All Articles
18326 articles total
Veriphi: Attack-Guided Neural Network Verification with Dataset-Dependent Training Methods
The Veriphi system provides GPU-accelerated neural network verification, demonstrating that the effectiveness of training methods is dependent on the dataset used.
What Does the Weight Norm Control in Grokking? Logit-Scale Mediation under Cross-Entropy
Study reveals that the 'grokking' phenomenon in neural networks is primarily driven by logit scale rather than weight norm, contrary to previous assumptions.
Structured Representation Learning with Locally Linear Embeddings and Adaptive Feature Fusion
A new reinforcement learning framework inspired by neuroscience uses locally linear embeddings and adaptive gating to improve learning efficiency.
MagpieTTS-LF: Inference-Time Long-Form Speech Generation Without Training on Long-Form data
MagpieTTS-LF introduces an inference-time approach to generate coherent long-form speech without requiring retraining on long-form data.
SFT Overtraining Predicts Rank Inversion via Entropy Collapse Under RLVR
Research indicates that overtraining during SFT can lead to rank inversion and entropy collapse when using RLVR, specifically affecting models like Qwen2.5-Coder.
Neural Phase Correlation
Introduction of a learned generalization of phase correlation that enables the discovery of transformations between observations in the Fourier domain.
PSyGenTAB: A Privacy-Preserving Framework for Synthetic Clinical Tabular Data Generation via Constrained Optimization
PSyGenTAB is a privacy-preserving framework for generating synthetic clinical tabular data using constrained optimization to balance privacy and utility.
As You Wish: Mission Planning with Formal Verification using LLMs in Precision Agriculture
A mission planning system for precision agriculture uses LLMs and linear temporal logic (LTL) to formally verify natural language mission specifications.
Smashed Toilet Phone Web Server
A project involving the creation of a web server hosted on a 'Smashed Toilet Phone'.
From Specification to Execution: AI Assisted Scientific Workflow Management
Proposes an AI-assisted scientific workflow management approach using specification-driven generation and an LLM-based debugging agent integrated with Pegasus.
CAOA -- Completion-Assisted Object-CAD Alignment
Introduces CAOA, a method for aligning CAD models to indoor RGB-D scans using point cloud completion and symmetry-aware relative pose estimation.
Self-CTRL: Self-Consistency Training with Reinforcement Learning
Introduces Self-CTRL, a reinforcement learning method to improve consistency between a language model's self-explanations and its actual behavior.
SafeClawBench: Separating Semantic, Audit-Evidence, and Sandbox Harm in Tool-Using LLM Agents
Presents SafeClawBench, a staged benchmark for evaluating security failures in tool-using LLM agents across various attack vectors.
Guava: An Effective and Universal Harness for Embodied Manipulation
Proposes Guava, a harness framework for embodied tool use that enables compact open-source models to perform complex manipulation tasks.
Redact or Keep? A Fully Local AI Cascade for Educational Dialogue De-Identification
Describes a fully local AI cascade framework for de-identifying PII in educational dialogues, outperforming commercial APIs on a single laptop.
RankGraph-2: Lifecycle Co-Design for Billion-Node Graph Learning in Recommendation
Details RankGraph-2, Meta's lifecycle co-design framework for billion-node graph learning in recommendation systems to optimize retrieval and serving.
LLMZero: Discovering Adaptive Training Strategies for RL Post-Training via LLM Agents
Introduces LLMZero, a system using LLM agents to automatically discover adaptive training strategies for reinforcement learning post-training.
Learning-Based Decision Making for Combustion Phasing Control in Multi-Fuel CI Engines with Latent Fuel Reactivity Estimation
Proposes a GRU-guided RL framework for combustion phasing control in multi-fuel engines, managing latent fuel reactivity estimation.
Deep Learning-Driven Inverse Design of Doherty Power Amplifiers Using Pixelated Combiners and Dual-State Impedance Synthesis
Utilizes CNNs and genetic algorithms for the inverse design of high-efficiency Doherty power amplifiers using pixelated combiners.
Deep-Learning-Based Pixelated Microwave Filter Design and Characterization using Electro-Optical Electric-Field Measurements
Applies deep learning and genetic algorithms to automate the design and characterization of pixelated microwave filters.