LLM Agents Interview Questions #25 - The Diversity Scaling Trap

Increasing sample count without increasing trajectory entropy breaks the core assumption behind self-consistency gains.

You’re in a Senior AI Engineer interview at DeepMind. The interviewer sets a trap:

“You’ve implemented self-consistency (majority voting) to improve an agent’s math reasoning. But scaling the sample size from 10 to 40 yields zero performance gain. What is silently killing your scaling laws?”

90% of candidates walk right into it.

Most candidates immediately…