Source-linked AI summary
Environment-Adaptive Dynamic KV Cache for Multi-Path Parameter Prediction
Zhaoyang Li, Yuhao Chen, Qianqian Yang, Zhaoyang Zhang
TL;DR
Existing Transformer-based multipath predictors repeatedly encode the same environment for different BS–UE links. This paper reuses environment-side representations and KV states, with dynamic cache compression. On WAIR-D, KV memory drops from 0.2500 MB to 0.1406 MB while delay NMSE changes from 0.0673 to 0.0678.
Problem
Transformer-based predictors repeatedly encode environmental input for each BS–UE query, although many links share an essentially unchanged environment.
Source: §1 · p. 1Method
The framework separates environment encoding from geometry-dependent prediction, reuses environment-side representations and KV states across queries, and compresses the cache using environment-side attention scores and layer-wise budget allocation.
Source: §1 · p. 1, §2.1 · p. 2, §2.2 · p. 2Results
On WAIR-D, KV memory decreases from 0.2500 MB to 0.1406 MB while delay NMSE changes from 0.0673 to 0.0678.
Source: §1 · pp. 1–2Takeaways & Limitations
The reported results demonstrate a memory–accuracy tradeoff, with cache capacity providing a controllable option for deployment under different memory constraints.
Source: §3.4 · pp. 3–4
Abstract
from arXiv · showhide
Multipath parameter prediction provides a structured and physically interpretable representation of wireless propagation for environment-aware communication systems. However, existing Transformer-based methods repeatedly encode the same environment when predicting multiple BS--UE links, resulting in redundant computation. To address this issue, we propose an environment-conditioned dynamic KV caching framework for multipath parameter prediction. Specifically, we decouple environment-dependent computation from BS--UE geometry, enabling environment-side representations and intermediate KV states to be reused across multiple link queries while only the geometry branch is updated. Building on this reusable cache, we further introduce a dynamic compression mechanism that uses environment-side attention scores to retain informative tokens and adaptively allocates a fixed cache budget across Transformer layers. Experiments on the WAIR-D dataset show that the proposed method reduces KV memory from 0.2500 MB to 0.1406 MB, while the delay NMSE increases only slightly from 0.0673 to 0.0678. Code is available at https://github.com/LZhaoyang/kv.
1. INTRODUCTION
Path-wise multipath parameters offer a compact, physically interpretable alternative to aggregate channel representations. Because many links can share a slowly changing environment, the paper proposes reusing environment-side Transformer computations and dynamically compressing their KV cache.
1. INTRODUCTION
- Path-wise representations describe dominant multipath components by gain, delay, AoD, and AoA, whereas aggregate CSI, covariance, or beam-index representations obscure individual paths’ physical relationship to the environment.The paper presents path-wise prediction as compact and physically interpretable, with potential use in channel reconstruction, beam management, localization, and sensing.Source: §1 · p. 1
- Ray tracing becomes more computationally costly as environments grow complex, while signal-processing parameter estimation typically needs online observations and predefined propagation or sparsity assumptions [7, 8] [9–11].Neural networks and Transformer models have also been studied for wireless channel estimation and prediction.Source: §1 · p. 1
- Existing Transformer predictors jointly process environmental input and BS–UE geometry for each query, repeating environment encoding even when links share a slowly changing setting.The paper notes that this redundancy burdens base stations and edge devices with limited computational and memory resources.Source: §1 · p. 1
- The proposed environment-conditioned dynamic KV cache reuses environment representations and intermediate KV states across link queries, while dynamic compression selects informative tokens and allocates cache budget across layers.The introduction reports WAIR-D results: KV memory falls from 0.2500 MB to 0.1406 MB, while delay NMSE changes from 0.0673 to 0.0678.Source: §1 · pp. 1–2
2. METHOD
The predictor encodes an environmental image and BS–UE geometry to predict dominant multipath components. It reuses environment-side representations and KV states across link queries, then compresses the cache using environment-side attention scores and a layer-adaptive budget.
