English

CausalGate: Causal Importance Distillation for Transformer Module Pruning

Machine Learning 2026-07-21 v1 Computation and Language Computer Vision and Pattern Recognition Machine Learning

Abstract

Existing adaptive inference methods for Large Language Models rely on observational heuristics, such as hidden-state similarity or activation magnitudes, to drop redundant modules. However, these correlation-based metrics often fail to capture subtle, non-linear structural computations vital for semantic accuracy. We introduce CausalGate, an intervention-guided framework for compute-efficient transformer inference. During a calibration phase, CausalGate isolates individual Attention and MLP sub-layers, zeros out their respective outputs, and measures the exact semantic damage via the Kullback-Leibler divergence of the final logit distribution. To eliminate runtime routing overhead, this structural importance hierarchy is distilled into a global set of static, lightweight scalar gates using an Exponential Moving Average smoothing objective paired with a differentiable pairwise ranking loss. Evaluated on TinyLlama-1.1B, Qwen2.5-3B, and Llama-3.1-8B across language modeling and commonsense reasoning benchmarks, CausalGate consistently outperforms prominent dynamic routing and layer-skipping baselines, translating theoretical compute savings into concrete hardware latency reductions with zero operational overhead.

Cite

@article{arxiv.2607.22720,
  title  = {CausalGate: Causal Importance Distillation for Transformer Module Pruning},
  author = {Kiran Nair and Smriti Regmi and Rodrigue Rizk},
  journal= {arXiv preprint arXiv:2607.22720},
  year   = {2026}
}

Comments

In-Review at a conference