Mutual Feature Refinement Module (MFRM)

• MFRM is a module that mutually refines feature representations by leveraging intra-modal self-attention and learnable inter-modal fusion for enhanced discrimination.
• In multispectral object detection, MFRM aligns RGB and thermal features, contributing to improved detection metrics such as a +1.7 mAP gain.
• In CTR prediction, MFRM refines embedding vectors through multi-head self-attention and gating, yielding consistent AUC improvements with minimal overhead.

Searching arXiv for the cited papers and closely related context. Mutual Feature Refinement Module (MFRM) is an acronym used for distinct feature-refinement mechanisms in recent arXiv literature. In multispectral object detection, MFRM denotes the first half of the IRDFusion core, where a visible and a thermal feature-map are refined through intra-modal self-attention and a learnable inter-modal Value-fusion step to obtain feature representations that are better aligned across modalities and more discriminative (Shen et al., 11 Sep 2025). In click-through rate (CTR) prediction, the same acronym denotes a feature refinement module placed between embedding and interaction layers, where feature embeddings are mutually refined through attention and gating on a per-instance basis (Wang et al., 2023). The shared name reflects a common emphasis on context-dependent refinement, but the two modules arise in different tasks, use different inputs, and are optimized under different downstream objectives.

1. Terminological scope and core definition

The acronym MFRM is not unique to a single architecture. In "IRDFusion: Iterative Relation-Map Difference guided Feature Fusion for Multispectral Object Detection" (Shen et al., 11 Sep 2025), the Mutual Feature Refinement Module operates on a pair of feature-maps from the visible (RGB) branch and the thermal (IR) branch. Its stated purpose is to take those two inputs and, by means of intra-modal self-attention plus a small inter-modal Value-fusion step, produce two enhanced feature-maps that are better aligned across modalities and more discriminative.

In "A Comprehensive Summarization and Evaluation of Feature Refinement Modules for CTR Prediction" (Wang et al., 2023), MFRM is defined in a different sense. There it is a feature refinement module for CTR models, motivated by the claim that fixed feature embeddings learned only through the embedding layer limit performance. The module lets features mutually refine one another on a per-instance basis so that the refined embedding of field $i$ depends not only on itself but also on other co-occurring features.

A useful distinction is therefore between a cross-modal MFRM in multispectral detection and a feature-embedding MFRM in CTR prediction. This is an editor's term introduced only to disambiguate the two uses of the same acronym.

2. MFRM in multispectral object detection

In IRDFusion, MFRM is described as the first operation inside each iteration of the iterative fusion loop (Shen et al., 11 Sep 2025). Its inputs are $F_v \in \mathbb{R}^{N\times d}$, the flattened and positional-encoded RGB feature-map, and $F_t \in \mathbb{R}^{N\times d}$, the corresponding IR feature-map. Each $F_i$, for $i\in\{v,t\}$, is linearly projected into Query, Key, and Value using learnable matrices $W_i^q$, $W_i^k$, and $W_i^v \in \mathbb{R}^{d\times d}$.

The module first computes intra-modal self-attention independently for each modality. The visible branch forms $A_v$, sometimes written $A_{v2v}$, as an $F_v \in \mathbb{R}^{N\times d}$0 attention map encoding pairwise affinities among visible pixels; the thermal branch forms $F_v \in \mathbb{R}^{N\times d}$1 analogously for thermal pixels. The module then performs inter-modal Value fusion rather than cross-attention at the relation-map stage. Specifically, each Value tensor is augmented by a learnable fraction of the other modality’s Value tensor: $F_v \in \mathbb{R}^{N\times d}$2 The two fusion weights $F_v \in \mathbb{R}^{N\times d}$3 and $F_v \in \mathbb{R}^{N\times d}$4 are learnable.

The refined outputs are then obtained by applying the intra-modal relation maps to the fused Values: $F_v \in \mathbb{R}^{N\times d}$5 These $F_v \in \mathbb{R}^{N\times d}$6 and $F_v \in \mathbb{R}^{N\times d}$7 are the refined, cross-modally aligned features that leave the module. The paper notes that a small feed-forward block, LayerNorm, or Dropout can optionally follow each $F_v \in \mathbb{R}^{N\times d}$8 in a Transformer pre-norm style, but also states that the core is the two equations above (Shen et al., 11 Sep 2025).