2.1. Multipath Prediction Formulation
- The model predicts up to Kmax candidate paths, each with an existence confidence and six propagation parameters: delay, two-dimensional AoA, two-dimensional AoD, and gain.The implementation sets Kmax = 16.Source: §2.1 · p. 2
2.2. KV Caching for Environment Reuse
- For a fixed environment, the model caches environment-side representations and layer-wise KV states, recomputing only the geometry-dependent branch for subsequent BS–UE queries.The first query evaluates the image encoder and environment branch to construct the cache; later queries skip those computations.Source: §2.2 · p. 2
2.3. Environment-Conditioned Dynamic KV Compression
- Dynamic compression selects historical environment tokens using attention-derived importance scores, while retaining a local window and allocating a fixed total cache budget across layers.The compressed environment-side states replace the full KV states for subsequent BS–UE queries.Source: §2.3 · pp. 2–3
3. EXPERIMENTS
Experiments evaluate prediction performance, reusable KV caching and compression, cache-capacity tradeoffs, and cross-scenario generalization. Dynamic caching lowers cache memory, while the model maintains prediction accuracy across the tested cache settings and scenarios.
3.1. Experimental Setup
- Experiments use WAIR-D Scene 1, with five BSs and 30 UEs per environment, and compare the four-layer, eight-head Transformer against adapted ULCCME and SICCME baselines.The setup uses 28 GHz, samples 64 of 384 subcarriers, and runs on one NVIDIA RTX A6000 GPU.Source: §3.1 · p. 3
3.2. Transformer-Based Prediction Performance
- At ten paths, per-path accuracy is 0.793 for the proposed model, versus 0.564 without environmental images and 0.628 without geometry.Across most path counts, the Transformer without cache also achieves lower delay, AoA, AoD, and gain NMSE than adapted ULCCME and SICCME baselines.Source: §3.2 · p. 3, Figure 5 · p. 4, Figure 4 · p. 4
3.3. KV Cache Reuse and Compression Gains
- KV reuse reduces mean online latency from 9.20 ms to 7.16 ms, while dynamic compression cuts KV memory from 0.2500 MB to 0.1406 MB (43.75%) with marginal changes in four NMSE metrics.No-cache, full-cache, and dynamic-cache NMSE curves nearly overlap; full KV reuse reproduces uncached NMSE values.Source: §3.3 · p. 3, Figure 4 · p. 4, Table 1 · p. 4
3.4. Dynamic KV Cache Capacity
- With W = 4, reducing Cmax from 12 to 6 lowers KV memory from 0.1875 MB to 0.0938 MB while NMSE changes from 0.0677 to 0.0683.At Cmax = 6, dynamic caching reduces memory by 62.50% relative to the full cache, providing a controllable memory–accuracy tradeoff.Source: §3.4 · pp. 3–4, Table 2 · p. 4
3.5. Cross-Scenario Generalization
- On Scenario 2, the model trained on Scenario 1 achieves per-path accuracies of 0.666, 0.795, and 0.755 for eight, nine, and ten paths, exceeding adapted non-Transformer baselines.Evaluation is without finetuning and includes unseen environments and a different UE-location distribution.Source: §3.5 · p. 4, Figure 6 · p. 4, Table 3 · p. 4
4. CONCLUSION
The framework reuses environment-side representations and KV states across BS–UE queries, with dynamic compression reducing memory while preserving a small delay-NMSE increase on WAIR-D.
4. CONCLUSION
- KV memory fell by 43.75% on WAIR-D, while delay NMSE increased by only 0.0005.The authors describe this as an effective memory–accuracy tradeoff for repeated multipath parameter prediction.Source: §4 · p. 4
- Environment-dependent computation is decoupled from geometry-dependent prediction, so environment representations and layer-wise KV states can be reused across multiple BS–UE queries.Source: §4 · p. 4
- Dynamic compression selects informative environment tokens using environment-side attention and adaptively allocates the cache budget across Transformer layers.Source: §4 · p. 4