Table 2
Comparative summary of MLLM compression techniquesa)
| Sec. | Technique | Core idea | Reduction type | Typical gain | Hardware |
|---|---|---|---|---|---|
| Modality Enc. | Visual Token Comp. | Resample / merge visual tokens before LLM | Token ↓ | 4x–16x tokens | GPU |
| Token Pruning | Remove redundant tokens via saliency/attention | Token ↓ | 3x–18x tokens; 1.2x–1.8x latency | GPU | |
| Sparse Attention | Restrict attention patterns (local/global) | Complexity ↓ | Moderate speedup | Custom kernels preferred | |
| Audio Modality | Downsample or compress temporal frames | Token ↓ | 2x–10x tokens | GPU | |
| Cross-Modal Conn. | Linear Projection | Linear mapping for modality alignment | Token/Dim | ~1x–4x | Universal |
| Query-based | Bottleneck queries (e.g., Q-Former) | Token ↓ | 8x–20x tokens | Universal | |
| Other Connectors | pooling/routing variants | Token/Dim ↓ | 2x–10x | Universal | |
| LLM Backbone | Small LMs | Replace backbone with smaller LLM | Params↓ | 2x–10x params | Edge/GPU |
| MoE | Activate subset of experts per token | Compute ↓ | Throughput ↑ | High bandwidth | |
| Non-typical | RNN / state-space architectures | Complexity ↓ | Long-seq speedup | Kernel support helpful | |
| General Opt. | Distillation | Transfer knowledge to smaller model | Params ↓ | Variable | Universal |
| Quantization | Low-bit weights (INT8/4, NF4) | Memory ↓ | 2x–4x memory | HW-dependent | |
| Pruning | Remove weights / channels | Params ↓ | ≤2x | Sparse kernels |
a) (1) “Reduction type" distinguishes token, parameter, memory, or computational complexity reduction. (2) Sparse attention, MoE, and state-space models are efficiency-oriented rather than strict compression. (3) Reported gains are component-level; end-to-end latency depends on system implementation.
Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.
Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.
Initial download of the metrics may take a while.