The problem
State space models process long sequences efficiently, but applying the same computation to every token can still waste work on tokens that contribute little to the task.
What we did
We studied dynamic token skipping in the Mamba state space model. ACE routes semantically insignificant tokens past compute-heavy state space blocks, allocating computation according to token importance rather than processing every token identically.
What we found
Experiments reported in the resume show that roughly half of the tokens can bypass these blocks without performance degradation, while improving inference throughput. This connects input-adaptive model execution with the broader goal of compute-efficient language processing.
Paper
Ashiqur Rasul, Abhijit Chatterjee. “ACE: Adaptive Compute-Efficient State Space Models with Dynamic Token Routing.” AAAI 2026, under review, as listed in the latest resume.