HICOM: Multidisciplinary Hierarchical Methods

Updated 6 July 2026

HICOM is an acronym family denoting distinct hierarchical strategies across domains, from graph compression to multimedia and wireless systems.
Each variant employs structured intermediate representations to reduce computational complexity while preserving domain-specific interactions.
Applications span text-rich graph node classification, multi-face deepfake detection, dynamic 3D scene rendering, video token compression, and cooperative traffic control.

HICOM, also rendered as HiCom, HiCoM, or HICom, is not a single standardized technical term but an acronymic family used for several unrelated research frameworks in contemporary arXiv literature. Across these usages, the label denotes methods for hierarchical compression of text-rich graphs, human-inspired multi-face deepfake detection, hierarchical coherent motion for streamable dynamic scene reconstruction, instruction-conditioned video token compression in multimodal LLMs, and related hierarchical or holographic control architectures in transportation and wireless systems (Zhang et al., 2024, Hu et al., 20 Jul 2025, Gao et al., 2024, Liu et al., 20 Mar 2025, Wang et al., 15 Jul 2025, Zhang et al., 9 May 2026).

1. Acronymic scope and nomenclature

The term appears in multiple fields with different expansions, architectural assumptions, and objective functions. The shared lexical pattern is the combination of hierarchy, compression, coherence, or cooperation, but the underlying mathematical objects differ sharply: token sequences in LLMs, graph neighborhoods, 3D Gaussians, traffic agents, or electromagnetic apertures.

Variant	Definition in source	Domain
HiCom	“Hierarchical Compression”	Text-rich graph node classification
HICOM	framework “designed to detect every fake face in multi-face scenarios”	Multi-face deepfake detection
HiCoM	“Hierarchical Coherent Motion”	Streamable dynamic scenes with 3D Gaussian Splatting
HICom	“Hybrid-level Instruction Injection Strategy for Conditional Token Compression in MLLMs”	Video token compression in multimodal LLMs
HCOMC	“Hierarchical Cooperative On-Ramp Merging Control”	Mixed-traffic on-ramp merging
HICOM	“holographic communications” as instantiated through RHS-enabled HISAC in the supplied material	Integrated sensing and communications

A common misconception is that HICOM denotes a single cross-domain architecture. The literature instead uses the acronym polysemously, with only loose methodological analogies across fields. In most cases, the decisive motif is a structured intermediate representation: a neighborhood summary, a multi-cue forensic signal, a hierarchy of regional motions, a compressed visual token set, or a beamforming surface constraint.

2. HiCom in text-rich graph learning

In graph machine learning, HiCom denotes the method introduced in "Hierarchical Compression of Text-Rich Graphs via LLMs" (Zhang et al., 2024). The problem setting is a text-rich graph $G=(V,X)$ , where each node $v_i\in V$ carries raw text $x_i=[x_i^1,\dots,x_i^t]$ , and node classification depends on both node-local language and graph neighborhood context. The method targets a setting in which standard GNN pipelines are limited because they require LM pre-encoding into fixed vectors, while flat LLM concatenation of multi-hop neighborhoods is constrained by context length and $O(n^2)$ attention complexity.

HiCom constructs a rooted neighborhood hierarchy around a target node $v_0$ , controlled by fanouts $[n_1,\dots,n_L]$ . Level $L$ contains the target node, level $L-1$ sampled 1-hop neighbors, level $L-2$ sampled 2-hop neighbors, and so forth. At each level, an LLM-based compressor $Comp(\cdot)$ receives child texts and lower-level summaries, appends learnable soft prompt tokens $v_i\in V$ 0, and returns summary vectors $v_i\in V$ 1. The final target summary $v_i\in V$ 2 is then concatenated with the target text as $v_i\in V$ 3 and passed to the same LLM backbone with a classification head.

