Guided Module (GM) in Machine Learning

• Guided Module (GM) is a design philosophy that uses auxiliary signals to alter inference paths in machine learning models.
• GM approaches span diverse applications including video captioning, semantic segmentation, and sensorimotor control with tailored contextual cues.
• Implementations range from training-only modules to inference-time enhancements and hybrid solvers, illustrating modular integration in complex architectures.

“Guided Module” (GM) does not denote a single canonical architecture in the literature. Instead, it refers to a family of modules that inject auxiliary structure, supervisory signals, or external context into a base computational process. Across the cited work, GM-like designs guide prediction with past and future words in encoder–decoder video captioning, modulate selective scan propagation with boundary and centripetal-flow cues in semantic segmentation, use multi-scale outputs to steer attention in salient object detection, translate language-level intentions into sensorimotor control in hierarchical action systems, and couple complementary solvers in scientific simulation and tracking (Zhang et al., 2020, Chan et al., 7 Jun 2026, Zhu et al., 2019, Qi, 2020, Izquierdo et al., 2023, He et al., 2023). A unifying property is that the module does not merely add capacity; it constrains or redirects computation using signals unavailable to the unguided core model.

1. Conceptual scope and defining characteristics

In the machine learning literature represented here, a GM is typically an auxiliary or embedded component that alters how a host model aggregates context, selects features, or executes a sequential process. The guidance signal may be linguistic, geometric, visual, probabilistic, or derived from another solver. This suggests that “guidance” is best understood functionally: it changes the trajectory of inference by introducing a structured prior or an additional decision pathway.

Several recurring patterns appear. First, many GM designs are training-time auxiliaries that improve a primary model without being retained at inference. The guidance module in GMNet for video captioning is attached to a standard encoder–decoder during training only; at test time, inference is done using the standard encoder–decoder whose parameters were influenced by the GM (Zhang et al., 2020). Second, some GMs are drop-in replacements for standard context heads. Geometry-Guided Mamba (G-Mamba) is studied as a plug-and-play context aggregation head that can replace ASPP, PPM, or attention heads in CNN-based semantic segmentation while leaving the backbone and other pipeline components unchanged (Chan et al., 7 Jun 2026). Third, some modules guide internal attention or feature reuse using outputs or parallel branches rather than relying solely on the current feature tensor, as in the Output-Guided Attention Module (OGAM) of OGNet and the two-level merging strategy of GM-Net (Zhu et al., 2019, Chen et al., 2017).

A common misconception is to treat GM as a fixed algorithmic template. The cited work does not support that view. The term spans modules for token-sequence conditioning, dense prediction, multimodal reasoning, sensorimotor planning, hybrid transport solvers, and multi-model target tracking. What remains consistent is the insertion of an explicit guiding signal into an otherwise standard pipeline.

2. Sequence guidance in encoder–decoder captioning

A clear, early formulation of a guidance module appears in video captioning. The paper “Guidance Module Network for Video Captioning” introduces a guidance module to address a limitation of teacher-forced encoder–decoder training: the model makes the prediction probability of each word close to a $0$-$1$ distribution and ignores other words (Zhang et al., 2020). The GM explicitly incorporates both past decoded words and future groundtruth words during training.

At decoding timestep $i$, the module processes the already-decoded words $\{y_1,\ldots,y_{i-1}\}$ and the future groundtruth words $\{y_{i+1}^*,\ldots,y_I^*\}$ using two LSTM encoders. If $\mathbf{A}_p$ and $\mathbf{A}_f$ denote the hidden states for past and future words, linear transformations and dimensionality reduction produce

$$\mathbf{A}_{e} = RD(\mathbf{W}_{p} \mathbf{A}_{p}) + RD(\mathbf{W}_{f} \mathbf{A}_{f}).$$

This representation is fused with attention-based visual features $\mathbf{A}_{att}$ using layer normalization,

$$\mathbf{A}_{F} = \mathrm{Norm}\left(\mathrm{Norm}(\mathbf{A}_e) + \mathrm{Norm}(\mathbf{A}_{att})\right),$$

and a parallel LSTM decoder then generates a guiding caption distribution (Zhang et al., 2020).

