Resting Neurons, Active Insights: Robustifying Activation Sparsity in LLMs via Spontaneity

Activation sparsity offers a compelling route to accelerate large language model (LLM) inference by selectively suppressing hidden activations, yet existing approaches exhibit severe accuracy degradation at high sparsity. We show that this failure stems from representational instability: activation sparsity disrupts input-dependent activation learned during pretraining, inducing distribution shifts in hidden states. We address this issue by reframing activation sparsity as a representational alignment problem and introducing Spontaneous Neurons (SPON), a lightweight mechanism inspired by spontaneous neural activity in biological systems. SPON injects a small set of learnable, input-independent activation vectors that act as persistent representational anchors for sparse computation. These vectors are trained via distribution matching to the dense model and can be absorbed into bias terms after training, incurring negligible inference overhead. Across multiple LLM backbones, SPON consistently restores performance, stabilizes latent representations, and preserves generalization. Our results establish SPON as an effective and principled solution for reliable activation-sparse inference, and offer new insights into knowledge retention in LLMs.

Activation sparsity can make large language models much faster by skipping many neuron activations during inference, but this often causes severe accuracy loss because the model’s internal representations become unstable. In this work, we show that sparsity disrupts the activation patterns learned during pretraining, leading to a drift in hidden representations and degraded reasoning ability. To address this problem, we introduce \emph{Spontaneous Neurons (SPON)}, a lightweight mechanism inspired by spontaneous neural activity in biological brains. SPON injects a small set of learned, input-independent activations that act as stable “anchors” for sparse computation, helping the model preserve its original knowledge even when many neurons are inactive. These activations can be merged into bias terms after training, resulting in virtually zero additional inference cost. Across multiple LLM families, tasks, sparsity levels, and scales up to 70B parameters, SPON consistently improves sparse model performance while adding less than 0.016% extra parameters. Our findings suggest that maintaining stable internal representations is essential for reliable sparse inference and may inspire future LLM architectures to reconsider the role of seemingly redundant components such as bias terms.