2.5pt

1 Introduction↩︎

The YOLO family of single-stage detectors [1], [2] continuously evolves through architectural refinements. YOLO26, released in January 2026 [2], adopts a non-maximum suppression (NMS)-free dual-head training strategy and simplifies the regression head. At the end of its backbone, YOLO26 employs the C2PSA block, which performs global feature aggregation through a position-sensitive self-attention (PSA) mechanism built on the cross-stage partial (CSP) topology [3]. While effective, self-attention [4], [5] scales quadratically with token count, constraining its use in lightweight edge deployments.

State space models (SSMs), notably Mamba [6], recently emerged as efficient sequence models with linear complexity through a selective scan mechanism, building on structured SSMs such as S4 [7]. Vision adaptations such as Vision Mamba [8] and VMamba [9] have shown that SSMs rival Transformer-based backbones on classification, and recent works integrate Mamba into the YOLO family [10]. To our knowledge, however, whether Mamba can directly replace the C2PSA block in YOLO26 has not been examined in the public literature; we therefore propose MambaPSA, a Mamba-based replacement that preserves the CSP wrapper. We additionally insert a bidirectional Vision Mamba (BiViM) module at three neck levels (P3, P4, P5) to identify effective positions for SSM-based feature aggregation. Our contributions are:

• We propose MambaPSA, a lightweight Mamba-based replacement for the C2PSA block of YOLO26.

• We conduct a placement study of SSM modules at P3, P4, and P5 of the neck, revealing a non-monotonic relationship between placement and accuracy.

• Empirical evaluation on PASCAL VOC shows that MambaPSA reduces FLOPs by 12.1% and improves CPU inference throughput by 17.6% with negligible accuracy degradation, while P4 BiViM yields the best accuracy gain.

2 Related Work↩︎

Mamba [6] introduces selective scanning that enables SSMs to model long sequences in linear time, building on S4 [7]. Vision Mamba [8] and VMamba [9] extend SSMs to image data via bidirectional or four-directional scan strategies. Several recent works integrate SSMs into YOLO detectors: Mamba-YOLO [10] replaces the C2f block in the PAFPN neck of YOLOv8, and subsequent works target the C3k2 block in YOLO11. The integration of SSMs with the C2PSA block of YOLO26—an NMS-free framework—remains, to our knowledge, unexplored.

3 Method↩︎

Fig. 1 illustrates the overall YOLO26 pipeline with the two SSM integration points, and Fig. 2 details the internal structure of the two proposed blocks, MambaPSA and BiViM.

3.1 MambaPSA↩︎

MambaPSA preserves the CSP wrapper of C2PSA but replaces the internal self-attention branch with a Mamba core (Fig. 2, left). Given an input \(\mathbf{x} \in \mathbb{R}^{C \times H \times W}\), a \(1\times1\) convolution projects \(\mathbf{x}\) to \(2C'\) channels split into halves \(\mathbf{a}\) and \(\mathbf{b}\). The first half \(\mathbf{a}\) is flattened into \(HW\) tokens, processed by a single Mamba block [6], and reshaped back; the second half \(\mathbf{b}\) is preserved through identity. The two halves are concatenated and projected back to \(C\) channels by a second \(1\times1\) convolution. The Mamba core is configured as lightweight (\(d_{\text{state}}{=}8\), \(e{=}1\), mono-directional scan) so that the overall block is approximately parameter-neutral relative to C2PSA.

3.2 Mamba in the Neck↩︎

For the placement study, a bidirectional Vision Mamba (BiViM) block is applied immediately after the upsample-concat-C3k2 stage at the P3, P4, or P5 level (Fig. 2, right). The block scans the flattened token sequence in both directions, sums the two outputs, applies a linear projection, and adds a residual connection (\(d_{\text{state}}{=}16\), \(e{=}2\)).

4 Experiments↩︎

4.1 Setup↩︎

We evaluate on PASCAL VOC. The training set combines VOC 2007 and 2012 trainval (16,551 images); evaluation is on VOC 2007 test (4,952 images, 20 classes) [11]. All models are trained for 100 epochs using AdamW (lr \(10^{-3}\), batch 32, image size \(640{\times}640\)) with a 5-epoch warmup, last-15-epoch mosaic close, and mixed-precision (fp16). Training is on a single NVIDIA RTX 4090; parameters are reported post-fusion. CPU throughput is measured on an Intel Core i7 workstation CPU (batch size 1, single run).

4.2 Main Results↩︎

Table 2 compares the YOLO26-Nano baseline, three placement variants, and the proposed MambaPSA. MambaPSA reduces parameters by 2.9% and FLOPs by 12.1% relative to the baseline with only a 0.1 mAP\(_{50:95}\) decrease. Among the placement variants, P4 yields the highest accuracy gain (\(+0.9\)), P3 degrades accuracy substantially (\(-1.5\)), and P5 provides moderate improvement (\(+0.7\)) at a parameter overhead of \(+43.8\%\). In terms of parameter cost per \(1\%\) relative mAP\(_{50:95}\) change (relative parameter change divided by relative mAP change), P4 is the most efficient (\(+5.3\%\)), P5 is over five times more expensive (\(+31.3\%\)), whereas MambaPSA reduces parameters rather than expending them (\(-14.5\%\)), placing it on the opposite side of the efficiency–accuracy trade-off.

On CPU inference, MambaPSA reaches 20 FPS versus the baseline’s 17 FPS, a 17.6% throughput improvement broadly consistent with the FLOP reduction. Precise mean \(\pm\) std across multiple runs, alternative deployment runtimes (ONNX Runtime, OpenVINO), and edge-device profiling are deferred to future work.

2pt

4.3 Per-class Analysis↩︎

Table 1 reports per-class mAP\(_{50:95}\) across the 20 PASCAL VOC categories. The P4 variant attains the best accuracy on 9 categories, including large objects (bus +2.1, train +2.4, aeroplane +2.5) and small objects (bottle +0.7, pottedplant +2.2), supporting its role as a balanced sweet spot. The P5 variant excels on categories with rich global context (bird +2.9, horse +1.8, sofa +2.3) but degrades on small objects such as pottedplant (\(-\)​1.4). The P3 variant regresses on nearly all categories, suggesting that one-dimensional scanning over high-resolution feature maps may disrupt the local spatial structure required for fine-grained detection. MambaPSA exhibits a category-dependent trade-off with gains on animals and large vehicles (bird +2.5, bus +1.8, sheep +1.8) and losses on indoor and elongated structures (diningtable \(-\)​3.2, boat \(-\)​2.1, train \(-\)​1.9).

5 Discussion and Conclusion↩︎

We propose MambaPSA, a lightweight Mamba-based replacement for the C2PSA block in YOLO26, achieving \(\sim\)​12% FLOP reduction and 17.6% CPU throughput improvement with negligible accuracy impact on PASCAL VOC. The placement study identifies P4 as a favorable position for SSM-based aggregation. All results reflect a single training seed; inter-seed variance on VOC-nano scale can approach the magnitudes of the per-variant differences reported here, so the observed gaps should be interpreted as directional trends. Future work will address multi-seed stability, the combination of MambaPSA with the P4 BiViM placement, precise CPU/edge-device profiling across deployment runtimes, and MS COCO validation.

Acknowledgements↩︎

This research work is partially supported by National Science and Technology Council, Taiwan, under grant number 114-2221-E-032-011.

References↩︎