acceptodds
Under review as a conference paper at ICLR 2027

XiYOLO: Energy-Aware Architecture Search for Object Detection from Sparse Hardware Measurements

Abstract

Object detection on heterogeneous edge devices must satisfy strict energy, latency, and memory constraints while still providing reliable perception for downstream autonomy. Existing energy-aware NAS methods often target limited deployment settings, while real energy remains difficult to optimize because it is highly device-dependent and costly to measure. We address these challenges with an energy-adaptive framework that combines an energy-aware XiResOFA search space, a two-stage energy estimator, and iterative multi-objective architecture search. The search yields an accuracy–energy Pareto frontier from which an operating point can be selected according to mission-specific energy requirements. We select a balanced operating point from this frontier as the reference architecture and apply post-search scaling to construct the XiYOLO family. Experiments on PascalVOC, COCO, and real-device deployment show that XiYOLO achieves a stronger energy–accuracy tradeoff than YOLO baselines. On PascalVOC, the medium XiYOLO model reaches 86.15 mAP50 while reducing energy relative to YOLOv12m by 20.6% on GPU and 35.9% on NPU. On COCO, XiYOLO reduces energy relative to YOLOv12 by up to 53.7% on GPU and 51.6% on NPU at the small scale. With only 20 target-device measurements, the proposed two-stage estimator recovers 88.6% and 93.0% of dense-reference search hypervolume on COCO and PascalVOC, respectively, corresponding to 38.1% and 27.6% improvements over joint adaptation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Energy-Aware Neural Architecture Search, Object Detection, Edge AI, Energy-Efficient Inference

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_5924OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposed object detection on edge devices, and requires satisfying simultaneous constraints on energy, latency, and memory. This paper introduces a framework that combines the "XiResOFA" search space (featuring an elastic structure), a two-stage energy estimator, and a multi-objective iterative search process to derive an accuracy-energy Pareto frontier. The resulting achieves accuracy comparable to YOLOv12 on the PascalVOC and COCO datasets, while reducing GPU/NPU energy consumption by up to 53.7%/51.6%.

Critical strengths

A strength of this paper is its reliance on direct energy measurements taken from actual hardware. They measures real-world power consumption on an Adreno 650 GPU, a 15 TOPS NPU, and a Jetson Orin Nano (over 1,000 times trial).

This enhances the credibility of the findings. the XiResOFA search space is well-designed: it consistently applies the same family of blocks across the backbone, FPN, and PAN components. By avoiding disparate search rules for different modules, the framework facilitates natural weight sharing among subnets; the entire search space is governed by a unified design philosophy.

Critical weaknesses

It has a limitation that the search space is restricted to architectures derived from YOLOv12 (XiResOFA uses YOLOv12n as its base), the search results inherit the inductive biases of this specific family.

The authors report that the search space itself is concentrated in a region that, on average, consumes 21.1% less energy than the original YOLOv12; this raises a question that requires further analysis: is the efficiency discovered through the search truly a result of the search process itself, or is it merely a consequence of the biases embedded in the design of the search space?

Questions for authors
  1. Does the post-search scaling process affect other metrics, such as energy consumption or accuracy?
  2. It is difficult to verify whether accuracy is being effectively captured by the proxy; is there any additional supporting evidence for this?
  3. The paper appears to focus on multi-device applicability—is there any evidence to support this claim?
1 comment
AuthorsOpenReview ↗

Response to Reviewer 5924

We thank Reviewer 5924 for the helpful feedback. We first address the questions on post-search scaling, accuracy-proxy reliability, and multi-device applicability, then isolate the contribution of the search process from the XiResOFA search space.

1. Does Post-Search Scaling Affect Accuracy and Energy? (Q1)

Yes. Scaling improves trained accuracy while increasing physical inference energy. We trained five representative architectures across nano, small, and medium scales (15 configurations) on PascalVOC under a matched 300-epoch protocol with early stopping:

  • Trained Accuracy (mAP50): Increases from 78.9–80.3% (nano) to 83.1–83.7% (small) and 84.9–85.6% (medium).
  • Physical Energy (45 hardware measurements): Across Adreno GPU, Adreno NPU, and Jetson Orin Nano, average gross energy rises relative to nano by 1.99$\times$ / 3.91$\times$ (GPU), 1.99$\times$ / 3.89$\times$ (NPU), and 1.80$\times$ / 3.51$\times$ (Orin Nano) for small / medium scales, respectively. These measurements quantify the accuracy–energy scaling behavior across five architectures and three hardware backends.

