Auxiliary-loss-free load balancing (ALF-LB) achieves promising performance in Mixture-of-Experts large language model (MoE-LLM) pre-training as it decouples load balancing from the main objective. However, we observe an unexplored phenomenon: routing imbalance is concentrated in shallow MoE layers. We attribute this phenomenon to the interaction between representation homogeneity, which makes shallow-layer routing decisions highly sensitive, and expert-bias staleness, which applies stale load corrections to the current router state. To break the cycle, we propose Dual-Bias Load Balancing (DLB), which augments the standard global expert bias with an intra-step micro-bias residual. DLB introduces the effective bias 𝐛_eff=𝐛_g+𝛥𝐛, where 𝐛_g is updated once per gradient step and 𝛥𝐛 is updated across micro-batches within the same step. DLB does not add forward or backward passes and integrates naturally into existing large-scale MoE training pipelines. Experiments on large-scale MoE pre-training show that DLB mitigates expert-bias staleness, alleviates shallow-layer imbalance, and yields lower test perplexity and stronger average downstream performance.