General
- AI and Memory Wall. Argues that AI scaling transitioned from compute-limited to memory-limited.
Research Insights
- ArcMemo: Abstract Reasoning Composition with Lifelong LLM Memory. Learning in reasoning traces is distilled into reusable language that can be saved and retrieved.
- MTQA:Matrix of Thought for Enhanced Reasoning in Complex Question Answering. Deploys inference compute along several parallel though strategies.
- Nvidia argues: Small Language Models are the Future of Agentic AI.
- Talk Isn’t Always Cheap: Understanding Failure Modes in Multi-Agent Debate. Shows how weaker agents can disrupt multi-agent work.
- Emergent Hierarchical Reasoning in LLMs through Reinforcement Learning. RL leads to a two-layer hierarchy without this being explicitly designed.
LLM
Image Synthesis
- ByteDance Seedream 4.0 (examples).
Video
World Synthesis