The design is explicitly characterized as using single-branch self-attention to build $F_v \in \mathbb{R}^{N\times d}$9 and $F_t \in \mathbb{R}^{N\times d}$0, with cross-modal coupling introduced only after those relation maps are computed. In the formulation provided in the paper summary, one can view $F_t \in \mathbb{R}^{N\times d}$1 as a "where" map over visible pixels, after which thermal content from $F_t \in \mathbb{R}^{N\times d}$2 is borrowed through the fused Value term.

3. Mathematical formulation and relation-map construction

The mathematical formulation of the multispectral MFRM is given in equations summarized from §3.2 of the IRDFusion paper (Shen et al., 11 Sep 2025): $F_t \in \mathbb{R}^{N\times d}$3

The inter-modal Value-fusion step is

$F_t \in \mathbb{R}^{N\times d}$4

with learnable fusion weights further specified as

$F_t \in \mathbb{R}^{N\times d}$5

The output equation is

$F_t \in \mathbb{R}^{N\times d}$6

The accompanying explanation in the source distinguishes the role of each component. $F_t \in \mathbb{R}^{N\times d}$7 is the intra-modal relational map, or self-attention, with size $F_t \in \mathbb{R}^{N\times d}$8. The cross-modal component is confined to the $F_t \in \mathbb{R}^{N\times d}$9-fusion term, where a small amount of the other modality’s Value feature is added. The learnable parameters $F_i$0, together with $F_i$1, allow the network to adaptively scale how much IR enters the RGB branch and vice versa (Shen et al., 11 Sep 2025).

This construction differs from a direct cross-attention formulation. The relation maps remain modality-specific, while the Value tensors carry the cross-modal transfer. A plausible implication is that the module attempts to preserve modality-specific spatial affinity structure while introducing complementary signal only at the content aggregation stage.

4. Position within IRDFusion and optimization behavior

MFRM is not a standalone model; it is embedded in the iterative mechanism termed Iterative Relation-Map Differential Guided Feature Fusion, or IRDFusion (Shen et al., 11 Sep 2025). At iteration $F_i$2, the inputs $F_i$3 pass through MFRM to produce $F_i$4. The Differential Feature Feedback Module (DFFM) then computes cross-modal difference features from those refined outputs and feeds them back to update the features for iteration $F_i$5.

The summary gives the visible-branch update as

$F_i$6

followed by

$F_i$7

The same process occurs symmetrically for the thermal branch. The updated $F_i$8 become the next inputs to MFRM (Shen et al., 11 Sep 2025).

The training objective is likewise defined at the full detector level rather than at the module level. The source states that there is no standalone MFRM loss. Instead, MFRM is trained end-to-end under the standard CoDETR detection loss, which includes Quality Focal Loss for classification plus $F_i$9 and GIoU for bounding-box regression, with auxiliary heads in the RPN and ROI. According to the paper summary, MFRM’s intended effects—sharper contours and better alignment—appear in the final detection metrics rather than through a bespoke auxiliary loss inside the module.

The reported hyperparameter choices further delimit the module’s intended operating regime. The design uses learnable fusion weights $i\in\{v,t\}$0 initialized by a small constant $i\in\{v,t\}$1, exemplified as $i\in\{v,t\}$2. The paper uses a single-head design with $i\in\{v,t\}$3 chosen to match the backbone’s $i\in\{v,t\}$4-dimensional output, exemplified as $i\in\{v,t\}$5. The iteration count $i\in\{v,t\}$6 is cross-validated, and the authors report that four iterations are optimal in their ablation study (Shen et al., 11 Sep 2025).

5. Empirical evidence and stated effects in IRDFusion

The IRDFusion summary reports ablation results on the multispectral detection setting that isolate the contribution of MFRM and DFFM (Shen et al., 11 Sep 2025).

Configuration	Reported metrics
baseline (no MFRM/DFFM)	mAP50=84.8, mAP75=44.0, mAP=46.9
+ MFRM only	mAP50=86.3 (+1.5), mAP75=46.0 (+2.0), mAP=48.6 (+1.7)
+ DFFM only	mAP50=87.5 (+2.7), mAP75=46.0 (+2.0), mAP=49.2 (+2.3)
MFRM + DFFM	mAP50=88.3 (+3.5), mAP75=48.0 (+4.0), mAP=50.7 (+3.8)

The same source states that, from these numbers, MFRM by itself contributes approximately $i\in\{v,t\}$7 mAP and does so primarily by sharpening weakly represented classes such as bicycle. It also states that the interaction of MFRM with DFFM yields the largest overall gain. The full IRDFusion framework is reported to achieve state-of-the-art performance on FLIR, LLVIP, and M$i\in\{v,t\}$8FD datasets and to outperform existing methods across diverse challenging scenarios (Shen et al., 11 Sep 2025).

