AI/ML arXiv cs.AI

ShortOPD: Recovering Pruned LLMs with Short-to-Long On-Policy Distillation

Introduction of ShortOPD, a method to recover pruned LLMs using a short-to-long on-policy distillation schedule to improve generation quality.

AI/ML arXiv cs.AI

Boogu-Image-0.1: Boosting Open-Source Unified Multimodal Understanding and Generation

Release of Boogu-Image-0.1, an open-source unified multimodal model for image generation and understanding with efficient training costs.

AI/ML arXiv cs.AI

Active Beyond-Diagonal RIS Empowered Heterogeneous Edge Computing: A Distributional Reinforcement Learning Approach

Proposed DSAC-T framework using distributional reinforcement learning to optimize resource allocation in active BD-RIS empowered edge computing.

AI/ML arXiv cs.AI

What Models Express, Suppress, and Resist: Auditing Open-Weight LLMs with Persona Vectors

Research using persona vectors to audit open-weight LLMs, revealing how post-training influences expressed and latent behaviors.

AI/ML arXiv cs.AI

SteinGate: Tail-Sensitive Safe Reinforcement Learning via Stein Discrepancy

Introduction of SteinGate, a safety certificate for reinforcement learning that uses Stein Discrepancy to better detect catastrophic tail events.

AI/ML arXiv cs.AI

RAGthoven at SemEval-2026 Task 1: A Multi-Stage Pipeline Walks Into a Benchmark and Barely Clears the Bar

Evaluation of RAGthoven, a multi-stage pipeline for constrained humor generation, suggesting diminishing returns for complex agentic scaffolding with frontier models.

AI/ML arXiv cs.AI

Analyzing Curricular Pattern Complexity Using AI to Improve On-Time Graduation Rates

Researchers are using Large Language Models to analyze and revise undergraduate software engineering curricula to reduce graduation bottlenecks and improve on-time graduation rates.

AI/ML arXiv cs.AI

Full-Pipeline Inference Optimization for MiMo-V2.5 Series: Pushing Hybrid SWA Efficiency to the Limit

The authors present a full-pipeline inference optimization for the MiMo-V2.5 model family, focusing on Hybrid SWA and distributed KVCache infrastructure called GCache.

AI/ML arXiv cs.AI

WaterMoE: Expert-Routing-based Watermarking for High Fidelity and Efficiency

WaterMoE is a new watermarking scheme for MoE LLMs that embeds signals into expert selection to ensure high fidelity and low inference overhead.

AI/ML arXiv cs.AI

TSSM: Triaxial State Space Model for Global Station Weather Forecasting with Temporal-Variable-Historical Modeling

The Triaxial State Space Model (TSSM) is introduced for global station weather forecasting, utilizing a history-enhanced paradigm to improve accuracy in extreme event prediction.

AI/ML arXiv cs.AI

Disentangling Knowledge States with Ability and Proficiency Modeling for Knowledge Tracing

Phase-Aware Knowledge Tracing (PAKT) decomposes student interactions into ability and proficiency phases to better predict future learning performance.

AI/ML arXiv cs.AI

STKAN: Kolmogorov-Arnold Networks for Spatio-Temporal Forecasting

STKAN introduces Taylor-polynomial Kolmogorov-Arnold Network modules into spatio-temporal forecasting for traffic data, offering a complement to architectural design.

AI/ML arXiv cs.AI

A Hybrid Mamba for Audio-Visual Navigation

Samba is a hybrid Mamba-based architecture for audio-visual navigation that replaces GRUs with Mamba State Encoders to improve generalization and efficiency.

Software Engineering arXiv cs.AI

SemaDiff: Identifying Semantic-Changing Commits with Generated Code and Tests

SemaDiff is a novel approach to identify semantic-preserving commits by generating tests for modified code using LLMs to detect behavioral differences.

AI/ML arXiv cs.AI

CoDiffGRN: Rethinking Gene Regulatory Network Inference via the BEELINE-KGC Benchmark and Co-evolutionary Discrete Diffusion

CoDiffGRN is a co-evolutionary discrete diffusion framework for inferring gene regulatory networks, outperforming existing methods in inductive generalization.

Cybersecurity arXiv cs.AI

AI in Cyberpsychology: A systematic literature review of Cybersecurity enhancement by using AI for analyzing psychology of Victims, Attackers, and Defenders

A systematic review examines the intersection of AI and cyberpsychology, analyzing how AI can enhance cybersecurity by decoding behavioral patterns of victims and attackers.

Tech Business/VC Hacker News

Where are YC founders now? OpenAI and Anthropic, mostly

A discussion on the current trajectory of Y Combinator founders, noting a heavy concentration in OpenAI and Anthropic.

Hardware/Chips TechCrunch

Phone maker OnePlus says it won’t release new phones in the U.S. and Europe

OnePlus announces it will stop releasing new phones in the U.S. and Europe, with potential winding down of India operations by 2027.

AI/ML arXiv cs.AI

The Entanglement Wall: Activation-Space Probes as Risk Detectors, Not Context Adjudicators

Research explores using activation-space probes to detect harmful requests in LLMs, finding they act as broad risk detectors rather than context-specific adjudicators.

AI/ML arXiv cs.AI

The Hitchhiker's Guide to Monoculture

A study on AI coding assistants shows they cause syntactic homogenization (standardizing code structure) but not semantic homogenization (problem-solving strategies).