The loss is explicitly additive: $1$0 This design makes the GM an auxiliary supervisory pathway rather than a replacement decoder. Its technical role is to encourage the model to generate words related to the past and future words in a caption, thereby increasing sentence coherence.

On MSVD, the reported ablation compares SA, SA_LN, and GMNet. GMNet achieves BLEU-4 $1$1, METEOR $1$2, ROUGE_L $1$3, and CIDEr $1$4, compared with SA_LN at BLEU-4 $1$5, METEOR $1$6, ROUGE_L $1$7, and CIDEr $1$8 (Zhang et al., 2020). The gains are modest but targeted: BLEU-4 increases by $1$9, ROUGE_L by $i$0, and CIDEr by $i$1, while METEOR remains unchanged. Within this formulation, guidance is fundamentally bidirectional sequence conditioning applied during training.

3. Guided context aggregation in visual dense prediction

In dense prediction, guidance modules are used to make context propagation structure-aware. The most explicit example is G-Mamba for CNN-based semantic segmentation. The module is adapted from the Directional Geometric Mamba block in DGM-Net and repurposed as a modular context head that can replace ASPP, PPM, or attention heads in DeepLabV3+, DANet, CCNet, PSPNet, PSANet, and OCRNet, with the ResNet-101 backbone unchanged (Chan et al., 7 Jun 2026).

Given an intermediate feature map $i$2, G-Mamba predicts three geometric priors: a centripetal potential map $i$3, a directional flow field $i$4, and a coarse morphological boundary map $i$5. A residual branch refines the boundary map: $i$6 For each scan direction $i$7, the projected flow $i$8 produces a scan-aware spatial prompt

$i$9

and the input to the selective scan is reweighted as

$\{y_1,\ldots,y_{i-1}\}$0

This geometry-guided mechanism modulates long-range propagation using object boundary information and centripetal-flow directionality, with the stated aim of enhancing boundaries and suppressing cross-object leakage (Chan et al., 7 Jun 2026).

The paper also defines Cascade G-Mamba. Early stages use isotropic multi-directional scans to build macro context, and the final stage applies geometry-guided scan to refine details. This separates context from precision. On Cityscapes, with ResNet-101, output stride $\{y_1,\ldots,y_{i-1}\}$1, input crop $\{y_1,\ldots,y_{i-1}\}$2, and evaluation at up to $\{y_1,\ldots,y_{i-1}\}$3, the module yields consistent mIoU gains under a single RTX 3080 Ti setting (Chan et al., 7 Jun 2026).

Architecture	Baseline mIoU	+G-Mamba mIoU
DeepLabV3+	75.17	76.58
DANet	76.32	78.49
CCNet	75.54	77.83
PSPNet	77.52	78.76
PSANet	76.79	77.02
OCRNet	78.22	79.01

Cascade G-Mamba further improves these numbers to $\{y_1,\ldots,y_{i-1}\}$4, $\{y_1,\ldots,y_{i-1}\}$5, $\{y_1,\ldots,y_{i-1}\}$6, $\{y_1,\ldots,y_{i-1}\}$7, $\{y_1,\ldots,y_{i-1}\}$8, and $\{y_1,\ldots,y_{i-1}\}$9, respectively (Chan et al., 7 Jun 2026). The GFLOPs overhead is reported as moderate, approximately $\{y_{i+1}^*,\ldots,y_I^*\}$0–$\{y_{i+1}^*,\ldots,y_I^*\}$1 GFLOPs relative to total model cost.

A related but distinct visual guidance mechanism appears in salient object detection. OGNet argues that self-attention modules that take only the processed feature map as input may amplify early errors, a failure mode termed “blind overconfidence” (Zhu et al., 2019). Its Output-Guided Attention Module uses multi-scale outputs from deeper decoder layers to guide attention. For an input feature map $\{y_{i+1}^*,\ldots,y_I^*\}$2, channel attention is

$\{y_{i+1}^*,\ldots,y_I^*\}$3

while spatial attention combines pooled current features with deeper outputs: $\{y_{i+1}^*,\ldots,y_I^*\}$4 The final attentive output is

$\{y_{i+1}^*,\ldots,y_I^*\}$5

On SOD and DUTS, the ablation in Table 2 shows MAE improving from $\{y_{i+1}^*,\ldots,y_I^*\}$6 to $\{y_{i+1}^*,\ldots,y_I^*\}$7 and from $\{y_{i+1}^*,\ldots,y_I^*\}$8 to $\{y_{i+1}^*,\ldots,y_I^*\}$9 when replacing the baseline attention with OGAM (Zhu et al., 2019).

