Keep up with tech in 5 minutes

LLM-as-a-Judge is a common framework in which an LLM is tasked with reviewing an input, assigning it a score, and then selecting the highest-probability token. This framework frequently produces ties because the scoring granularity is too coarse. The authors introduce LLM-as-a-Verifier as an alternative framework that achieves higher decision accuracy and no ties by scaling the number of repeated verifications, the granularity of scoring tokens, and the decomposition of evaluation criteria.

LinkedIn is cutting roughly 5% of its staff.

Performance metrics like AI token usage or project complexity can become "perverse incentives" that encourage engineers to prioritize visible activity over actual productivity. Examples like Goodhart's Law and the Cobra Effect show that companies must carefully align incentive structures to focus on meaningful outcomes rather than easily gamed inputs.