What changed

RoRA allocates visual-token budgets across semantic core, complementary context, and local detail, retaining 96.5% of LLaVA-1.5 performance at 88.9% pruning in author-run tests.

Why it matters

Training-free pruning can cut multimodal prefill and cache load without retraining, if accuracy survives on the actual visual workload.

What to do

Replay the public patches on your models, prompts, and hardware, measuring end-to-end latency, memory, accuracy, and token-selection overhead together.

Caveat

The initial repository has three commits, no weights or data, model-specific patching, and only author-run H800 performance results.