Cross-Modal StateKV Injection for Communication in Multi-LLM Systems across Temporal Segments
Authors: Xiaoqian Zhang, Zhihao Zhang
Organizations: Department of Ultrasound & Faculty of Medical Imaging Technology, Ruijin Hospital, College of Health Science and Technology, Shanghai Jiao Tong University School of Medicine, Shanghai, China · School of Mathematics, Shanghai University of Finance and Economics, Shanghai, China · School of Computing and Artificial Intelligence, Shanghai University of Finance and Economics, Shanghai, China
Video understanding with multiple multimodal language models requires a communication interface that can transfer complementary visual observations. Text-based communication is interpretable, but introduces an intermediate generation stage and can omit information relevant to subsequent reasoning. We study Cross-Modal StateKV Injection, a provider--carrier framework that communicates through key--value (KV) states at visual-token positions. Provider models observe complementary video frames; their visual KV states are aligned and projected into the carrier representation space and integrated through gated residual updates. Injection preserves the carrier KV-cache sequence length, while retaining the additional computation required by the providers and the communication module. The reported experiments on MLVU and Video-MME compare carrier-only inference, text-based communication, and StateKV with two provider architectures. Results show modest, task-dependent accuracy gains and lower reported latency than the evaluated text-communication baseline, with some categories remaining unchanged or decreasing. Increasing provider-side frame coverage offers additional gains on selected tasks without extending the carrier cache. These findings motivate further study of visual KV communication under explicitly controlled observation and computation budgets.
Figures & tables
Figure 1: Text and visual-state communication. Text communication supplies an intermediate generated description, whereas StateKV transfers visual-context information through KV injection. A shared textual prompt does not imply identical internal states across models.
Figure 2: StateKV architecture. Provider and carrier caches feed a visual-state preparation module, followed by projection, modulation and gated residual injection into the carrier cache. The carrier performs the final autoregressive decoding.
Figure 3: Layer-wise rewrite ratios for keys (left) and values (right). The carrier is Qwen3-VL-2B; providers are Qwen3-VL-4B and InternVL3.5-4B.
Figure 4: Qualitative counting example. Carrier-only inference predicts one riding-mule instance, while the provider–carrier setting predicts four after observing additional frames. The additional observations are incorporated through visual KV injection while retaining the carrier cache length. This example illustrates the information pathway rather than establishing an aggregate accuracy gain.
Accuracy (%)
Reported time (s)
Task
Carrier
Text
StateKV
Carrier
Text
StateKV
Count
17.48
14.56
20.87
0.19
3.22
0.43
Ego
46.38
43.18
46.88
1.17
5.11
1.9
Needle
49.29
40.84
49.39
0.21
3.26
0.46
Order
35.90
33.97
38.22
0.4
3.43
0.71
Plot QA
49.92
45.64
50.72
0.64
4.19
1.23
Table 1: MLVU with the Qwen provider.
Accuracy (%)
Reported time (s)
Task
Carrier
Text
StateKV
Carrier
Text
StateKV
Count
17.48
16.01
19.90
0.19
12.83
0.86
Ego
46.38
42.26
48.01
1.17
15.2
1.32
Needle
49.29
40.56
48.73
0.21
13.97
0.91
Order
35.90
32.05
36.68
0.4
14.17
1.14
Plot QA
49.92
45.40
50.72
0.64
14.53
1.08
Table 2: MLVU with the InternVL provider.
Accuracy (%)
Reported time (s)
Task
Carrier
Text
StateKV
Carrier
Text
StateKV
Short
55.67
55.89
55.44
0.23
3.61
0.59
Medium
43.11
42.56
42.89
0.22
3.72
0.59
Long
41.56
39.00
42.22
0.23
3.7
0.6
Table 3: Video-MME with the Qwen provider.
Accuracy (%)
Reported time (s)
Task
Carrier
Text
StateKV
Carrier
Text
StateKV
Short
55.67
55.56
55.67
0.23
12.54
0.59
Medium
43.11
42.11
42.89
0.22
12.53
0.59
Long
41.56
38.78
42.44
0.23
12.54
0.6
Table 4: Video-MME with the InternVL provider.
