Comparison with Standard Token Insertion + KV Eviction?

[RealJosephus]

Hi, thanks for the impressive work!

The streaming benchmarks currently compare CASA against "Full Insertion" (accumulating infinite memory). Have you compared it against a Standard Insertion + KV Eviction baseline (processing images via FFN but dropping old visual KVs)?

While CASA has a clear compute advantage by skipping FFNs, both methods share similar memory characteristics and RoPE handling (keeping position ID "gaps").

Do you have any performance comparisons against a simple KV eviction strategy?

Thanks!

[ameroyer]

Hello,
Thanks for your interest and the very interesting question :)
Indeed KVCache compression (whether pruning, merging or quantising) is another very relevant approach for streaming VLM execution. We don't have final results yet but we are investigating this and will update the repo/report when available