Llama

The researchers present ComPO, a new alignment paradigm that uses comparison oracles to extract directional information from preference pairs. This method effectively mitigates the risks of likelihood displacement and model verbosity seen in traditional direct alignment methods.

The AI industry is seeing a surge in 'model routers' designed to intelligently distribute LLM queries to the most cost-effective and capable models. Companies including Cursor, Ramp, and Meta are developing these systems to optimize performance and reduce the high costs associated with frontier-grade models.

This paper presents a method for converting pretrained transformers into linear-time architectures by focusing on the efficiency of state update designs. By analyzing softmax attention through a first-order approximation, the authors prove that delta-style updates are superior for post hoc linearization.

A critical vulnerability in Ollama allows unauthenticated attackers to trigger an out-of-bounds heap read via malicious GGUF files. This exploit can expose sensitive information like user messages and system prompts by leaking them into newly created model files.