8 provider → 8 carrier
16 provider → 8 carrier
Task
Accuracy (%)
Time (s)
Accuracy (%)
Time (s)
Count
20.87
0.43
19.42
0.43
Ego
46.88
1.90
47.44
1.99
Needle
49.39
0.46
48.73
0.46
Order
38.22
0.71
38.22
0.67
Plot QA
50.72
1.23
50.83
1.32
Table 5: Provider frame-budget comparison with the Qwen provider and eight carrier frames. Results are organized by task.
Video vision-language models (VLMs) are increasingly used in long-horizon and streaming settings, yet most video encoders still rely on spatiotemporal self-attention, causing compute and latency to grow quadratically with the number of frames. Existing efficiency methods improve scalability but often lose accuracy relative to full self-attention, for example through aggressive frame/token dropping or coarse attention approximations. We introduce StateKV, an inference-time method that adapts pretrained long-video VLMs to linear-time video prefill by carrying cross-frame context in a fixed-capacity, importance-based recurrent state, paired with a second full per-frame cache used for decoding. Across three long-video benchmarks and seven models spanning three families and multiple scales, StateKV remains close to full self-attention and consistently outperforms dominant sliding-window / recency-based streaming approximations, without fine-tuning or architectural changes. StateKV also reduces video-prefill cost measured FLOPs, enabling stronger accuracy at a fixed compute budget by running larger models. These results suggest a practical step toward scalable long-video understanding.
Cristobal Eyzaguirre, Jiajun Wu, Juan Carlos Niebles
Understanding a video requires more than recognizing isolated moments, as humans continuously track entities, states, and events over time. This capacity for visual state tracking is fundamental to video understanding, yet remains underexplored in current evaluations of Multimodal Large Language Models (MLLMs). We introduce Visual STAte Tracking benchmark (VSTAT), a video-based benchmark designed to diagnose visual state tracking in MLLMs. VSTAT consists of 834 clips drawn from both synthetic and real-world videos, paired with 1,500 questions that cannot be answered from any single frame or short segment, requiring continuous perception and integration of events across the entire video stream. Despite their strong performance on existing video benchmarks, we find that state-of-the-art MLLMs perform far below humans and only modestly above answer-prior baselines. To analyze this gap, we compare MLLMs' thinking traces with the underlying video stream to understand why and when MLLMs fail on VSTAT. We find that MLLMs reason and track correctly in text, but fail at visually perceiving the events they need to track. Finally, our preliminary evaluation suggests that recent agentic approaches, including MLLM-based video agents and coding agents, do not readily resolve these failures, still falling short on VSTAT.
Multimodal Large Language Models (MLLMs) have advanced unified reasoning over text, images, and videos, but their inference is hindered by the rapid growth of key-value (KV) caches. Each visual input expands into thousands of tokens, causing caches to scale linearly with context length and remain resident in GPU memory throughout decoding, which leads to prohibitive memory overhead and latency even on high-end GPUs. A common solution is to compress caches under a fixed allocated budget at different granularities: token-level uniformly discards less important tokens, layer-level varies retention across layers, and head-level redistributes budgets across heads. Yet these approaches stop at allocation and overlook the heterogeneous behaviors of attention heads that require distinct compression strategies. We propose HybridKV, a hybrid KV cache compression framework that integrates complementary strategies in three stages: heads are first classified into static or dynamic types using text-centric attention; then a top-down budget allocation scheme hierarchically assigns KV budgets; finally, static heads are compressed by text-prior pruning and dynamic heads by chunk-wise retrieval. Experiments on 11 multimodal benchmarks with Qwen2.5-VL-7B show that HybridKV reduces KV cache memory by up to 7.9× and achieves 1.52× faster decoding, with almost no performance drop or even higher relative to the full-cache MLLM.
Bowen Zeng, Feiyang Ren, Jun Zhang +4
The State Key Laboratory of Blockchain and Data Security, Zhejiang University · Hangzhou High-Tech Zone (Binjiang) Institute of Blockchain and Data Security · Hangzhou Dianzi University, Hangzhou, China