These visual modules share a central design principle: guidance is injected into spatial aggregation itself rather than only at the loss or decoder output.

4. Guided reasoning and action

A more general interpretation of GM appears in multimodal reasoning and embodied control. In multimodal mathematical reasoning, GM-PRM denotes a Generative Multimodal Process Reward Model. The model takes a mathematical problem $\mathbf{A}_p$0, an image $\mathbf{A}_p$1, and a step-by-step reasoning process $\mathbf{A}_p$2, and for each step produces fine-grained critiques of step intent, image alignment, reasoning logic, and a final judgment (Zhang et al., 6 Aug 2025). Unlike binary verifiers, it also generates a corrected version of the first erroneous step it identifies. The evaluation mapping is given as

$\mathbf{A}_p$3

and the correction operator as

$\mathbf{A}_p$4

This corrective capacity supports Refined Best-of-$\mathbf{A}_p$5 (Refined-BoN): generate $\mathbf{A}_p$6 initial solutions, use GM-PRM to stop at the first erroneous step and refine it, continue generation from the corrected partial solution, then select among the combined $\mathbf{A}_p$7 candidates (Zhang et al., 6 Aug 2025). The reported training set contains approximately $\mathbf{A}_p$8 samples, and the paper states that the framework achieves state-of-the-art results on five multimodal math benchmarks while requiring only a $\mathbf{A}_p$9K-sample training dataset.

A broader action-oriented notion of guided modules appears in Language Guided Machine Action. LGMA is a hierarchical modular network with three systems: a primary sensory system, an association system, and a high-level executive system (Qi, 2020). The primary sensory system includes a visual autoencoder, a somatosensorimotor autoencoder, and a language autoencoder producing visual, somatosensorimotor, and language vectors. The association system includes Wernicke and Broca modules for language comprehension and synthesis, BA14/40 for translation between sensorimotor and language, a midTemporal module for conversion between language and vision, a superior parietal lobe module for integrating attended visual object and arm state into a cognitive map, pre-SMA for converting high-level intention into sequential atomic actions, and SMA for integrating atomic actions with current arm and attended object state into a sensorimotor vector. The executive system contains PFC for explicit inference and voluntary-action guidance based on language, and BG as the habitual action control center (Qi, 2020).

In both GM-PRM and LGMA, guidance is not limited to feature weighting. It governs sequential structure. In GM-PRM, the module evaluates and revises reasoning trajectories; in LGMA, language-level intentions are decomposed into action sequences and passed down a hierarchy. This suggests a broader category of GM in which the module acts as a collaborator or controller rather than merely a local attention block.

5. Guided hybridization in scientific computing and probabilistic tracking

Outside mainstream deep vision and language tasks, the term GM is also attached to hybrid mechanisms that coordinate complementary inference engines. In neutrino transport, the Guided Moments formalism runs truncated moments ($\mathbf{A}_f$0) and Monte Carlo ($\mathbf{A}_f$1) schemes in parallel (Izquierdo et al., 2023). The method uses MC to inform the pressure-tensor closure in optically thin regions and standard analytic closure in optically thick regions, while using the moment solution to guide or regularize MC sampling. The closure is interpolated as

$\mathbf{A}_f$2

with a switch function

$\mathbf{A}_f$3

Moment matching is enforced so that the lowest moments of the MC packets match those from the $\mathbf{A}_f$4 scheme. In the optically thick limit, the method recovers the exact $\mathbf{A}_f$5 closure; in the optically thin limit, it uses the $\mathbf{A}_f$6-based closure (Izquierdo et al., 2023). Comparative tests include diffusion in a moving medium, double beam, radiating and absorbing sphere, and radiating and absorbing torus, where the hybrid scheme matches $\mathbf{A}_f$7 in thick regions, preserves beam crossing like $\mathbf{A}_f$8 in thin regions, and suppresses statistical noise relative to pure MC.

