New Data Pricing Paradigm: Token-Level Quality Assessment Reshapes LLM Training Data Valuation
A latest arXiv paper proposes a utility-based dynamic data valuation framework that starts from token-level information …
📂 Research
Latest articles in Research
Accelerating Frequency-Domain Diffusion Model Inference: A Deep Dive into the E²-CRF Caching Strategy
A latest arXiv paper proposes the E²-CRF method, leveraging two key structural properties — spectral localization and mi…
Aviation Fuel Pump Simulation Benchmark Launched, Solving AI Fault Diagnosis Data Challenge
A research team has released a physics-informed, high-fidelity co-simulation system for aircraft main fuel pumps, provid…
Multi-Fidelity AI Model Accurately Predicts Wind Loads on Container Ships in Ports
A new study proposes a multi-fidelity surrogate model approach that leverages AI to fuse simulation data of varying accu…
AI Fuses Satellite and Reanalysis Data to Map Reliable PM2.5 Across Africa
A new arXiv study proposes a PM2.5 fusion system combining LightGBM, spatial cross-validation, and conformal prediction.…
Stochastic KV Routing: A New Paradigm for Cache Sharing Across the Depth Dimension
A latest arXiv paper proposes "Stochastic KV Routing" technology, enabling adaptive KV cache sharing across the depth di…
AutoCompress: Efficient Transformer Compression Through Critical Layer Isolation
A research team has proposed the AutoCompress method, discovering that Layer 0 in small Transformers carries over 60 tim…
Parameter Efficient ≠ Memory Efficient: Rethinking On-Device LLM Fine-Tuning
New research challenges the widespread assumption that "parameter efficient means memory efficient," revealing that whil…
Novel Neural Plasticity: A Single Experience Can Rewire the Brain
New neuroscience research has uncovered an entirely new neural plasticity mechanism that allows the brain to complete sy…
The 'Spectral Lifecycle' of Transformer Training Systematically Revealed for the First Time
A new study systematically tracks the singular value spectrum evolution of weight matrices during Transformer pre-traini…
KARL Framework: Teaching Large Models Knowledge Boundary Awareness to Reduce Hallucinations
Researchers propose the KARL framework, a knowledge-boundary-aware reinforcement learning approach that enables large la…
BiTA Model: Bidirectional Temporal Graph Networks Revolutionize Cybersecurity Early Warning
Researchers propose the BiTA framework, which integrates Bidirectional GRU with a Transformer Aggregator into Temporal G…