Attention Residuals: A Comprehensive Understanding
This paper addresses a fundamental problem in training deep transformer models: uncontrolled hidden-state magnitude growth as model depth increases. The auth...
Recent posts
The Web Is Not a Neutral Environment for Agents
Browser agents are getting better fast, but the web is full of things that try to steer behavior. If that already works on humans, why would agents be immune?
Modeling Long User Histories for Ads Ranking
How ads ranking systems went from aggregated feature counts to retrieve-and-compress architectures that handle 10,000+ user events under millisecond latency ...
The Evolution of Reward Hacking and Jailbreak Research in AI
From specification gaming in classical RL to deceptive alignment and jailbreaks in LLMs—a survey of how reward hacking has become a central challenge in AI s...
Reasoning in Large Language Models: A Research-Centric Overview
Reasoning in Large Language Models: A Research-Centric Overview