2. Does the Accuracy Proxy Preserve Useful Architecture Rankings? (Q2)

Yes, for ranking against inherited-supernet accuracy. NAS primarily requires reliable relative ordering to filter promising candidates rather than exact prediction of final standalone mAP. Below we reproduce the held-out validation from Section 4.4, Table 1 (200 held-out architectures per dataset):

ProxyDataset / DeviceRMSE ↓\downarrowSpearman ρ↑\rho \uparrowKendall τ↑\tau \uparrowRecall@10% ↑\uparrow
AccuracyCOCO0.01970.75030.56070.400
AccuracyPascalVOC0.00990.88530.71270.600
EnergyAdreno 65025.25 mJ0.90200.73300.500

The rank correlations (ρ=0.75–0.89\rho = 0.75\text{--}0.89) measure ordering consistency, while Recall@10% (0.40–0.600.40\text{--}0.60) measures recovery of top-performing candidates. These results support proxy-guided candidate selection rather than exact final-accuracy prediction. The 15 independently trained scaled models in Q1 provide complementary downstream validation confirming strong trained accuracy across scales, though they are not a separate 200-architecture fully trained rank test.

3. What Evidence Supports Multi-Device Applicability? (Q3)

The generic energy prior generalizes zero-shot to unseen hardware, while optimal architectures remain device-dependent. We measured 214 matched architectures across Adreno 650 GPU, Adreno NPU, Jetson AGX Orin, and Jetson Orin Nano. To test prediction transfer explicitly, we performed leave-one-device-out (LODO) evaluation on 98 common architectures, excluding all target-device data during generic-prior training. On 78 held-out architectures per target, zero-shot results are:

Unseen Target DeviceSpearman ρ↑\rho \uparrowGross-Energy Selection Regret ↓\downarrow
Adreno GPU0.8340.74%
Adreno NPU0.8960.45%
AGX Orin0.8111.37%
Orin Nano0.8150.23%
  • Gross-energy selection regret is averaged across nine matched accuracy thresholds and five evaluation draws.

The 0.811–0.896 rank correlations and 0.23–1.37% selection regret show that our energy predictor transfers well to unseen hardware without target-device calibration. However, directly transferring architectures selected on Adreno GPU to other devices incurs 4.81–7.32% net-energy regret on COCO under matched accuracy constraints (and 8.1–23.0% top-10% gross-energy regret across all 214 architectures; see Response KB9F). This motivates our two-stage design combining a transferable prior with sparse target calibration (note: full sparse-calibrated NAS in the paper was conducted on Adreno GPU, while the other devices evaluate cross-device transfer).

4. Is the Gain Due to Search or the XiResOFA Search Space? (W1)

Both contribute, and the search gain remains large after controlling for the search space. While Figure 8(a) shows that XiResOFA itself favors low-energy architectures, we compared our iterative NAS with random search inside that identical XiResOFA space across 15 matched seeds and 20,000 evaluations:

Search Method (in identical XiResOFA space)COCO HV ↑\uparrowPascalVOC HV ↑\uparrow
Random search0.065110.04704
Iterative NAS (ours)0.078870.05987
Improvement+21.1%+27.3%

Because both searches share the exact same YOLOv12-derived architectural biases, the +21.1% / +27.3% Pareto HV improvement isolates the additional benefit of guided NAS beyond the efficiency of XiResOFA. We will incorporate all new analyses into the revised manuscript.

Reviewer_BccsOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents XiYOLO, an energy-aware architecture search framework for object detection on heterogeneous edge devices. It combines a weight-sharing XiResOFA supernet across the backbone, FPN, and PAN with a two-stage energy predictor and stage-wise multi-objective search. From the resulting Pareto frontier, the framework selects a balanced reference architecture and scales it into nano, small, and medium detectors. Experiments benchmark XiYOLO against YOLOv8 and YOLOv12 on PascalVOC and COCO across an Adreno 650 GPU, a 15 TOPS NPU, and a Jetson Orin Nano. On PascalVOC, the paper reports 86.15 mAP50 for XiYOLOm with 20.6% lower GPU energy than YOLOv12m, while on COCO XiYOLOs reduces GPU energy by 53.7%.