In multi-target tracking, the interactive model fusion-based GM-PHD filter uses another guided hybridization strategy. Non-Gaussian measurement noise is approximated as a Gaussian mixture,

$\mathbf{A}_f$9

which induces a Gaussian-mixture measurement model

$$\mathbf{A}_{e} = RD(\mathbf{W}_{p} \mathbf{A}_{p}) + RD(\mathbf{W}_{f} \mathbf{A}_{f}).$$0

A GM-PHD filter is then run for each Gaussian component, and an interactive multi-model framework fuses the resulting target states (He et al., 2023). If $$\mathbf{A}_{e} = RD(\mathbf{W}_{p} \mathbf{A}_{p}) + RD(\mathbf{W}_{f} \mathbf{A}_{f}).$$1, the method reduces to the classical GM-PHD filter. In the reported two-dimensional MTT simulation, with true measurement noise $$\mathbf{A}_{e} = RD(\mathbf{W}_{p} \mathbf{A}_{p}) + RD(\mathbf{W}_{f} \mathbf{A}_{f}).$$2, clutter intensity $$\mathbf{A}_{e} = RD(\mathbf{W}_{p} \mathbf{A}_{p}) + RD(\mathbf{W}_{f} \mathbf{A}_{f}).$$3, and Gaussian process noise, the interactive fusion method yields lower OSPA than the classical GM-PHD filter under non-Gaussian noise (He et al., 2023).

These examples broaden the meaning of guidance. The guiding signal can be another solver’s moments, or the latent model identity associated with a Gaussian-mixture approximation. In both cases, guidance mediates between multiple inferential views of the same underlying process.

6. Guided feature reuse and modular efficiency

In CNN architecture design, GM can also refer to a structural motif for efficient feature learning. GM-Net, or Grouped Merging Net, uses grouped convolution in nearly every convolutional layer and combines it with Basic Units and a two-level merging strategy (Chen et al., 2017). The paper explicitly describes the “Guided Module” as the architectural pattern combining BUs and merging to guide efficient and discriminative feature reuse.

Two basic units are defined. BU_A uses dense connection via elementwise summation rather than concatenation, while BU_B uses straight mapping similar to residual connections. Grouped convolution reduces parameter count according to

$$\mathbf{A}_{e} = RD(\mathbf{W}_{p} \mathbf{A}_{p}) + RD(\mathbf{W}_{f} \mathbf{A}_{f}).$$4

where $$\mathbf{A}_{e} = RD(\mathbf{W}_{p} \mathbf{A}_{p}) + RD(\mathbf{W}_{f} \mathbf{A}_{f}).$$5 is kernel size, $$\mathbf{A}_{e} = RD(\mathbf{W}_{p} \mathbf{A}_{p}) + RD(\mathbf{W}_{f} \mathbf{A}_{f}).$$6 the input channels, $$\mathbf{A}_{e} = RD(\mathbf{W}_{p} \mathbf{A}_{p}) + RD(\mathbf{W}_{f} \mathbf{A}_{f}).$$7 the output channels, and $$\mathbf{A}_{e} = RD(\mathbf{W}_{p} \mathbf{A}_{p}) + RD(\mathbf{W}_{f} \mathbf{A}_{f}).$$8 the group number (Chen et al., 2017). To offset information loss from aggressive grouping, GM-Net introduces operation-level merging through non-grouped bottleneck convolutions and unit-level merging through dense summation or skip connections and by merging outputs from parallel units BU_f and BU_s.

