AI/ML arXiv cs.AI

Mixtures of SubExperts for Large Language Continual Learning

Introduction of MoSEs, a modular and sparse framework for LLM continual learning to resolve the stability-plasticity dilemma.

Software Engineering Hacker News

Lobste.rs is now running on SQLite

The community forum Lobste.rs has migrated its database to SQLite.

Cybersecurity VentureBeat

Capital One releases VulnHunter, an open-source AI tool that finds software flaws before hackers do

Capital One has open-sourced VulnHunter, an AI-powered security tool that uses attacker-first forward analysis and a falsification engine to identify and fix software vulnerabilities.

AI/ML arXiv cs.AI

"Skill Issues'': Data-Centric Optimization of Lakehouse Agents

Researchers propose a data-centric optimization pipeline for coding agents operating on a branching lakehouse (Bauplan), improving task reward by up to 28.6%.

AI/ML arXiv cs.AI

Building Agent Harnesses for Scientific Curation from Multimodal Sources

Introduction of Beaver, an agent harness designed for structured scientific curation from multimodal sources, significantly outperforming frontier agents in evidence extraction.

AI/ML arXiv cs.AI

When Does Belief-Based Agent Memory Help? Reliability-Conditional Updating and Provenance-Capped Poisoning Defense

The Nous architecture explores belief-based long-term memory for LLM agents, finding that probabilistic updating is most effective when dealing with contradictory or unreliable evidence.

AI/ML arXiv cs.AI

Decoupled Alignment for Robust Plug-and-Play Adaptation

DAPA is a training-free safety enhancement method for LLMs using knowledge distillation and model fusion to prevent shadow alignment during adaptation.

AI/ML arXiv cs.AI

Empirical evidence of Large Language Model's influence on human spoken communication

Research indicates that LLM-generated lexical patterns (e.g., 'delve', 'meticulous') are increasingly being internalized and adopted into spontaneous human speech.

AI/ML arXiv cs.AI

Reinforcement Learning in Switching Non-Stationary Markov Decision Processes: Algorithms and Convergence Analysis

The SNS-MDP framework addresses RL in switching non-stationary environments, proving convergence for TD learning and Q-learning in time-varying systems.

AI/ML arXiv cs.AI

Generalized Fisher-Weighted SVD: Scalable Kronecker-Factored Fisher Approximation for Compressing Large Language Models

GFWSVD is introduced as a scalable LLM compression technique that uses Kronecker-factored Fisher approximations to account for parameter correlations.

AI/ML arXiv cs.AI

Fully Offline Reinforcement Learning

SOReL and TOReL are introduced as fully offline Bayesian RL methods that enable hyperparameter selection without online interaction.

Other Hacker News

Flock CEO Apologizes for Calling Activists 'Terrorists'

Flock CEO issues an apology after labeling activists as 'terrorists'.

Tech Business/VC Hacker News

Meta trying to destroy whistleblower Sarah Wynn-Williams, US senator says

A US senator claims Meta is attempting to silence or destroy a whistleblower, Sarah Wynn-Williams.

Other Hacker News

Lego building instructions through time

A look at how Lego building instructions have evolved over time.

Hardware/Chips TechCrunch

Agility Robotics plants its flag in Tesla’s backyard

Agility Robotics is opening a new training center for its Digit robots in Fremont, California.

AI/ML VentureBeat

Intuit scrapped its own AI agent architecture twice in four months. At VB Transform 2026, its AI VP called that the fast path

Intuit discusses the evolution of its AI agent architecture, moving from specialist agents to a skills-and-tools based system to reduce compounding errors in natural language handoffs.

AI/ML arXiv cs.AI

Subjective functions

A research paper proposing 'subjective functions' as higher-order objective functions endogenous to an agent to mimic human intelligence in goal synthesis.

AI/ML arXiv cs.AI

Large language models can effectively convince people to believe conspiracies

Study finds LLMs can be used to both promote and debunk conspiracy theories, though specific guardrails can mitigate the spread of misinformation.

AI/ML arXiv cs.AI

MedBeads: An AI-Native Clinical Context Graph Built from Immutable Beads and Reconstructable Clinical Links

Introduction of MedBeads, an AI-native clinical context graph using immutable beads and Merkle DAGs to provide auditable patient data for LLMs.

AI/ML arXiv cs.AI

Animating Petascale Time-varying Data on Commodity Hardware with LLM-assisted Scripting

A framework for animating petascale time-varying data on commodity hardware using a Generalized Animation Descriptor and LLM-assisted scripting.