Critical strengths
  • The paper focuses on the edge detector design around a mission speicific energy budget. It directly addresses the misalignment between analytical FLOPs and real hardware energy. Rather than relying on synthetic cost tables, the evaluation measures physical board power on embedded robotics hardware.
  • Energy is measured on an actual drone platform with a shunt resistor oscilloscope, with idle power subtracted and results averaged over 1000 inferences. This is more credible than most energy-aware NAS papers, which rely on FLOPs or simulators.
  • Separating a transferable architecture-level energy ranking prior from a lightweight target-device residual is a well motivated approach to reducing hardware profiling overhead.
Critical weaknesses
  • The framework combines pre-existing components by adding a residual connection to XiNet blocks inside an Once-for-All supernet.
  • The main gains may come from the block, not the search. In Table 7, swapping XiNet convolution alone cuts energy from 8.24 J to 4.15 J.
  • The baselines are only YOLOv8 and YOLOv12. There are no NAS-based detectors, e.g., ELASTIC, MobileDets, or other efficient detectors.
  • There is no latency-aware or FLOPs-aware search baseline. Without one, it’s unclear that optimizing energy finds different architectures.
Questions for authors
  • Is the “XiRes” small model in Table 7 the same network as XiYOLOs in Figure 5?
  • How does an FLOPs- or MAC-based proxy compare when calibrated with the same 20 Adreno samples?
1 comment
AuthorsOpenReview ↗

Response to Reviewer Bccs

We thank Reviewer Bccs for the constructive feedback. Below we clarify Table 7 vs. Figure 5, isolate the search contribution from block design, evaluate MACs/latency proxies and search objectives, and compare against EfficientDet.

1. Is XiRes in Table 7 the Same Network as XiYOLOs in Figure 5? (Q1, W1, W2)

Yes. The "XiRes" small model in Table 7 is the exact same XiYOLOs network evaluated in Figure 5. Table 7 isolates its block-level components: XiNet convolution cuts energy from 8.24 to 4.15 J per 30 inferences, while residual connections restore accuracy to 84.38 mAP50 at 4.36 J.

NAS further improves efficiency beyond block design. To isolate the search contribution from the XiRes block, we compared our iterative NAS against random search within the identical XiResOFA space (where all candidates use XiRes blocks) across 15 matched seeds and 20,000 evaluations:

Search Method (in identical XiResOFA space)COCO HV ↑\uparrowPascalVOC HV ↑\uparrow
Random search0.065110.04704
Iterative NAS (ours)0.078870.05987
Improvement+21.1%+27.3%

Pareto hypervolume (HV) measures the accuracy–energy region dominated by the Pareto frontier under a common reference point. Our iterative NAS improves HV by +21.1% / +27.3% on COCO/PascalVOC under identical search spaces and budgets, demonstrating the independent contribution of guided search beyond the XiRes block.

2. MACs/Latency Proxies vs. Direct Energy Optimization (Q2, W4)

MACs-based proxies can provide competitive energy rankings under sparse calibration, but optimizing MACs does not necessarily minimize hardware energy. While calibrating MACs on 20 samples captures coarse scale trends, MACs assigns identical cost to architectures with equal compute but different operator mixes, kernel sizes, and memory behavior. By combining an architectural prior with a 20-shot residual correction, our two-stage estimator retains 88.6% / 93.0% of dense-reference search HV on COCO/PascalVOC (vs. 64.1% / 72.9% for single-stage joint adaptation).

To directly test whether optimizing energy finds better and different architectures than optimizing MACs or latency (W4), we compared Energy-, Latency-, and MACs-aware NAS using identical search spaces, accuracy predictors, evolutionary settings, and 5 matched seeds (using dense cost predictors to isolate the optimization objective from calibration variance):

Search ObjectiveCOCO Energy HV (%) ↑\uparrowPascalVOC Energy HV (%) ↑\uparrow
Energy (ours)100.0100.0
MACs94.196.6
Latency90.191.8
  • HV is evaluated using a common dense-energy reference and normalized to energy-aware search.
  • Frontier & Architectural Differences: Energy-aware search achieves 5.9 / 3.4 percentage points higher normalized HV (+6.3% / +3.5% relative) than MACs-aware search and 9.9 / 8.2 points higher than latency-aware search on COCO/PascalVOC. Across 9 matched accuracy thresholds, selected architectures differ in 84% / 48% of block configurations from MACs-aware search and 85% / 58% from latency-aware search (achieving 0.84%/0.48% and 0.86%/0.84% lower average predicted energy across matched thresholds while expanding the overall Pareto frontier).
  • Physical Hardware Evidence (Appendix D, Tables 5–6): Pruning YOLOv12n reduces latency from 7.97 to 7.42 ms (and lowers MACs) but increases measured inference energy from 81.0 to 86.0 mJ, confirming that latency/MACs proxies can diverge from physical energy.

