Efficient serving of multimodal large language models (MLLMs) is increasingly bottlenecked by high-resolution visual inputs, which can generate thousands of visual tokens per request and severely increase latency, memory pressure, and serving cost. Visual token pruning is a natural solution, but existing methods are difficult to deploy in production MLLM serving: many operate inside LLM layers and conflict with engine invariants such as KV-cache granularity, fused kernels, prefix caching, and batching, while fixed-ratio pruning can collapse accuracy on dense-information tasks such as document and chart understanding.

We present SieveServe, a serving-compatible framework for adaptive visual token pruning. SieveServe establishes pre-LLM vision-centric pruning as an engine-compatible paradigm and introduces a request-aware dual-guard policy: a Query Complexity Router (QCR) estimates query-side visual evidence demand, and a Dense Information Guard (DIG) detects image-side pruning risk using lightweight channel-wise embedding variance. Together, they allocate visual-token budgets per request, aggressively compressing pruning-robust inputs while conservatively preserving dense/high-risk ones.

Experiments with Qwen2.5-VL-7B and InternVL3-8B across eight benchmarks show that SieveServe retains at least 95% of baseline accuracy on all benchmarks and at least 97.8% on dense-information tasks, while compressing pruning-robust tasks by 3.2x and recovering over 20 accuracy points lost by fixed-ratio pruning. On high-resolution serving workloads, SieveServe reduces P90 TTFT by 5.4x and sustains 2.2-2.7x higher throughput under latency SLO constraints compared with vanilla vLLM.