Pith. sign in

REVIEW 3 major objections 3 minor

A Mamba state-space block can replace YOLO26’s final attention module, cutting parameters and FLOPs while matching PASCAL VOC accuracy and speeding CPU inference.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-15 04:11 UTC pith:D5EB4X2Y

load-bearing objection Abstract-only engineering swap of Mamba for C2PSA in YOLO26; concrete VOC deltas look useful but cannot yet support the general claim. the 3 major comments →

arxiv 2607.12681 v1 pith:D5EB4X2Y submitted 2026-07-14 cs.CV

MambaPSA: A Mamba-based Replacement for C2PSA in YOLO26

Sheng-Wei Chan , Chia-Min Lin , Hsin-Jui Pan , Ching-Yu Tsai , Chih-Hsiang Yang , Yung-Che Wang , Jen-Shiun Chiang This is my paper
classification cs.CV
keywords Mambastate space modelsYOLO26object detectionC2PSAVision MambaNMS-free detectorsefficient inference
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper aims to show that a lightweight Mamba state-space module (MambaPSA) can stand in for the C2PSA attention block at the end of the YOLO26 backbone without meaningful accuracy loss, while shrinking the model and speeding inference. On PASCAL VOC 2007+2012 the swap cuts parameters by 2.9 percent, FLOPs by 12.1 percent, and raises CPU throughput from 17 to 20 FPS, with only a 0.1 drop in mAP50:95. The authors also test a bidirectional Vision Mamba module inserted at the neck’s P3, P4, or P5 levels and find the P4 placement gives the largest accuracy lift (+0.9 mAP50:95). If the result holds, NMS-free lightweight detectors can trade quadratic attention for linear-complexity state-space layers and run faster on constrained hardware at nearly the same detection quality. A reader who cares about real-time edge detection would see this as a practical efficiency lever that does not force a redesign of the whole detector.

Core claim

MambaPSA, a Mamba-based replacement for the C2PSA block at the end of the YOLO26 backbone, reduces parameters by 2.9 percent, FLOPs by 12.1 percent, and improves CPU inference throughput by 17.6 percent (17 to 20 FPS) with a negligible accuracy change of −0.1 mAP50:95 on PASCAL VOC 2007+2012. Separately, inserting a bidirectional Vision Mamba module at the P4 neck level yields the best accuracy gain of +0.9 mAP50:95. The authors conclude that state-space models offer a favorable efficiency–accuracy trade-off when they replace attention-based blocks in NMS-free lightweight detectors.

What carries the argument

MambaPSA—a lightweight Mamba state-space module that replaces the C2PSA attention block at the backbone terminus of YOLO26—supplies linear-complexity sequence modeling in place of quadratic self-attention while preserving the rest of the NMS-free detection head.

Load-bearing premise

The claim that state-space models are a generally favorable drop-in for attention blocks rests on single-run PASCAL VOC 2007+2012 numbers without multi-seed variance or results on larger modern benchmarks such as COCO.

What would settle it

Train the YOLO26 baseline and the MambaPSA variant under an identical protocol on COCO and check whether the reported parameter and FLOP savings still leave mAP50:95 within roughly 0.2 points of the baseline; a substantially larger accuracy drop would falsify the favorable trade-off claim.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • NMS-free YOLO-style detectors can replace the final backbone attention block with a Mamba module and gain CPU speed without meaningful mAP loss on VOC-scale data.
  • A bidirectional Vision Mamba insertion at the mid-resolution (P4) neck level can raise accuracy more than the backbone-only swap.
  • Edge detectors can reclaim roughly 3 percent of parameters and 12 percent of FLOPs by this single late-stage substitution.
  • Linear-complexity state-space layers become a practical drop-in candidate for late attention stages in one-stage detectors.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar Mamba substitutions earlier in the backbone might compound FLOP savings if receptive-field requirements still hold.
  • The near-zero accuracy cost observed on PASCAL VOC may shrink or reverse on denser, larger-scale benchmarks such as COCO where long-range attention could matter more.
  • The measured CPU throughput gain points to particular value for GPU-free edge or mobile deployments.
  • A joint configuration that pairs MambaPSA with the best BiViM neck placement is a natural next experiment for simultaneous efficiency and accuracy gains.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 3 minor

Summary. The manuscript proposes MambaPSA, a lightweight Mamba-based replacement for the C2PSA block at the end of the YOLO26 backbone, and additionally inserts bidirectional Vision Mamba (BiViM) modules at the P3, P4 and P5 levels of the neck. On PASCAL VOC 2007+2012 the authors report that MambaPSA alone reduces parameters by 2.9 %, FLOPs by 12.1 % and raises CPU throughput from 17 to 20 FPS (+17.6 %) with a negligible accuracy change of -0.1 mAP50:95, while the P4 BiViM placement yields the largest accuracy gain of +0.9 mAP50:95. From these numbers they conclude that state-space models offer a favourable efficiency–accuracy trade-off when substituting attention blocks inside NMS-free lightweight detectors.