3. Comparison with Efficient & NAS-Based Detectors (W3)

We additionally evaluated EfficientDet-D0–D3 (which pairs a NAS-searched backbone with BiFPN scaling) on COCO val2017:

Detector ComparisonBaseline APXiYOLO APBaseline / XiYOLO GMACs
EfficientDet-D0 / XiYOLO-N@51234.337.32.55 / 2.55
EfficientDet-D1 / XiYOLO-S@41640.140.86.13 / 5.41
EfficientDet-D2 / XiYOLO-S@51243.444.011.03 / 8.33
EfficientDet-D3 / XiYOLO-M@51247.048.124.94 / 24.66

XiYOLO achieves 0.6–3.0 higher COCO AP than EfficientDet at equal or up to 24.5% lower GMACs. We will incorporate all new comparisons and ablations into the revised manuscript.

Reviewer_KB9FOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes XiYOLO, an energy-aware neural architecture search (NAS) framework for object detection on heterogeneous edge devices. The framework consists of three main components: (1) XiResOFA, a searchable detector architecture with flexible channel widths, convolution kernel sizes, and attention configurations; (2) a two-stage energy estimator that combines a transferable architecture-level energy predictor with a hardware-specific residual model; and (3) an iterative multi-objective search procedure that identifies a Pareto frontier balancing detection accuracy and energy consumption. The paper addresses an important practical challenge: optimizing neural networks for actual energy consumption on target hardware rather than relying solely on computational proxies such as FLOPs or parameter count.

Critical strengths

The proposed method maintains an accuracy-energy Pareto frontier over optimizing a single scalar objective when deployment requirements vary.

The two-stage estimator is a useful component of the work. Separating a transferable architecture-level prior from hardware-specific residual adaptation is intuitively motivated and offers a practical way to reduce the number of expensive target-device measurements.

Critical weaknesses

The accelerations and advancements are only limited to bottleneck block applications like YOLO. It is difficult to conclude how the block can transfer to attention based settings like transformers.

The proposed framework combines several established ideas: once-for-all weight sharing, elastic architectural choices, evolutionary multi-objective optimization, iterative stage-wise search, and hardware-aware energy prediction. The paper states that it extends ELASTIC with Pareto-based accuracy–energy optimization and introduces XiResOFA and a two-stage energy estimator. However, the distinct methodological contribution of each component is not yet sufficiently clear.

The evidence for generalization across heterogeneous hardware is limited. The sparse-calibration comparison focuses on one target device, the Adreno 650, while the deployment evaluation covers additional GPU and NPU platforms. It is unclear whether the same estimator and sparse-calibration procedure were used to guide search on these other devices, or whether their measurements were used only for evaluating already-selected architectures. The choice of 20 measurements is unclear.

The paper also does not sufficiently investigate the relative importance of different architectural blocks in the search space. Identifying which blocks are essential for maintaining detection accuracy and which can be simplified or removed with minimal performance degradation could provide valuable insights into the architecture. Such an analysis would help reveal the dynamic contribution of individual blocks and enable more adaptive architectural configurations tailored to different accuracy and energy constraints.

Questions for authors

The paper is an interesting read. Thank you! Please identify how the proposed method can be scaled across other devices - is it necessary to calibrate for every device? Or is the rough order of performance roughly similar across all architectures?

1 comment
AuthorsOpenReview ↗

Response to Reviewer KB9F

We thank Reviewer KB9F for the constructive feedback. We address cross-device transfer and calibration first, followed by calibration budgets, methodological novelty, and block-level sensitivity.

1. Cross-Device Transfer: Is Calibration Necessary for Every Device? (Q1, W3)

Calibration is not required for initial zero-shot ranking, but is recommended for target-specific selection. In leave-one-device-out (LODO) tests across four platforms (Adreno 650 GPU, Adreno NPU, Jetson AGX Orin, Orin Nano), our generic energy prior excludes all target-device data and achieves Spearman ρ\rho of 0.811–0.896 with 0.23–1.37% gross-energy selection regret on 78 held-out architectures per device. This supports initial screening on unseen hardware without target-device profiling.

