Internalizing Outcome Supervision for Enhanced RL Reasoning
Discover a new reinforcement learning paradigm that internalizes outcome supervision into process supervision to boost AI reasoning and learning efficiency...
MACS: Boosting Multimodal MoE Inference Efficiency
Discover how MACS improves multimodal MoE inference with modality-aware scaling, enhancing AI efficiency and accuracy in processing text and images.
Enhancing Unlearnable Examples for Pretraining-Finetuning AI
Explore how Shallow Semantic Camouflage improves unlearnable examples' effectiveness in pretraining-finetuning paradigms to protect AI model privacy.
Overcoming Structural Instability in Feature Composition
Explore how geometric frameworks and Sparse Autoencoders address feature composition instability in transformer models for better AI control.
Adaptive Token Routing Boosts Transformer Efficiency
Discover how Token-Selective Attention enables adaptive computation in transformers, reducing token-layer operations by up to 23% with minimal overhead.