Zheng et al.

The LLM-as-judge pattern uses a more powerful language model to automatically evaluate another model's responses in production, addressing the scalability limits of human review and the semantic blind…

A new masterclass proposes a lattice-based, zero-knowledge cryptographic blueprint with a 106ms verification footprint to secure AI pipelines against quantum-era gradient inversion and collusion, addr…

A new training method called replacement-aware training produces sparse auto-encoders (SAEs) that retain language capabilities when used in a full replacement model of Gemma-2-2B, unlike standard SAEs…