• 2508, Bytedance, DuPO: Enabling Reliable LLM Self-Verification via Dual Preference Optimization
• 2507, Anthropic, Subliminal Learning: Language Models Transmit Behavioral Traits via Hidden Signals in Data
• 2505, Google, AlphaEvolve: A coding agent for scientific and algorithmic discovery
• 2503, Anthropic, Circuit Tracing: Revealing Computational Graphs in Language Models
• 2503, Anthropic, Auditing language models for hidden objectives
• 2502, DeepSeek, Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention
• Sora : Video generation models as world simulators