AI/ML arXiv cs.AI

Multi-Agent Routing as Set-Valued Prediction: A WildChat Benchmark and Cost-Aware Evaluation

A new benchmark and evaluation protocol based on WildChat is introduced for cost-aware multi-agent routing.

AI/ML arXiv cs.AI

DLR: Zero-Inference-Cost Latent Residuals for Low-Rank Pre-Training

DLR (Duplicated Latent Residuals) is a parameter-free plug-in that improves the efficiency and quality of low-rank pre-training for LLMs with zero inference cost.

AI/ML arXiv cs.AI

Machine-learnable Sets

This study proposes a formal definition and experiments on 'machine-learnable sets' using Boolean autoencoders.

Cybersecurity arXiv cs.AI

Clustering Unsupervised Representations as Defense against Poisoning Attacks on Speech Commands Classification System

A new defense mechanism using DINO unsupervised representations and clustering is proposed to protect speech command classification systems from poisoning attacks.

Tech Business/VC Hacker News

Department of Commerce has lifted export controls on Claude Fable 5 and Mythos 5

The US Department of Commerce has lifted export controls on Anthropic's Claude Fable 5 and Mythos 5 models.

Tech Business/VC The Verge

Anthropic’s long-sidelined Fable 5 is greenlit to return

Anthropic plans to restore access to Claude Fable 5 and Mythos 5 after the US Department of Commerce lifted export controls.

AI/ML arXiv cs.AI

HARD-KV: Head-Adaptive Regularization for Decoding-time KV Compression

Researchers introduce HARD-KV, a framework that improves LLM inference throughput by bridging dynamic KV compression with static memory constraints.

AI/ML arXiv cs.AI

Fisher-Routed Mixture of Experts for Federated Class-Incremental Learning

The FedFMX framework proposes Fisher-Routed Mixture of Experts to improve federated class-incremental learning by addressing capacity conflict and catastrophic forgetting.

AI/ML arXiv cs.AI

LAMP: Lean-based Agentic framework with MCP and Proof Repair

LAMP is a multi-agent framework using the Model Context Protocol to synthesize kernel-verified Lean 4 proofs, specifically enhancing Combinatorics on Words (CoW) formalization.

AI/ML arXiv cs.AI

The Heterogeneous Safety Impacts of Benign Multilingual Fine-Tuning

A study reveals that benign multilingual fine-tuning can unexpectedly increase an LLM's vulnerability to unsafe adversarial prompts across different languages.

AI/ML arXiv cs.AI

Perspectives on Latent Factor Indeterminacy and its Implications for Data Representation

This paper analyzes latent factor indeterminacy in generative models, suggesting that the factor model is highly suited for representation learning of very-high-dimensional data.

Other arXiv cs.AI

Building AI-Ready Data Systems for Space Life Sciences, Aerospace Medicine, and Deep Space Exploration

The authors propose a three-tier approach to transition space flight biological data from FAIR to AI-ready and space-ready formats to enable better AI analysis.

AI/ML arXiv cs.AI

Defeat Devices in AI Systems

Researchers define 'defeat devices' in AI—mechanisms where models behave differently in evaluation vs deployment—and propose a forensic detection protocol (TADP).

AI/ML arXiv cs.AI

An Integrated Machine Learning and Hierarchical Variance Decomposition Pipeline for Student Performance Prediction and Metacognitive Calibration on Multi-Signal Telemetry

The UBP-CAP pipeline integrates machine learning and variance decomposition to predict student performance and analyze metacognitive calibration in tutoring systems.

AI/ML arXiv cs.AI

5ting at SemEval-2026 Task 8: Strong End-to-End Multi-Turn RAG via LLM-Based Reranking and Faithfulness Control

Introducing 5ting, a system for multi-turn RAG that utilizes BGE-M3 retrieval, FAISS indexing, and LLM-based reranking to minimize context drift and hallucinations.

AI/ML arXiv cs.AI

Four Types of LLM Reliance and Their Predictors Among Undergraduate Writers: A Mixed-Methods Study at a Minority-Serving R1 University

A mixed-methods study identifying four types of LLM reliance among undergraduate writers: Strategic, Instrumental, Dialogic, and Dependent.

AI/ML arXiv cs.AI

X-Mind: Efficient Visual Chain-of-Thought via Predictive World Model for End-to-End Driving

X-Mind introduces a Visual Chain-of-Thought framework for autonomous driving, using a predictive world model and compact visual sketches to reduce latency and improve reasoning.

AI/ML arXiv cs.AI

Majority Vote Silences Minority Values: Annotator Disagreement at the Hate/Offensive Boundary in HateXplain

Research demonstrating that majority vote labeling in hate speech datasets silences minority views and creates false confidence in AI models at the hate/offensive boundary.

AI/ML arXiv cs.AI

Brownian Bridge Diffusion-Based Joint Channel Estimation and Data Detection for Jamming-Resilient Receivers

BBD-JCED is a new framework using Brownian bridge diffusion for joint channel estimation and data detection to make wireless receivers more resilient to jamming.

AI/ML arXiv cs.AI

BREIT: A Framework for Brain Stroke Reconstruction using Multi-Frequency 3D EIT

BREIT is a modular framework for 3D multi-frequency electrical impedance tomography (MF-EIT) for non-invasive brain stroke reconstruction.