating models with their system harnesses.

AI/ML arXiv cs.AI

No-Free-Fairness: Fundamental Limits and Trade-offs in Learning Systems

The 'No-Free-Fairness' theorems establish theoretical limits on fairness in learning systems, showing that disparity arises from intrinsic problem structures and finite data.

AI/ML arXiv cs.AI

Conservation Laws for Modern Neural Architectures

A unified framework to characterize conservation laws in modern neural architectures, including GELU/SiLU activations and Mixture-of-Experts.

AI/ML arXiv cs.AI

A Framework for Evaluating Agentic Skills at Scale

A framework for evaluating 'agent skills' at scale, demonstrating how structured knowledge artifacts influence LLM agent behavior and performance.

AI/ML arXiv cs.AI

Human-in-the-Loop Atlas-Based 3D Asset Segmentation for Interactive Content Workflows

A human-in-the-loop pipeline for generating 3D asset segmentation atlases using SAM 2 and Label Studio for interactive content workflows.

AI/ML arXiv cs.AI

When Multiple Scripts Matter: Evaluating ASR in Clinical Settings

Introduction of MultiClin, a benchmark for evaluating ASR robustness to multiscript variability in clinical settings.

AI/ML arXiv cs.AI

Functional Equivalence in Attention: A Comprehensive Study with Applications to Linear Mode Connectivity

A study on functional equivalence in Transformers, explaining why rotary positional encodings (RoPE) enhance expressivity compared to sinusoidal encodings.

AI/ML arXiv cs.AI

Perceptual compensation for tonal context in self-supervised speech models

Research suggests that self-supervised pre-training in wav2vec2.0 may not be sufficient for phonological regularities without supervised objectives.

Hardware/Chips Hacker News

[x86] AI Compute Extensions (ACE) Specification

Introduction of the x86 AI Compute Extensions (ACE) Specification to enhance AI workloads on x86 architectures.

Other Hacker News

Biological evolution and information acquisition

A theoretical discussion on biological evolution and how information is acquired through evolutionary processes.

Hardware/Chips The Verge

Midjourney Medical goes from generating ‘cat images’ to full-body ultrasound scans

Midjourney is expanding into medical hardware with a full-body ultrasound scanner designed to provide MRI-like quality imaging.

AI/ML arXiv cs.AI

Vision-language models for chest radiography do not always need the image

Research indicates that some medical vision-language models may rely on text priors rather than actually analyzing the radiographic images for diagnosis.

Cybersecurity arXiv cs.AI

Structured Adversarial Camouflage via Voronoi Diagrams

Proposed a new method for adversarial camouflage using Voronoi diagrams to deceive person detection AI models effectively.

AI/ML arXiv cs.AI

ED3R: Energy-Aware Distributed Disaster Detection Enabled by Cooperative Robotic Agents

Introduction of ED3R, an energy-aware distributed framework for cooperative robotic agents to detect wildfires under uncertainty.

Other arXiv cs.AI

Symplectic Transversality and Endpoint Green Estimates for Finite-Horizon Pontryagin Systems

A mathematical study on finite-horizon Pontryagin systems focusing on symplectic transversality and endpoint Green estimates.

AI/ML arXiv cs.AI

Talking to Your Data: Exploring Embodied Conversation as an Interface for Personal Health Reflection

Exploring an embodied conversational agent as a user interface for reflecting on personal health data from wearables.

AI/ML arXiv cs.AI

A Neuromorphic Trigger for Efficient Audio Event Detection

A neuromorphic trigger using spiking neural networks to efficiently gate audio events, significantly reducing computational FLOPs for downstream models.

Hardware/Chips arXiv cs.AI

MIVE: A Minimalist Integer Vector Engine for Softmax LayerNorm and RMSNorm Acceleration

MIVE is a minimalist integer vector engine designed to accelerate Softmax, LayerNorm, and RMSNorm operations in LLM hardware accelerators.

Other Hacker News

Show HN: Spin Lab

Spin Lab is a new tool shared on Hacker News, though the provided snippet contains no details.