Zhipu AI MLE Interview Questions (2026)

Free questions with model answers (3)

• #48 Describe mixture-of-experts and how it grows model capacity without growing per-token compute.

  A sparse MoE router sends each token to a small subset of expert MLPs.

• #89 Explain RLVR and where it beats a learned reward model.

  Verifiable, not learned: RLVR (reinforcement learning with verifiable rewards) replaces a learned reward model with a deterministic checker: a math answer matches the gold solution, code passes held-out unit tests, a proof type-checks.

• #296 Implement a top-k MoE router, then contrast auxiliary-loss load balancing with a loss-free approach.

  Compute router logits, select top-k experts per token, normalize selected weights, dispatch tokens, run experts, and scatter-add outputs.