The core computational claim is that hierarchical segmentation replaces a flat cost of $v_i\in V$ 4 with

$v_i\in V$ 5

while preserving token-level interactions inside local graph neighborhoods. The model also introduces summary accumulation, which includes summaries from all previous levels rather than only the immediate lower level, resembling skip connections and multi-hop filter combinations.

Empirically, HiCom outperforms both GNNs and LLM backbones on node classification for Amazon product graphs and MAPLE citation graphs. In dense regions, defined via $v_i\in V$ 6-core subgraphs, HiCom-OPT achieves a 3.48% average performance improvement on five datasets relative to the second-best method, with improvements of +0.75% on Cloth, +1.49% on Sports, +3.83% on Economics, +5.48% on Mathematics, and +5.83% on Geology. The paper further reports a positive correlation between improvement and average degree, with Pearson 0.667, and shows that with comparable neighbor count, HiCom [2,2] is ~44% faster than OPT-NConcat (31.24 versus 55.46 seconds per epoch) while improving test F1 (0.6716 versus 0.6250) (Zhang et al., 2024).

3. HICOM in multi-face deepfake detection

In multimedia forensics, HICOM is the human-inspired multi-face detector introduced in "Seeing Through Deepfakes: A Human-Inspired Framework for Multi-Face Detection" (Hu et al., 20 Jul 2025). Its task differs from conventional single-face detection: each frame may contain multiple faces, only some of which are manipulated, and correctness is measured not only per face but also at the frame-level complete multi-face detection level, where a frame is correct only if every face is correctly labeled.

The framework is explicitly grounded in human studies. In the cue-discovery phase, 4 university students analyzed 2,000 multi-face deepfake images/videos sampled from OpenForensics, FFIW, and DF-Platter; in the controlled ablation phase, 20 Amazon Mechanical Turk workers evaluated 920 stimuli with cue-blocking manipulations. The resulting cue taxonomy has four dominant components: scene-motion coherence (34.2%), inter-face appearance compatibility (31.5%), interpersonal gaze alignment (25.0%), and face-body consistency (7.5%), with minor cues at 1.8% excluded from modeling.

These findings map directly onto four modules. M1 models scene-motion coherence using multi-scale features, RoIAlign, and a spatio-temporal dynamic inference network, optimized with

$v_i\in V$ 7

M2 models inter-face appearance compatibility with a Transformer backbone and a contrastive pairwise loss on face embeddings, using $v_i\in V$ 8 and $v_i\in V$ 9. M3 is a rule-based gaze module built on a ResNet gaze classifier pretrained on the Columbia Gaze Dataset and refined on an internal dataset. M4 detects face-body age/gender mismatch using face crops and body crops with GaussianBlur applied to the face region inside the body box. Module outputs are fused with an XOR-like fusion strategy that functions as anomaly-OR at the face level, and an LLM layer produces human-readable explanations.

Evaluation uses FAC, FAU, FCAC, and FCAU. On in-dataset evaluation, HICOM reports +3.3% FCAC and +3.1% FCAU over the second-best method. Under OpenForensics perturbations, it reports +1.5% FCAC and +2.8% FCAU over the previous best method. On unseen ManualFake, it reports +5.8% FCAC over prior methods. The framework also exceeds average human performance on 300 randomly selected samples from FFIW, OpenForensics, and DF-Platter, with statistical significance reported as $x_i=[x_i^1,\dots,x_i^t]$ 0 (Hu et al., 20 Jul 2025).

4. HiCoM in streamable dynamic scene reconstruction

In neural rendering, HiCoM denotes the framework in "HiCoM: Hierarchical Coherent Motion for Streamable Dynamic Scene with 3D Gaussian Splatting" (Gao et al., 2024). The setting is online reconstruction of dynamic scenes from synchronized multi-view video streams, typically with 13–21 cameras, where the representation must support free-viewpoint video while remaining efficient in training, rendering, and storage.

