Cybersecurity arXiv cs.AI

FakeIDet3-DB: Refining Digital Attacks and Patch Extraction for Secure ID Benchmarking

Introduction of FakeIDet3-DB, a database of digital manipulations on real IDs and the PACE algorithm for privacy-aware patch extraction.

AI/ML arXiv cs.AI

FPSGen: Flexible Point Cloud Scene Generation with BEV-Supported Transport Flows

FPSGen provides a flexible framework for point cloud scene generation using BEV-supported transport flows to improve outdoor scene completion.

AI/ML arXiv cs.AI

The Art of Not Forgetting A Local Learning Architecture for Continual Learning

Introduces CMP (Cognitive Memory Primitive), a continual-learning architecture using sparse relational codes and local updates to mitigate catastrophic forgetting in language modeling.

AI/ML arXiv cs.AI

Shared Symbolic Backbones for Physically Consistent Multi-Output Symbolic Regression

Presents a neuro-evolutionary symbolic regression method that uses shared symbolic backbones to ensure physical consistency across multiple coupled outputs.

AI/ML arXiv cs.AI

AgentGFM: A Graph Foundation Model with Node-Agent Information-Flow Control

Proposes AgentGFM, a Graph Foundation Model where each node acts as an agent to autonomously control information flow for better transferability across diverse graph topologies.

AI/ML arXiv cs.AI

A Persona-based Rate Action Index

Develops a persona-based index using generative AI to predict FOMC interest rate decisions by simulating the behavior of specific Federal Reserve members.

AI/ML arXiv cs.AI

ServerlessT2I: Efficient Text-to-Image Workflow Serving on a Serverless Platform

Introduces ServerlessT2I, a system that decomposes text-to-image workflows into independent model functions to optimize GPU utilization and scaling on serverless platforms.

Cybersecurity arXiv cs.AI

Recover, Decode, Reguard: Guard-Agnostic Defense Amplification againstEncoded VLM Jailbreaks

Analyzes the effectiveness of 'recover-and-decode' amplifiers to defend Vision-Language Models (VLMs) against encoded jailbreaks, revealing significant safety-utility trade-offs.

AI/ML arXiv cs.AI

Classification of Disease from Lungs X-ray Images using VGG16, VGG19 and ResNet50 Models

Evaluates the performance of VGG16, VGG19, and ResNet50 models for classifying lung diseases from X-ray images, finding ResNet-50 to be the most effective.

AI/ML arXiv cs.AI

One Run Is Not an Idea: The Implementation Lottery in Automated Research

Identifies the 'implementation lottery' in automated research, where conclusions are often based on a single implementation rather than the underlying idea's reliability.

AI/ML arXiv cs.AI

A Physics-Informed Framework for PID Tuning of Chemical Processes Using Large Language Model Agents

Presents a physics-informed framework using LLM agents and PI-GRPO for the automated PID tuning of chemical processes, significantly improving first-attempt success rates.

AI/ML arXiv cs.AI

Decoupled Visual Processing: Efficient Multimodal Adaptation via Modality-Specific Transformer Substitution

Proposes Decoupled Visual Processing (DVP), a parameter-efficient training framework that replaces upper decoder layers with a specialized visual transformer block for MLLMs.

Software Engineering Hacker News

Obfuscation (Part II): Diamond IO

A discussion thread on Hacker News focusing on the concept of 'Diamond IO' in the context of obfuscation.

AI/ML arXiv cs.AI

Reinforcement Learning on Cost-Constrained Quadrupedal Hardware

Researchers use a forward model of actuator delay and a time-aware neural network to close the sim-to-real gap for low-cost quadrupedal robots.

AI/ML arXiv cs.AI

Mergeable Model-Side Aggregation States for Long-Context Language Models

Introduces a model-side aggregation interface using HyperLogLog sketch states to improve non-additive aggregation in long-context LLMs.

AI/ML arXiv cs.AI

ForgetBench: Benchmarking Forgetting Dynamics of Long-Term Parametric Memory in Language Models

ForgetBench is a new benchmark designed to systematically characterize the forgetting dynamics of long-term parametric memory in LLMs.

Hardware/Chips arXiv cs.AI

PUDA: An AI-Native Hardware Harness for Self-Driving Laboratories

PUDA is an AI-native hardware harness providing a command-line runtime for agentic self-driving laboratories, separating orchestration from physical operation.

AI/ML arXiv cs.AI

Audio-Anchored Fusion of Multi-Ratio DiT Reconstruction Residuals for Cross-Domain Audio Deepfake Detection

A new audio deepfake detection method using Diffusion Transformer (DiT) reconstruction residuals and WavLM auditory representations.

Hardware/Chips arXiv cs.AI

LLMET: Enabling Cross-Layer Evaluation of Emerging M3D Memories for Energy-Efficient LLM Serving

LLMET is a cross-layer simulation framework used to demonstrate that scaling on-chip M3D memory can significantly reduce energy consumption in LLM serving.

AI/ML arXiv cs.AI

Collaborative Weighting with Pessimistic Critic for Mitigating Overestimation in Off-Policy Reinforcement Learning

The Collaborative Weighting Actor-Critic (CWAC) framework mitigates overestimation bias in off-policy RL by accounting for predictive uncertainty.