It maintains VideoMME benchmark accuracy without model retraining.
Jiajun Wu#358
1/ The biggest problem in video understanding today isn't the models. It's that we can barely run them.
Introducing StateKV: an inference-time method that makes pretrained video VLMs scale linearly with video length.🧵
🔗 http://ceyzaguirre4.github.io/StateKV
Processing long videos with VLMs shouldn't scale quadratically. Enter StateKV! 🎬💡
By framing streaming prefill as a fixed-capacity temporal state, we unlock linear-time prefill while keeping full per-frame detail.
Paper by @CristbalEyzagu2 and team👇
https://arxiv.org/abs/2605.31598
1/ The biggest problem in video understanding today isn't the models. It's that we can barely run them.
Introducing StateKV: an inference-time method that makes pretrained video VLMs scale linearly with video length.🧵
🔗 http://ceyzaguirre4.github.io/StateKV
It maintains VideoMME benchmark accuracy without model retraining.
Jiajun Wu#3581/ The biggest problem in video understanding today isn't the models. It's that we can barely run them.
Introducing StateKV: an inference-time method that makes pretrained video VLMs scale linearly with video length.🧵
🔗 http://ceyzaguirre4.github.io/StateKV