Algorithm Innovation and New Architecture Emergence Breaking Through Reinforcement Learning Performance Stagnation and Attention Operation Bottlenecks
Advanced Problem-Solving Capabilities for Complex Real-World Challenges Through Self-Play Learning and Brain Model-Based Reasoning

This week''s META-X AI paper review covers advances in novel architectures, video generation, reinforcement learning, and multimodal reasoning.

The Dragon Hatchling: The Missing Link between the Transformer and Models of the Brain proposes a new LLM architecture (BDH) inspired by the brain''s operation, overcoming existing transformer limitations. Based on locally interacting neuron network structures with spiking neurons and Hebbian learning for working memory implementation, BDH achieves performance equivalent to GPT-2 with similar parameter counts while showing enhanced interpretability as relevant synapses strengthen when processing specific concepts.

LongLive: Real-time Interactive Long Video Generation addresses speed and quality degradation in real-time interactive long video generation through KV-recache mechanism for natural scene transitions, streaming tuning, and short-window attention for computational efficiency. Achieves 20.7 FPS generation speed for up to 4-minute high-quality videos on single GPU.

Additional papers reviewed cover: reinforcement learning algorithmic improvements addressing reward shaping challenges and exploration-exploitation tradeoffs in sparse reward environments; efficient attention approximation methods enabling practical deployment of 100K+ context window models; self-play learning systems achieving expert-level performance in complex strategy games without human demonstration data; and multimodal reasoning advances enabling more coherent integration of visual, textual, and structured data for complex analytical tasks requiring cross-modal inference.