The method has three principal components. First, it constructs the initial frame’s 3D Gaussian Splatting representation using perturbation smoothing, adding Gaussian noise to means during training,

$x_i=[x_i^1,\dots,x_i^t]$ 1

to reduce overfitting under sparse-view capture and to favor fewer, larger Gaussians. Second, it introduces Hierarchical Coherent Motion, in which non-empty spatial regions at multiple levels share translation and rotation parameters $x_i=[x_i^1,\dots,x_i^t]$ 2. A Gaussian’s motion is the sum across levels,

$x_i=[x_i^1,\dots,x_i^t]$ 3

The default design uses 3 levels—coarse, medium, and fine—with $x_i=[x_i^1,\dots,x_i^t]$ 4 maximum Gaussians per region at the finest level. Third, it performs continual refinement by cloning Gaussians with large accumulated gradients, merging new Gaussians into the evolving base model, and removing an equal number of low-opacity Gaussians before the next frame.

The optimization schedule uses $x_i=[x_i^1,\dots,x_i^t]$ 5 motion steps and $x_i=[x_i^1,\dots,x_i^t]$ 6 refinement steps per frame after the first frame, while retaining the standard 3DGS photometric loss

$x_i=[x_i^1,\dots,x_i^t]$ 7

A further systems contribution is parallel learning of multiple future frames from a shared reference frame. The HiCoM-P4 variant learns four frames in parallel and reduces average wall time per frame to < 2 seconds with negligible quality degradation.

On the N3DV benchmark, HiCoM reports 31.17 dB PSNR, 0.7 / 0.9 MB storage per frame excluding/including the first frame, 6.7 s/frame average training time, and 274 FPS rendering, compared with 30.15 dB, 7.6 / 7.8 MB, 8.2 s/frame, and 210 FPS for 3DGStream. On Meet Room, it reports 26.73 dB, 0.4 / 0.6 MB, 3.9 s/frame, and 284 FPS, compared with 25.96 dB, 4.0 / 4.1 MB, 4.7 s, and 212 FPS for 3DGStream. The paper summarizes these gains as about 20% improvement in learning efficiency and 85% reduction in data storage (Gao et al., 2024).

5. HICom in multimodal video token compression

In multimodal language modeling, HICom is the module presented in "Hybrid-Level Instruction Injection for Video Token Compression in Multi-modal LLMs" (Liu et al., 20 Mar 2025). Its purpose is to reduce the computational burden of long-video MLLMs by conditioning token compression on the user instruction rather than applying unconditional average pooling or frame downsampling.

The pipeline begins with a frozen SigLIP vision encoder that maps video frames to

$x_i=[x_i^1,\dots,x_i^t]$ 8

and a shared SigLIP text encoder that produces pooled and token-level instruction embeddings $x_i=[x_i^1,\dots,x_i^t]$ 9 and $O(n^2)$ 0. Compression then occurs at two levels. In the local branch, video tokens are partitioned into 3D groups controlled by $O(n^2)$ 1; each group is pooled to $O(n^2)$ 2 and compressed through instruction-conditioned attention,

$O(n^2)$ 3

In the global branch, learnable tokens $O(n^2)$ 4, with $O(n^2)$ 5, attend over the full video with 3D positional embeddings,

$O(n^2)$ 6

The paper studies three injection mechanisms—direct injection, coarse injection based on LayerNorm with scale and shift, and fine injection via cross-attention—and selects direct injection for the local branch and coarse injection for the global branch.

A distinctive training contribution is the conditional pre-training stage on HICom-248K, a dataset of 248K video clips and 739K QA-style instruction-followed descriptions, with 25.67 s average video length, sourced from Panda-70M and Ego4D. This stage is inserted between image-text alignment and instruction tuning, and is specifically designed to train instruction-conditioned compression rather than generic projector alignment.