Significance. If the reported deltas prove robust under fuller evaluation, the work supplies a concrete, drop-in architectural alternative that improves the resource profile of a modern NMS-free YOLO variant without sacrificing accuracy. The dual exploration of a backbone replacement and neck insertions, together with the explicit numerical trade-offs, constitutes a useful engineering contribution for practitioners targeting CPU or edge deployment. Significance is presently constrained by the exclusive reliance on the older PASCAL VOC benchmark and by the absence of statistical or multi-benchmark corroboration.

major comments (3)
  1. [Abstract] The three headline efficiency claims (params -2.9 %, FLOPs -12.1 %, FPS 17 o20) and the accuracy deltas rest on a single dataset (PASCAL VOC 2007+2012) under an unspecified training protocol. Without multi-seed averages, confidence intervals, or results on a contemporary large-scale benchmark such as MS COCO, the general assertion that SSMs constitute a favourable replacement for attention blocks in NMS-free lightweight detectors remains under-supported.
  2. [Abstract] FLOPs and FPS figures are stated without identifying the hardware platform, batch size, numerical precision or profiling tool. These measurement conditions are load-bearing for the claimed 12.1 % FLOP reduction and 17.6 % throughput gain; identical conditions for the baseline and the proposed model must be demonstrated.
  3. [Abstract] The abstract asserts that the P4 BiViM placement yields the best accuracy gain (+0.9 mAP50:95) yet supplies no comparative numbers for the P3 and P5 placements, nor any ablation that isolates the interaction between MambaPSA and BiViM. Without those controls the dual-modification claim cannot be evaluated.
minor comments (3)
  1. [Abstract] The acronym C2PSA is never expanded; readers unfamiliar with the internal blocks of YOLO26 cannot immediately understand what is being replaced.
  2. [Abstract] CPU FPS numbers are given without naming the processor model, software stack or input resolution, reducing reproducibility of the throughput claim.
  3. [Abstract] No reference is made to prior Mamba-based YOLO or detection works; situating the contribution relative to those efforts would clarify novelty.

Circularity Check

0 steps flagged

No circularity: empirical architecture swap with measured efficiency/accuracy deltas; abstract-only review shows no self-definitional or fitted-as-prediction loop.

full rationale

The available material is an abstract-only report of an empirical architecture experiment: replace C2PSA with MambaPSA (and optionally insert BiViM) inside YOLO26, then measure parameters, FLOPs, CPU FPS, and mAP50:95 on PASCAL VOC 2007+2012. There is no derivation chain that folds fitted constants or self-defined quantities back into a claimed first-principles result. The reported deltas (params -2.9%, FLOPs -12.1%, FPS +17.6%, mAP -0.1 or +0.9) are presented as experimental outcomes of the swap, not as quantities forced by construction from a prior fit of the same quantities. No uniqueness theorem, self-citation load-bearing premise, or ansatz smuggled via prior author work appears in the abstract. Self-citation of YOLO26 or Mamba is ordinary background, not circular. Circularity burden is therefore zero; remaining risks (single-benchmark protocol, unspecified training recipe, lack of multi-seed variance) are external-validity / correctness concerns, not circular reasoning. Score 0 with empty steps is the honest finding for this abstract-only empirical paper.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Abstract-only review. No free parameters are fitted in a mathematical sense; the work is an empirical architecture swap. Domain assumptions are the standard ones of the YOLO/VOC detection literature. No new physical or mathematical entities are invented.

axioms (2)
  • domain assumption PASCAL VOC 2007+2012 mAP50:95 under the authors' training protocol is a valid proxy for the efficiency-accuracy trade-off of NMS-free lightweight detectors.
    The abstract reports all numbers exclusively on VOC and draws a general conclusion about SSMs in NMS-free detectors from that single benchmark.
  • domain assumption Mamba / BiViM modules can be inserted into YOLO26's backbone and neck without breaking the NMS-free training or inference pipeline.
    Implicit in the experimental design; the abstract does not discuss any architectural incompatibilities.

pith-pipeline@v1.1.0-grok45 · 6133 in / 2270 out tokens · 18938 ms · 2026-07-15T04:11:14.455189+00:00 · methodology

0 comments
read the original abstract

State space models (SSMs), notably Mamba, have recently emerged as efficient alternatives to self-attention with linear computational complexity. We investigate the integration of Mamba into YOLO26, the latest non-maximum suppression (NMS)-free object detection framework, by proposing MambaPSA, a lightweight Mamba-based replacement for the C2PSA block at the end of the backbone. To complement this study, we additionally insert a bidirectional Vision Mamba (BiViM) module at the P3, P4, and P5 levels of the neck. Experiments on PASCAL VOC 2007+2012 show that MambaPSA reduces parameters by 2.9%, FLOPs by 12.1%, and improves CPU inference throughput by 17.6% (from 17 to 20 FPS) with negligible accuracy change (-0.1 mAP50:95), while the P4 BiViM placement yields the best accuracy gain (+0.9 mAP50:95). These results suggest that SSMs offer a favorable efficiency-accuracy trade-off when replacing attention-based blocks in NMS-free lightweight detectors.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.