Researchers note input length directly affects activation sparsity, call for sequence-compression methods

antoine_chaffin · x · 2026-09-29

antoinechaffin and a collaborator discuss activation sparsity in LLMs, noting it is well established that input sequence length directly affects sparsity. He floats an open research idea: a method that reduces sequence embedding length without hurting the underlying representations.

Original post →

More from Research

Research channel →