Reducing Peak Memory Usage for Modern Multimodal Large Language Model Pipelines (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing methods to reduce memory usage of multimodal large language models rely on storing large numbers of vision tokens in the key–value cache . however, such compression is typically only applied after all inputs are processed, resulting in high peak memory usage during the prefill stage. |
| Approach: | They propose a sequential input-compression mechanism that enforces a fixed memory budget by performing structure-aware key–value cache compression during the prefill stage. |
| Outcome: | The proposed approach reduces peak memory usage while maintaining generative performance with only minimal degradation, enabling more practical and memory-efficient multimodal inference. |
Similar Papers
HybridKV: Hybrid KV Cache Compression for Efficient Multimodal Large Language Model Inference (2026.acl-long)
Copied to clipboard
| Challenge: | Multimodal Large Language Models (MLLMs) are hindered by the rapid growth of key–value (KV) caches. |
| Approach: | They propose a hybrid KV cache compression framework that reduces KV memory by 7.9 and speeds up decoding by 1.52. |
| Outcome: | Experiments on 11 multimodal benchmarks show that HYBRIDKV cuts KV cache memory by 7.9 and speeds up decoding by 1.52. |
MEDA: Dynamic KV Cache Allocation for Efficient Multimodal Long-Context Inference (2025.naacl-long)
Copied to clipboard
| Challenge: | Long-context Multimodal Large Language Models (MLLMs) require substantial computational resources as their multimodal Key-Value (KV) cache grows with increasing input lengths, challenging memory and time efficiency. |
| Approach: | They propose a dynamic multimodal KV cache allocation strategy that dynamically allocating KV size based on attention entropy to better adapt to multimodal interactions. |
| Outcome: | The proposed model achieves up to 72% KV cache memory reduction and 2.82 faster decoding speeds while maintaining or enhancing performance on various multimodal tasks in a long context. |
LOOK-M: Look-Once Optimization in KV Cache for Efficient Multimodal Long-Context Inference (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Long-context Multimodal Large Language Models (MLLMs) require substantial computational resources for inference . the growth of their multimodal Key-Value (KV) cache challenges memory and time efficiency. |
| Approach: | They propose a fine-tuning-free approach that efficiently reduces the multimodal KV cache size while maintaining performance comparable to a full cache. |
| Outcome: | The proposed method reduces the multimodal KV cache size while maintaining performance comparable to a full cache. |
MM-ShiftKV: Decode-Aware Prefill-Stage KV Selection for Multimodal Large Language Models (2026.findings-acl)
Copied to clipboard
| Challenge: | Recent work suggests a prefill-stage KV cache selection method to estimate KV importance from prefilling statistics. |
| Approach: | They propose a training-free, decode-aware and strictly prefill-only KV selection method that retains key-value caching for decoding . |
| Outcome: | The proposed method outperforms existing methods under tight cache budgets on multimodal benchmarks. |
Efficient Inference for Large Vision-Language Models: Bottlenecks, Techniques, and Prospects (2026.findings-acl)
Copied to clipboard
Jun Zhang, Yicheng Ji, Feiyang Ren, Yihang Li, Bowen Zeng, Zonghao Chen, Ke Chen, Lidan Shou, Gang Chen, Huan Li
| Challenge: | Large Vision-Language Models are hindered by a systemic efficiency barrier known as visual token dominance. |
| Approach: | They propose a systematic taxonomy of efficiency techniques structured around the inference lifecycle . they examine visual encoding, prefilling, and decoding to understand bottlenecks . |
| Outcome: | The proposed techniques reveal how upstream decisions dictate downstream bottlenecks . the proposed techniques include hybrid compression and modality-aware decoding . |
Token Pruning in Multimodal Large Language Models: Are We Solving the Right Problem? (2025.findings-acl)
Copied to clipboard
| Challenge: | Multimodal large language models have shown remarkable performance for cross-modal understanding and generation, yet suffer from severe inference costs. |
| Approach: | They propose to prune redundant tokens in MLLMs to reduce computation and storage costs. |
| Outcome: | The proposed method reduces the computational and storage costs of MLLMs by identifying redundant tokens and pruning them. |
Look and Think: Efficient Multimodal Reasoning via Modality-Decoupled Compression (2026.findings-acl)
Copied to clipboard
| Challenge: | Multimodal large language models have strong performance on visual question answering benchmarks . however, their inference efficiency is severely constrained by the rapidly growing context . |
| Approach: | They propose a modality-decoupled compression method that enables efficient multimodal inference . they propose to evict visual tokens whenever visual grounding is unnecessary . |
| Outcome: | The proposed method reduces the average context length by up to 57% while maintaining comparable performance to the standard MLLM baseline. |
From Verbatim to Gist: Distilling Pyramidal Multimodal Memory via Semantic Information Bottleneck for Long-Horizon Video Agents (2026.acl-long)
Copied to clipboard
| Challenge: | Existing multimodal large language models struggle with long-horizon video understanding due to limited context windows and static memory mechanisms that fail to mirror human cognitive efficiency. |
| Approach: | They propose a pyramidal multimodal memory architecture grounded in Fuzzy-Trace Theory that structures memory hierarchically into a *Sensory Buffer*, *Episodic Stream*, and *Symbolic Schema*. |
| Outcome: | The proposed architecture achieves state-of-the-art on both offline and streaming tasks, demonstrating robust generalization and validating the effectiveness of cognition-inspired memory organization. |
Learning Flexible Large Multimodal Models with Arbitrary Modality Combinations (2026.findings-acl)
Copied to clipboard
| Challenge: | Multimodal Large Language Models (MLLMs) have potential for cross-modal understanding . but extending MLLM to handle diverse modalities introduces two challenges . |
| Approach: | They propose a dual-stage compression mechanism to reduce the number of modality tokens per modality and condense it into a single, compact token sequence. |
| Outcome: | Experiments show that Flex-M3 outperforms its counterpart trained on only full-modality data. |
Probing Multimodal Large Language Models for Global and Local Semantic Representations (2024.lrec-main)
Copied to clipboard
| Challenge: | Existing studies have focused on the ability of MLLMs to generate single tokens one by one, while lacking studies about how their representation vectors can encode global multimodal information. |
| Approach: | They propose to use image-caption corpus to train Multimodal Large Language Models (MLLMs) . they find that the topmost layers encode more global semantic information . |
| Outcome: | The proposed models can encode more global semantic information, rather than the topmost layers, and perform better on visual-language entailment tasks. |