AI/ML arXiv cs.AI

A Neuro-Symbolic Approach for Probabilistic Reasoning on Graph Data

A new neuro-symbolic framework integrates Graph Neural Networks (GNNs) with Relational Bayesian Networks (RBNs) to combine deep learning with probabilistic reasoning.

AI/ML arXiv cs.AI

Human-Aligned Procedural Level Generation Reinforcement Learning via Text-Level-Sketch Shared Representation

VIPCGRL is a deep reinforcement learning framework that uses text, levels, and sketches to create human-aligned procedural content for games.

AI/ML arXiv cs.AI

RL-Struct: A Lightweight Reinforcement Learning Framework for Reliable Structured Output in LLMs

RL-Struct is a lightweight framework using GRPO to align LLMs with structural constraints for reliable JSON output, reducing VRAM usage compared to PPO.

AI/ML arXiv cs.AI

Mechanistic Interpretability of Cognitive Complexity in LLMs via Linear Probing using Bloom's Taxonomy

Study finds that cognitive complexity levels based on Bloom's Taxonomy are linearly separable in the internal representations of LLMs.

AI/ML arXiv cs.AI

The AI Fiction Paradox

The 'AI-Fiction Paradox' explains why autoregressive models struggle with long-form fiction due to failures in narrative causation and emotional architecture.

AI/ML arXiv cs.AI

SciVisAgentBench: A Benchmark for Evaluating Scientific Data Analysis and Visualization Agents

SciVisAgentBench is a new comprehensive benchmark for evaluating AI agents that translate natural language into scientific visualization tasks.

AI/ML arXiv cs.AI

FAIR_XAI: Improving Multimodal Foundation Model Fairness via Explainability for Wellbeing Assessment

Investigation into VLM fairness reveals significant racial and gender biases in wellbeing assessments, showing a gap between transparency and equitable outcomes.

Other Hacker News

When can a power company take your land for data center infrastructure?

A discussion on the legal aspects of power companies exercising eminent domain for data center infrastructure.

AI/ML arXiv cs.AI

LLM-Powered Agentic AI for 5G/6G Networks: A Tutorial and Survey on Architectures, Protocols, and Standardization

A tutorial and survey on integrating LLM-powered agentic AI into 5G/6G network architectures and protocols.

AI/ML arXiv cs.AI

JoyNexus: Service-Oriented Multi-Tenant Post-Training for VLA Models

Introduction of JoyNexus, a multi-tenant service for VLA model post-training that decouples training, inference, and environment services to improve GPU efficiency.

AI/ML arXiv cs.AI

HCIG: A Hierarchical Cross-Modal Incongruity Graph Network for Multimodal Sarcasm and Cyberbullying Detection

Proposed HCIG and GCCN frameworks for multimodal sarcasm and cyberbullying detection using hierarchical cross-modal incongruity graph networks.

AI/ML arXiv cs.AI

DADiff: Diffusion-Driven Cross-Domain Policy Adaptation for Reinforcement Learning

Introduction of DADiff, a diffusion-based framework for cross-domain policy adaptation in reinforcement learning to handle dynamics mismatch.

AI/ML arXiv cs.AI

Understanding Reasoning from Pretraining to Post-Training

Research using chess as a testbed to quantify how pretraining choices affect the returns to RL compute and the nature of reasoning improvements.

AI/ML arXiv cs.AI

A Methodology for Auditable Trustworthiness Levels in AI Lifecycle Governance

A lightweight methodology for creating auditable trustworthiness levels and lifecycle governance in AI systems.

AI/ML arXiv cs.AI

ToolSciVer: Multimodal Scientific Claim Verification with Visual Tool Augmented Reinforcement Learning

ToolSciVer, a framework for multimodal scientific claim verification that uses visual tools and GRPO for enhanced reasoning on scientific visuals.

AI/ML arXiv cs.AI

When Do Multi-Agent Systems Help? An Information Bottleneck Perspective

Theoretical analysis of multi-agent systems versus single-agent systems from an information bottleneck perspective, exploring the trade-off between context reduction and information loss.

AI/ML arXiv cs.AI

An Exam for Active Observers

Introduction of ActiveVision, a benchmark demonstrating that current MLLMs struggle significantly with active visual observation tasks compared to humans.

Other Hacker News

US gas prices hit an average of $4 a gallon again

US gas prices have returned to an average of $4 per gallon.

AI/ML arXiv cs.AI

Robustness of Reinforcement Learning-Based Congestion Management in Low-Voltage Grids

Researchers propose a framework combining random-forest pre-classifiers with actor-critic controllers to manage congestion in low-voltage grids more robustly.

AI/ML arXiv cs.AI

DPNeXt: A Lightweight Multi-Scale Feature Fusion Framework for Efficient ViT-Based Multi-Task Dense Prediction

Introduction of DPNeXt, a lightweight multi-scale feature fusion framework for efficient Vision Transformer-based multi-task dense prediction in robotics.