The paper reports that removing operation-level merging drops accuracy by approximately $$\mathbf{A}_{e} = RD(\mathbf{W}_{p} \mathbf{A}_{p}) + RD(\mathbf{W}_{f} \mathbf{A}_{f}).$$9, while removing unit-level merging also hurts performance by approximately $\mathbf{A}_{att}$0 (Chen et al., 2017). On CIFAR-10+, CIFAR-100+, SVHN, and MNIST, GM-Net BU_A/BU_B with $\mathbf{A}_{att}$1M parameters obtains error rates $\mathbf{A}_{att}$2, $\mathbf{A}_{att}$3, $\mathbf{A}_{att}$4, and $\mathbf{A}_{att}$5, respectively. The same table lists FractalNet at $\mathbf{A}_{att}$6M parameters with CIFAR-10+ error $\mathbf{A}_{att}$7 and CIFAR-100+ error $\mathbf{A}_{att}$8, and DenseNet ($\mathbf{A}_{att}$9) at $$\mathbf{A}_{F} = \mathrm{Norm}\left(\mathrm{Norm}(\mathbf{A}_e) + \mathrm{Norm}(\mathbf{A}_{att})\right),$$0M parameters with CIFAR-10+ error $$\mathbf{A}_{F} = \mathrm{Norm}\left(\mathrm{Norm}(\mathbf{A}_e) + \mathrm{Norm}(\mathbf{A}_{att})\right),$$1, CIFAR-100+ error $$\mathbf{A}_{F} = \mathrm{Norm}\left(\mathrm{Norm}(\mathbf{A}_e) + \mathrm{Norm}(\mathbf{A}_{att})\right),$$2, and SVHN error $$\mathbf{A}_{F} = \mathrm{Norm}\left(\mathrm{Norm}(\mathbf{A}_e) + \mathrm{Norm}(\mathbf{A}_{att})\right),$$3 (Chen et al., 2017). The contribution of the GM here is therefore not an external semantic signal but a merge-guided reuse mechanism that preserves discriminability under parameter-efficient grouping.

A plausible implication is that “guidance” can operate at multiple abstraction levels: semantic guidance in captioning, geometric guidance in segmentation, process guidance in reasoning, and architectural guidance in feature reuse.

7. Comparative themes, limitations, and misconceptions

Across the cited literature, several themes recur. One is the use of auxiliary signals to correct a known weakness of a baseline. In video captioning, the weakness is local teacher-forced prediction ignoring future words (Zhang et al., 2020). In semantic segmentation, it is boundary leakage and the isotropic nature of conventional context heads (Chan et al., 7 Jun 2026). In salient object detection, it is blind overconfidence from self-attention guided only by the current feature map (Zhu et al., 2019). In multimodal reasoning, it is the inability of binary PRMs to correct reasoning errors (Zhang et al., 6 Aug 2025). In neutrino transport and non-Gaussian tracking, it is the regime-specific failure of a single solver or a single Gaussian measurement model (Izquierdo et al., 2023, He et al., 2023).

A second theme is selective deployment. Some GMs are inference-time modules, such as G-Mamba and OGAM. Others are training-only auxiliaries, such as the guidance module in video captioning. Others still define the full runtime control loop, as in GM-PRM’s Refined-BoN or the Guided Moments transport scheme. Consequently, GM should not be equated with attention, nor with a plug-and-play block, nor with multimodal conditioning alone.

A third theme is modular decomposition. The papers frequently isolate the guided component from the rest of the system in order to attribute gains. G-Mamba replaces only the context head while keeping the ResNet-101 backbone and training setup unchanged (Chan et al., 7 Jun 2026). GMNet in video captioning adds an auxiliary decoder and loss while leaving the standard encoder–decoder inference path intact (Zhang et al., 2020). GM-Net studies merging separately from grouped convolution (Chen et al., 2017). This experimental pattern indicates that GM is often conceived as a controllable insertion point rather than an end-to-end redesign.

The main limitation of the term is semantic ambiguity. “GM” names modules with substantially different mathematical objects: LSTM-encoded past/future context, geometry-derived scan prompts, multi-scale decoder outputs, textual critiques and corrected steps, cortical-style subsystems, hybrid closures, and interactive model probabilities. Any encyclopedia treatment must therefore distinguish the specific instantiation from the generic label. In the cited work, Guided Module is best understood as a design philosophy: guide a base model with structured side information so that computation becomes more coherent, more selective, or more robust than the unguided baseline.