However, architecture preferences vary across hardware. We measured 214 identical architectures across all four devices. Below is the top-10% net-energy selection regret when models selected on one device are evaluated on another:

Selection ↓\downarrow \ Target →\rightarrowAdreno GPUAdreno NPUAGX OrinOrin Nano
Adreno GPU0.0%23.0%16.4%20.2%
Adreno NPU17.0%0.0%13.5%15.2%
AGX Orin18.8%17.3%0.0%11.4%
Orin Nano20.3%14.6%8.1%0.0%

Direct cross-device selection incurs 8.1–23.0% top-10% net-energy regret across the 214 architectures (which differs from the 4.81–7.32% COCO net-energy regret under matched accuracy thresholds in Response 5924). Thus, an unseen-device prior provides strong initial rankings, while device-specific measurements improve final selection. Our two-stage estimator combines this transferable prior with a sparse hardware-specific residual correction. Clarification on W3: The submitted sparse-calibrated NAS targets Adreno GPU; the other devices support transfer and deployment evaluations rather than separate sparse-calibrated NAS runs.

2. Why Use 20 Calibration Measurements? (W3)

We evaluated calibration budgets of 5–50 target-device measurements on the dense Adreno reference benchmark:

  • 10 →\rightarrow 20 measurements: Held-out Spearman ρ\rho improves from 0.519 to 0.724 and energy MAE drops from 38.7 to 31.9 mJ ( reaching 0.785 at 50 measurements, showing diminishing gains).
  • Search impact: With only 20 measurements ($\sim$2.2% of dense profiling), our two-stage calibrated search retains 88.6% / 93.0% of dense-reference Pareto hypervolume (HV) on COCO/PascalVOC, outperforming single-stage joint adaptation (64.1% / 72.9%).

3. Methodological Contributions and Scope (W1, W2)

While ELASTIC uses cyclic stage-wise search for single-objective loss minimization under size constraints, XiYOLO introduces a maintained accuracy–energy Pareto archive across an elastic detector supernet with two-stage energy calibration. Controlled ablations isolate each component:

  1. XiRes Block (Table 7): Swapping XiNet conv cuts energy from 8.24 to 4.15 J; adding residual connections restores mAP50 to 84.38 at 4.36 J.
  2. Guided Pareto Search vs. Space Bias: In the identical XiResOFA space (20,000 evaluations, 15 matched seeds), iterative NAS improves HV by +21.1% (COCO) and +27.3% (PascalVOC) over random search.
  3. Direct Energy Objective: Under matched dense-cost search settings, energy-aware optimization improves gross-energy HV by +6.3% / +3.5% over MACs-aware search and +11.0% / +8.9% over latency-aware search.
  4. Two-Stage Calibration: Retains 88.6% / 93.0% of reference HV at 20 samples vs. 64.1% / 72.9% for joint adaptation. Scope on Transformers (W1): XiResOFA is detector-specific (searching widths, kernels, and attention configurations across backbone/FPN/PAN). Extending our Pareto search and two-stage estimator to standalone Vision Transformers requires ViT-specific architectural descriptors and profiling data, and remains unvalidated on standalone transformers.

4. Architectural Block Sensitivity (W4)

Freezing individual stages during search reduces COCO HV by 1.8% (backbone), 3.2% (FPN), and 5.1% (PAN). Single-block analysis identifies backbone block 1 as most accuracy-sensitive (predicted mAP50 drops of 3.52 / 5.17 points on COCO/PascalVOC), whereas several FPN blocks tolerate lower-energy settings and PAN blocks 6–7 offer substantial energy savings with minimal accuracy loss. We will incorporate all new analyses into the revised manuscript.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Follow-Up on Author Response and Reviewer Discussion

Dear Reviewers and Area Chairs,

Thank you again for your careful reviews and constructive feedback. We have posted detailed responses addressing the reviewers’ questions and concerns, including the requested clarifications and additional experimental results.

As the author–reviewer discussion period is approaching its conclusion, we would be grateful if the reviewers could let us know whether our responses have adequately addressed their principal concerns or whether any further clarification would be helpful. We remain available to respond promptly during the remainder of the discussion period.

Thank you for your time and consideration.

AuthorsOpenReview ↗

Follow-Up on Author Response and Reviewer

Dear Reviewers and Area Chairs,

As today marks the final day of the author–reviewer discussion period, we would like to kindly follow up on our previous responses and additional experimental results.

We would appreciate any final feedback on whether our responses have adequately addressed your concerns. If any questions or clarifications remain, we would be happy to address them before the discussion period closes.

Thank you again for your time and thoughtful consideration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.