Evaluate TopK-Guided at larger scales and on other architectures

Determine how TopK-Guided behaves in decoder-only language models larger than 8 billion parameters and in architectures other than those evaluated in the study.

Background

TopK-Guided is evaluated only on the decoder-only Llama-2-7B and Llama-3-8B models. The proposed method is described as making no architecture-specific assumptions, but its behavior outside these model sizes and architectures has not been established. The authors identify evaluation at larger scale and on other architectures as an unresolved limitation.

References

We evaluate decoder-only models up to 8B parameters, so the method's behaviour at larger scale and on other architectures remains to be checked, though it makes no architecture-specific assumptions.

— TopK-Guided: Adaptive, Budget-Aware Activation Sparsity for Efficient LLM Inference  (2610.01763 - Agarwalla et al., 1 Oct 2026) in Section Conclusion and Limitations