Structured-Token Attention for Fine-Grained mmWave Liquid Identification

Junseo Min; Joongheon Kim

cs.AIOct 20, 2026

Structured-Token Attention for Fine-Grained mmWave Liquid Identification

Authors: Junseo Min, Joongheon Kim

Organizations: Department of Defense Convergence Technology, Korea University, Seoul, Republic of Korea · Department of Electrical & Computer Engineering, Korea University, Seoul, Republic of Korea

Abstract

Fine-grained liquid identification requires a classifier to distinguish similar substances from compact millimeter-wave measurements. We investigate structured-token attention, which represents a radar feature tensor through complementary antenna, feature, branch, and global views. The proposed StructAttn model embeds these views into eleven tokens and combines them with a three-layer transformer. Experiments on the public FG-LiquID data cover 30 liquids and 3,000 samples. Over five seeds on the released split, StructAttn achieves 96.70% mean best-epoch accuracy, matching the original-recipe MMNet baseline and exceeding a scalar-token attention model by 1.50 percentage points. A five-seed probability ensemble reaches 97%, corresponding to 18 errors among 600 test samples. Final-epoch and five-fold results complement the checkpoint-selected comparison, with StructAttn obtaining 95.30% out-of-fold accuracy. Token and feature analyses show that structured views are useful for prediction, while residual confusions concentrate in water, soft-drink, and edible-oil classes. These results support structured tokenization as a practical design choice for attention-based liquid recognition.

Figures & tables

Fig. 1: Structured views and attention processing in StructAttn-S3.

View	Count	Values	Grouping
Antenna	4	12	One receiver
Feature	4	12	One observable
Location	1	36	Distance, angles
Reflection	1	12	RSS branch
Global	1	48	Complete tensor
Total	11	–	Without [CLS]

TABLE I: Token construction for StructAttn-S3.

Model	Parameters	Best epoch	Final epoch	Macro-F1	Five-seed ensemble
Nearest centroid	0	70.17	70.17	69.31	–
MLP	50K	$94.47\pm 0.22$	$93.73\pm 0.45$	94.23	–
MMNet	238K	$96.70\pm 0.70$	$\mathbf{96.30\pm 0.76}$	96.66	96.33 (22 errors)
MMNet-R	238K	$\mathbf{96.93\pm 0.09}$	$96.23\pm 0.19$	96.85	96.67 (20 errors)
NoAttn	283K	$92.37\pm 0.14$	$91.43\pm 0.25$	92.10	–
AttnNet	548K	$95.20\pm 0.76$	$94.53\pm 0.92$	94.84	–

TABLE II: Released-split performance over five seeds. Accuracy and macro-F1 are percentages.

Category	Classes	Train	Test
Liquor	12	958	242
Edible oil	5	393	107
Soft drink	5	409	91
Water	2	161	39
Milk	2	157	43
Beer and wine	2	162	38

TABLE III: Composition of the released FG-LiquID split.

Fig. 2: Best- and final-epoch accuracy over five seeds.

Variant	Tokens	Width/layers	Params	Best	Final
S1	14	192/3	1.24M	96.33	96.00
S2	14	256/4	2.56M	96.83	96.00
S3	11	192/3	1.19M	97.00	96.50
S4	3	192/2	782K	96.33	95.00

TABLE IV: Token variants on seed 0. Accuracy is in percent.

Fig. 3: Feature masking with MMNet-R at the final epoch.

Metric or category	MMNet-R	StructAttn-S3
Fold mean $\pm$ std	$95.77\pm 0.40$	$95.30\pm 0.64$
Liquor	99.8	99.9
Edible oil	90.6	88.6
Soft drink	92.0	91.8
Water	84.5	82.5
Other categories	98.8	98.8

TABLE V: Final-epoch five-fold results on all 3,000 samples.

True liquid	Correct/100	Predicted instead	Count
Mineral water	74	Tap water	14
		Soda	12
Olive oil	81	Soybean oil	19
Soybean oil	83	Olive oil	17
Soda	84	Tap water	10
		Mineral water	6

TABLE VI: Representative S3 out-of-fold confusions.

Fig. 4: Accuracy across sampled liquid subsets. Bands show ranges.

Classification

Classification maps observations to discrete labels, including familiar, unknown, hierarchical, or multiple categories.

See all 408 papers

Can Language Models Understand mmWave Data? Benchmarking Large Language Models for mmWave Radar-Based Human Understanding
Aug 14, 2026Jeongwan Shin, Jaehyeon Kim, Donguk Ko +1mmWave Radar Sensing LLM Evaluation
DGIST · KAIST InnoCORE LLM
Teaching Foundation Models to Read mmWave: Pose-Guided Kinematic Representation for Human Behavior Understanding
Aug 4, 2026Duo Zhang, Zhehui Yin, Zhiyun Yao +8Representation Learning Human Activity Recognition
1Peking University · 2ETH Zurich · 3Institut Polytechnique de Paris
Weather-Robust Scene Semantics with Vision-Aligned 4D Radar
May 8, 2026Kali Hamilton, Christoffer HeckmanCross-Modal Learning VLM Robustness
Autonomous Robotics and Perception Group at the University of Colorado Boulder, Boulder, CO 80309, USA.

Aug 14, 2026J/K move · Enter open · S save

Can Language Models Understand mmWave Data? Benchmarking Large Language Models for mmWave Radar-Based Human Understanding

Jeongwan Shin, Jaehyeon Kim, Donguk Ko, Jaeho Choi

DGIST · KAIST InnoCORE LLM

mmWave Radar Sensing LLM Evaluation

Open paper AskPDF arXiv

Large language models (LLMs) have shown remarkable reasoning and generative capabilities, motivating their use as universal reasoning engines for perception. While modern approaches such as vision-language models (VLMs) have attempted to incorporate reasoning capabilities into visual sensing, the integration of LLMs with the millimeter-wave (mmWave) modality-despite its unique advantages under low light and occlusion-remains largely unexplored. The principal bottlenecks stem from the scarcity of radar language pairs, severe cross-dataset heterogeneity, and the absence of a foundational mmWave encoder. We address this gap through a minimal textualization interface that serializes each mmWave point cloud into concise natural language, allowing off-the-shelf LLMs to operate in a question answering (QA) setting. Building on this, we present mmWave-QA, the first benchmark for language-conditioned mmWave human perception. mmWave-QA aggregates heterogeneous public mmWave datasets and harmonizes them via calibration-aware preprocessing and global taxonomy alignment, while providing natural language QA. Spanning six scenarios and five QA tasks, the benchmark enables standardized evaluation across diverse mmWave hardware and experimental conditions, establishing a foundation for scalable research on mmWave-LLM integration. We further evaluate and analyze LLMs on our mmWave-QA, highlighting their zero-shot reasoning potential for radar perception, as well as their robustness under visual degradation.

Abstract

Classification

Explore similar work

Can Language Models Understand mmWave Data? Benchmarking Large Language Models for mmWave Radar-Based Human Understanding

Teaching Foundation Models to Read mmWave: Pose-Guided Kinematic Representation for Human Behavior Understanding

Weather-Robust Scene Semantics with Vision-Aligned 4D Radar