The main empirical claim is that HICom can save 78.8% tokens relative to LLaVA-Video-7B while improving average accuracy by 2.43% on VideoMME, MVBench, and EgoSchema. Concretely, HICom-7B with 64 frames and 1328 tokens outperforms LLaVA-Video-7B with 32 frames and 6272 tokens. With 128 frames and 2624 tokens, it reports +3% average improvement over LLaVA-Video while still saving 58.2% tokens. On the systems side, the reported total inference time for 32 frames is 137.7 ms for HICom-7B versus 567 ms for LLaVA-OV-7B, with the LLM stage dropping from 553.7 ms to 102.7 ms (Liu et al., 20 Mar 2025).

A related transportation usage is HCOMC, the Hierarchical Cooperative On-Ramp Merging Control framework in "HCOMC: A Hierarchical Cooperative On-Ramp Merging Control Framework in Mixed Traffic Environment on Two-Lane Highways" (Wang et al., 15 Jul 2025). Although the acronym differs, the supplied material explicitly places it in the same conceptual space. HCOMC layers an IDM-based heterogeneous traffic flow model for HDVs and CAVs, a modified virtual vehicle model for hierarchical cooperative planning, a Stackelberg-game discretionary lane-changing model, and a multi-objective NSGA-II optimizer. Its objectives are safety, fuel economy, and efficiency, with safety measured through a critical-acceleration surrogate,

$O(n^2)$ 7

The reported simulations show that HCOMC improves critical distance, reduces stabilization time, lowers low-speed region volume, and often reduces fuel consumption relative to FIFO and game-theoretic baselines (Wang et al., 15 Jul 2025).

A second related usage is HICOM as holographic communications, represented in the supplied material by RHS-enabled HISAC in "Holographic Surface Enabled Integrated Sensing and Communications" (Zhang et al., 9 May 2026). Here the acronym refers to communication and sensing through reconfigurable holographic surfaces rather than conventional phased arrays. The key physical distinction is the leakage power constraint induced by serial feeding: $O(n^2)$ 8 which couples aperture amplitudes along each row. The tutorial describes holographic beamforming, joint communication-sensing optimization, and implementations supporting downlink communication, radar sensing, and JCAS. The supplied results include a DVB-T real-time 1080p video transmission experiment with ~30 dB IF SNR and ~5 W total RHS power, range estimation up to 20 m with ≤10 cm error, and JCAS operation with 5 Mbit/s communication while recovering target ranges (Zhang et al., 9 May 2026).

7. Cross-domain structure and technical distinctions

Taken together, these usages suggest that the acronymic family is less about a shared application area than about a recurring design strategy: replacing flat processing with a structured intermediate layer. In graph learning, the structure is a sampled multi-hop neighborhood hierarchy. In deepfake detection, it is a four-cue modular decomposition derived from human studies. In 3D Gaussian Splatting, it is a hierarchy of non-empty spatial regions with shared motion parameters. In video MLLMs, it is a hybrid local-global compression lattice conditioned on the instruction. In traffic control, it is a planning–game–control stack. In holographic communications, it is an aperture-level optimization problem constrained by feed physics.

This suggests a common methodological theme: hierarchical bottlenecks are introduced when naive end-to-end processing is either computationally prohibitive or physically mismatched to the signal domain. A second plausible implication is that the acronym tends to appear where the bottleneck itself is made semantically or physically meaningful: summary tokens in HiCom, interpretable human cues in HICOM for deepfakes, regional motions in HiCoM, instruction-conditioned tokens in HICom, cooperation modes in HCOMC, and leakage-constrained surface patterns in holographic HICOM.

The same surface similarity can obscure decisive differences. HiCom in graph learning is an LLM compressor over raw text neighborhoods (Zhang et al., 2024); HICOM in deepfake detection is a modular forensic system with human studies and an explanatory LLM (Hu et al., 20 Jul 2025); HiCoM in neural rendering is a dynamic 3DGS updater (Gao et al., 2024); HICom in MLLMs is a token-compression module between SigLIP and Qwen2.5 (Liu et al., 20 Mar 2025). For technical reading, the capitalization and expansion are therefore not cosmetic details but essential disambiguators.