AI/ML VentureBeat

Xiaomi's HarnessX rewrites its own AI scaffolding mid-task — and smaller models gain the most

Xiaomi introduces HarnessX, a framework that allows AI agents to autonomously rewrite and optimize their own software scaffolding to improve performance, especially for smaller models.

AI/ML arXiv cs.AI

EMAgnet: Parameter-Space EMA Regularization for Policy Gradient Self-Play in Large Games

EMAgnet is introduced as a parameter-space regularization method for policy gradient self-play in large games to achieve lower exploitability compared to PPO.

AI/ML arXiv cs.AI

Towards Spec Learning: Inference-Time Alignment from Preference Pairs

Spec learning is proposed as an inference-time alignment framework that uses small preference sets to compile natural-language specifications, avoiding expensive fine-tuning.

AI/ML arXiv cs.AI

Fast and Slow Variational Continual Learning

The CoVON optimizer is introduced to implement fast and slow adaptation in variational continual learning, improving stability and plasticity in LLM fine-tuning.

AI/ML arXiv cs.AI

Towards Version-aware Operations and Transaction Memories for Multi-layer MeMo

This paper proposes a version-aware operation layer for MeMo language models, allowing knowledge updates via explicit memory edits instead of full retraining.

Hardware/Chips arXiv cs.AI

Rapid FinFET Modelling Using an Autoencoder

A machine learning framework using autoencoders is presented for the rapid modeling and characterization of FinFET device physics.

AI/ML arXiv cs.AI

RAVEN: A Regime-Aware Variable-context Expert Network for Financial Time Series Forecasting

RAVEN is a Mixture-of-Experts framework for financial time series forecasting that adaptively determines temporal context to handle non-stationary data.

AI/ML arXiv cs.AI

Selective Capability Unlearning in End-to-End Spoken Language Understanding

The Binding Subspace (BSU) framework is proposed to selectively unlearn specific capabilities in spoken language understanding systems to prevent capability persistence.

AI/ML arXiv cs.AI

Token Complexity of Certifying Stochastic-Oracle Reliability

This research establishes the token complexity bounds for certifying the reliability of stochastic oracles using a SOTM framework.

Cybersecurity Hacker News

NSA lost access to Mythos amid Anthropic dispute

The NSA reportedly lost access to Mythos, a tool associated with Anthropic, following a dispute between the two entities.

Software Engineering Hacker News

How a Microsecond-Level Low-Latency Engine Works

A deep dive into the architecture and implementation of a microsecond-level low-latency engine.

Other Hacker News

Journalism is rearranging the deckchairs. It needs to reinvent itself

An opinion piece arguing that modern journalism needs a fundamental reinvention to survive and remain relevant.

Other Ars Technica

Formula E reveals first calendar for GEN4 with lots of real race tracks

Formula E reveals its first calendar for the GEN4 era, featuring more traditional racing tracks for 2027.

Tech Business/VC Ars Technica

Google starts lowering Play Store fees, making good on Epic Games settlement

Google is lowering Play Store fees as part of a legal settlement with Epic Games.

Other Ars Technica

Elon Musk denies Tesla’s Autopilot caused crash that killed grandmother

Elon Musk denies claims that Tesla's Autopilot system caused a fatal crash involving a grandmother.

AI/ML VentureBeat

Stanford researchers will discuss their agentic 'scientists' that are on course to reshape drug discovery at VB Transform 2026

Stanford researchers are using multi-agent AI systems to automate the pharmaceutical drug discovery lifecycle, from initial discovery to clinical trial design.

AI/ML arXiv cs.AI

Faithful by Construction: Claim-Anchored Attribution for Multi-Document Summarization

Introduction of CAMS, a framework for multi-document summarization that uses claim-anchored attribution to reduce hallucinations and improve faithfulness.

AI/ML arXiv cs.AI

RASC+: Retrieval-Constrained LLM Adjudication for Clinical Value Set Authoring

RASC+ utilizes retrieval-constrained LLM adjudication to improve the accuracy and safety of clinical value set authoring.

AI/ML arXiv cs.AI

Learning to Trigger: Reinforcement Learning at the Large Hadron Collider

Researchers demonstrate the first use of Reinforcement Learning (RL) to control real-time event triggering at the Large Hadron Collider, improving signal efficiency.

Hardware/Chips Hacker News

The Xteink X4 E-Ink Reader

The article discusses the Xteink X4, an E-Ink reader, though specific details are limited in the snippet.