Within that framework, MFRM is explicitly described as a lightweight Transformer-style block whose novelty is twofold: it computes strong intra-modal relation maps $i\in\{v,t\}$9 and $W_i^q$0, and then borrows a learnable fraction of the other modality’s Value vectors. The source further characterizes the resulting outputs as more tightly aligned and more discriminative, serving as the foundation for the subsequent DFFM.

A common misunderstanding would be to treat MFRM as a generic cross-attention module. In the IRDFusion formulation, that is not the stated mechanism. The relation maps are computed intra-modally, while inter-modal transfer occurs in the Value-fusion term. Another misunderstanding would be to expect an auxiliary refinement loss; the source is explicit that no standalone MFRM loss is used (Shen et al., 11 Sep 2025).

6. MFRM in CTR prediction

In the CTR literature summarized in (Wang et al., 2023), MFRM refers to a different architecture and problem formulation. The module is defined for the embedding-to-interaction stage of CTR models, where the input is a stack of $W_i^q$1 feature embeddings

$W_i^q$2

and the output is

$W_i^q$3

Its motivation is that, in classic CTR models, each categorical feature is embedded once, so the same feature always has the same embedding regardless of the other co-occurring features in a particular impression. MFRM remedies this by allowing every pair or set of features to mutually refine each other’s embeddings on a per-instance basis.

The architecture is described as a two-branch mutual refinement block composed of a multi-head self-attention submodule followed by a gating and residual submodule. Queries, Keys, and Values are linear projections of $W_i^q$4: $W_i^q$5 For head $W_i^q$6, the attention map and contextual output are

$W_i^q$7

The heads are concatenated and projected back: $W_i^q$8 A per-feature gating vector is then generated: $W_i^q$9 and the final refined embedding is

$W_i^k$0

The paper summary also gives a compact pairwise interpretation: $W_i^k$1 It then notes that the attention-based construction handles all $W_i^k$2 features in parallel. The claimed outcome is a fully context-aware, bit-level, non-linear representation that can be inserted into existing CTR backbones (Wang et al., 2023).

Empirically, the summary reports consistent AUC gains of $W_i^k$3 on Criteo and Frappe when MFRM is inserted between the embedding layer and the downstream feature interaction and prediction tower. It also reports that MFRM outperforms simpler vector-gating or field-only methods such as FEN, SENET, and FWN; incurs extra parameters of $W_i^k$4; adds approximately $W_i^k$5 to $W_i^k$6 seconds per epoch; converges in fewer epochs; and, in parallel CTR models such as DeepFM’s FM-tower plus DNN-tower, yields a further approximately $W_i^k$7 AUC lift when two independent MFRMs are assigned to separate sub-networks rather than sharing a single MFRM (Wang et al., 2023).

7. Cross-domain comparison and interpretive significance

The two MFRM formulations share the language of mutual refinement but differ in what is being refined, how mutuality is implemented, and where the module is inserted in the model stack.

Setting	Inputs	Mutuality mechanism
Multispectral detection (Shen et al., 11 Sep 2025)	Visible and thermal feature-maps $W_i^k$8	Intra-modal relation maps plus learnable inter-modal Value fusion
CTR prediction (Wang et al., 2023)	Feature embedding stack $W_i^k$9	Multi-head self-attention plus mutual gating and residual mixing

In the multispectral case, mutuality is defined across sensing modalities. Each branch computes its own self-attention map, and complementary information is injected through the fused Value tensor. In the CTR case, mutuality is defined across feature fields within a single instance, and the refinement is expressed through attention over the feature set followed by gating at the embedding level.

This comparison helps delimit the meaning of the acronym. MFRM does not designate a single canonical block across machine learning. Rather, it denotes a family of modules whose common theme is refinement through reciprocal contextual influence. The specific mathematical form, training objective, and empirical role depend on the application domain. In IRDFusion, MFRM is tied to iterative cross-modal fusion and the downstream CoDETR loss (Shen et al., 11 Sep 2025). In CTR prediction, it is tied to feature embedding refinement prior to feature interaction and is evaluated through AUC and computational overhead metrics (Wang et al., 2023).

A plausible implication is that the recurrence of the name reflects a broader architectural pattern: refinement is treated not as a one-shot preprocessing step but as an adaptive operation in which each representation is modified in light of another representation or of the full feature set. The supplied literature, however, also shows that identical nomenclature should not be taken as evidence of architectural equivalence.