Can Multimodal Learning Improve Fake Profile Detection?
A new arXiv preprint asks whether combining multiple data types can beat strong metadata and graph baselines for spotting fake accounts.
Research Digest
Preprints and papers, rewritten in plain English.
GLOW: Ward's Method for Neuroimaging General Linear Models
A new approach, GLOW, applies Ward's clustering to general linear models in neuroimaging, potentially avoiding the dilution or splitting of effects common in region-based analysis.
PeekKV: First Token Identifies Key Attention Heads for Cache Compression
A new arXiv paper proposes using the first content token's attention patterns to decide which heads to keep in the KV cache, targeting the memory bottleneck in long-context LLM serving.
Machine Learning Speeds Up Inner Sphere Tree Construction
A new arXiv preprint proposes using machine learning to accelerate the building of inner sphere trees for collision detection.
Group-Aware Comparison of Classical and Deep Learning for Sensor HAR
A new arXiv study examines how sensor-related factors influence reported performance in human activity recognition, comparing classical machine learning and deep learning methods.
DS-Net Proposes Dynamic Splitting for Edge-Cloud Inference
A new arXiv paper introduces a dynamic split network approach to address limitations in existing edge-cloud collaborative inference methods.
SPADE-KD: Peer-to-Peer Learning Without a Central Server
A new distributed deep learning method removes the central aggregator, avoiding single points of failure and bottlenecks.
Factorization-Guided Method for Diffusion-Drafted Speculative Decoding
A new arXiv paper proposes factorization-guided training and inference to improve diffusion-based drafters in speculative decoding.
TWIN-MoE Offers Two-Stage Imputation for Missing Modalities
A new Mixture-of-Experts approach tackles incomplete multimodal data by imputing missing inputs within each expert in two stages.
Physics-Aware Pretraining for Unlabeled Raman Spectra
A masked-autoencoder framework lets Raman spectroscopy models learn from large volumes of unlabeled data without manual labels.
Bookmarking Causal Chains to Fix Credit Assignment in Delayed-Reward RL
A new arXiv proposal uses visual causal-chain bookmarks and retrieval-weighted credit assignment to give delayed rewards to the actions that actually caused them.
LatLM: Deferred Commitment for Line-Parallel Decoding
A new framework relaxes the rigid commitment schedule of line-parallel language models to accommodate variable cross-line dependencies.
Auditing PINNs for Misspecified Equation Coefficients
A new preprint proposes a closed-form audit to detect when physics-informed neural networks are trained with incorrect equation coefficients, despite low training loss.
Audit Questions Test Labels in Graph Anomaly Detection Benchmark
A new audit of the UB-GOLD benchmark suggests that reported results may not reflect truly label-free graph anomaly detection, because test labels could have been used for model selection.
Early-Exit Vision Transformers Tailored for Heterogeneous Federated Learning
A new approach adapts early-exit ViT inference to client-specific data difficulty and compute budgets in federated learning.
Bridging the Execution Gap in Native NVFP4 Reinforcement Learning
A new diagnostic tool called miniTE targets the instability that arises when LLM reinforcement learning uses separate training and sampling engines with NVFP4 precision.
Stochastic Routing and Reinforcement Fusion for Multi-Model Reasoning
A new arXiv paper proposes combining stochastic routing with reinforcement learning to improve dynamic model selection in multimodal systems.
Generative Drug Design Models Often Overlook Selectivity
A new arXiv paper highlights how generative models for structure-based drug design may neglect selectivity, a key factor in reducing side effects.