AI/ML arXiv cs.AI

H3Former: Hypergraph-based Semantic-Aware Aggregation via Hyperbolic Hierarchical Contrastive Loss for Fine-Grained Visual Classification

H3Former is a novel token-to-region framework using hypergraph aggregation and hyperbolic contrastive loss for fine-grained visual classification.

Cybersecurity Hacker News

We Put an L7 Firewall in the Kernel

A discussion about implementing a Layer 7 firewall directly within the operating system kernel for improved performance and security.

Other Hacker News

An Infuriating Goodbye to Photoshop

A user's account of their frustration and eventual decision to stop using Adobe Photoshop.

AI/ML arXiv cs.AI

HiPO: Hierarchical Preference Optimization for Adaptive Reasoning in LLMs

Introduces HiPO, a Hierarchical Preference Optimization method that improves LLM reasoning by applying DPO to specific segments of a response.

AI/ML arXiv cs.AI

Heterogeneous Information-Bottleneck Coordination Graphs for Multi-Agent Reinforcement Learning

Proposes HIBCG, a framework for multi-agent reinforcement learning that uses information-bottleneck coordination graphs to optimize communication topology.

AI/ML arXiv cs.AI

Explaining is Harder Than Predicting Alone: Evaluating Concept-based Explanations of MLLMs as ICL Visual Classifiers

Evaluates whether multimodal LLMs can provide machine-verifiable explanations for their visual classifications, finding that formal explaining is harder than predicting.

AI/ML arXiv cs.AI

Latent Reward Steering: An Adaptive Inference-Time Framework that Implicitly Promotes Cognitive Behaviors in Reasoning LLMs

Presents Latent Reward Steering (LRS), an inference-time framework that uses sparse-autoencoders to implicitly promote correct cognitive behaviors in reasoning LLMs.

AI/ML arXiv cs.AI

Projection Methods for Operator Learning and Universal Approximation

Provides a theoretical framework for operator learning in Banach spaces using projection methods and universal approximation theorems.

AI/ML arXiv cs.AI

Multi-Attribute Steering of Language Models via Targeted Intervention

Introduces MAT-Steer, a framework for steering LLMs across multiple attributes simultaneously (e.g., helpfulness and toxicity) using orthogonal steering vectors.

AI/ML arXiv cs.AI

Transformer-Empowered Actor-Critic Reinforcement Learning for Sequence-Aware Service Function Chain Partitioning

Proposes a Transformer-based actor-critic RL framework to optimize the partitioning of Service Function Chains in 6G networks.

AI/ML arXiv cs.AI

M4V: Multimodal Mamba for Efficient Text-to-Video Generation

Introduces M4V, a multimodal Mamba-based framework for text-to-video generation that reduces computational complexity compared to Transformers.

Other Hacker News

A voxel Tokyo in real Japan time – ride the Yamanote line and study Japanese

A project visualizing Tokyo in real-time voxels, allowing users to virtually ride the Yamanote line and study Japanese.

AI/ML arXiv cs.AI

QAgent: An LLM-based Multi-Agent System for Autonomous OpenQASM programming

Introduction of QAgent, an autonomous multi-agent framework for end-to-end OpenQASM code generation for quantum circuits.

Cybersecurity arXiv cs.AI

Beyond Embeddings: Interpretable Feature Extraction for Binary Code Similarity

A new method for binary code similarity detection that uses LLM-based agents to generate interpretable, human-readable features for reverse engineering.

AI/ML arXiv cs.AI

Leveraging Multi-Agent System (MAS) and Fine-Tuned Small Language Models (SLMs) for Automated Telecom Network Troubleshooting

A multi-agent system combining LLMs and fine-tuned small language models (SLMs) to automate telecom network troubleshooting.

AI/ML arXiv cs.AI

Improving Language Agents through BREW: Bootstrapping expeRientially-learned Environmental knoWledge

BREW is a framework that allows LLM agents to learn from past interaction trajectories by distilling them into a structured knowledge base of reusable 'recipes'.

AI/ML arXiv cs.AI

Programming over Thinking: Efficient and Robust Multi-Constraint Planning

The Scalable COde Planning Engine (SCOPE) separates reasoning from execution to create reusable solver functions for efficient multi-constraint planning.

AI/ML arXiv cs.AI

PACE: A Personalized Adaptive Curriculum Engine for 9-1-1 Call-taker Training

PACE is a personalized adaptive curriculum engine designed to accelerate 9-1-1 call-taker training using a co-pilot system.

AI/ML arXiv cs.AI

A Self-Evolving Agentic Framework for Metasurface Inverse Design

A self-evolving agentic framework that uses a coding agent and skill files to automate metasurface inverse design in optics.

AI/ML arXiv cs.AI

Rectification Difficulty and Optimal Sample Allocation in LLM-Augmented Surveys

A framework for optimizing the allocation of human respondents in LLM-augmented surveys by predicting 'rectification difficulty'.