AI/ML arXiv cs.AI

A robust association between LLM use and scientific productivity: Assessing stopping-time selection

A study analyzing the association between LLM use and scientific productivity, countering arguments that results were based on stopping-time selection artifacts.

AI/ML arXiv cs.AI

RAID: Towards Robust AI-Generated Image Detection with Bit-Reversed Images

Introduction of RAID, a fast and robust AI-generated image detection pipeline using bit-reversed images and bit-planes.

AI/ML arXiv cs.AI

PARALLEL: A Prefrontal-Aligned Reinforcement inspired Approach for Language-Model Learning under Explicit Limits

Proposal of PARALLEL, a reinforcement-inspired approach that optimizes LLM adaptation by adjusting update intensity per sample.

AI/ML arXiv cs.AI

Adjudicated Captioning: Multi-Agent Alignment Scoring and Consensus-Distilled Beam Arbitration for Strict Zero-Shot Image Captioning

Introduction of Adjudicated Captioning, a multi-agent framework for zero-shot image captioning that improves grounding feedback during inference.

AI/ML arXiv cs.AI

Point2Radio: A Foundation Model for Cross-Scene Radio Fields from Material-Aware Point Clouds

Point2Radio is a foundation model that predicts radio propagation fields from material-aware point clouds without explicit path tracing.

AI/ML arXiv cs.AI

Auto-JEPA: A Latent World Model of Continuous Intent for End-to-End Autonomous Driving

Auto-JEPA is an action-oriented latent world model for autonomous driving that focuses on planning-relevant features rather than dense future prediction.

AI/ML arXiv cs.AI

Improving scDiffusion with Sparsity-Biased Classifier-Free Guidance

Research on SB-CFG, a sparsity-biased classifier-free guidance strategy to improve synthetic single-cell RNA sequencing data generation.

AI/ML arXiv cs.AI

RareSense: Rarity-Aware Similarity Search for Anomaly Retrieval in Transactional Data

RareSense is a new similarity search framework for sparse transactional data that uses rare itemsets and association rules to detect anomalies more effectively than standard IDF weighting.

AI/ML arXiv cs.AI

To Add Is Machine, To Delete Is Human: Measuring and Mitigating Deletion Avoidance in LLM Code Editing

Researchers identify 'deletion avoidance' in LLMs, where models fail to remove unnecessary code during edits, and introduce the CanItDelete benchmark to measure this tendency.

AI/ML arXiv cs.AI

Human-LLM Collaborative Inductive Coding for Conceptualizing K-12 Educator AI Use

A study describes a human-LLM collaborative pipeline for qualitative coding of large interaction corpora to develop hierarchical codebooks while maintaining human interpretive authority.

AI/ML arXiv cs.AI

Agreement Is Not Quality: Blind Expert Verification of Human and LLM Qualitative Coding When Human Consensus Is Not Ground Truth

This research challenges the use of human agreement as the gold standard for LLM qualitative coding, showing that blind expert verification can favor LLM interpretations over human consensus.

AI/ML arXiv cs.AI

TORUS: A Test of Rendering-Understanding Self-Coherence for Unified Audio Models

TORUS is introduced as a self-coherence test for unified audio models to determine if their audio generation and understanding capabilities are aligned.

Other arXiv cs.AI

Design Concept: Scaffolding Geopolitical Reflection Among Tech Workers

A speculative HCI design proposal for an AI-enabled narrative system aimed at encouraging geopolitical reflection among technology workers.

AI/ML arXiv cs.AI

Gated Q-learning: Add Off-Policy Bias to Taste

Gated Q-learning is proposed as a framework to manage off-policy bias in reinforcement learning by interpolating between eliminating bias and ignoring it via a gating mechanism.

AI/ML arXiv cs.AI

FairFund-Bench: Evaluating Distributive Bias in LLM Resource Allocation

FairFund-Bench is a new benchmark for evaluating distributive bias in LLM resource allocation, revealing that audit formats significantly influence bias detection.

Cybersecurity arXiv cs.AI

DiffAttack: Evasion Attacks Against Face Recognition via Latent Diffusion Models

DiffAttack uses latent diffusion models to create high-quality adversarial images that can evade facial recognition systems with high success rates.

Cybersecurity arXiv cs.AI

Retrieval-Driven Training-Free AI-Generated Video Attribution

A training-free attribution paradigm is introduced to identify the generative source of AI-generated videos using a generative fingerprint-based pipeline.

Other Hacker News

Less Coffee, Better Sleep

A discussion on lifestyle choices regarding caffeine consumption and its impact on sleep quality.

Cybersecurity Hacker News

What DMARC Protects You From, and What It Does Not

An explanation of the DMARC protocol and its specific capabilities and limitations in email security.

Other Hacker News

MPs demand answers on Fujitsu's inclusion in lucrative frameworks

Political scrutiny in the UK regarding Fujitsu's inclusion in government frameworks.