Neural Architecture Search (NAS) for object detection is severely bottlenecked by high evaluation cost, as fully training each candidate YOLO architecture on COCO demands days of GPU time. Meanwhile, existing NAS benchmarks largely target image classification, leaving the detection community without a comparable benchmark for NAS evaluation. To address this gap, we introduce YOLO-NAS-Bench, the first surrogate benchmark tailored to YOLO-style detectors. YOLO-NAS-Bench defines a search space spanning channel width, block depth, and operator type across both backbone and neck, covering the core modules of YOLOv8 through YOLO12. We sample 1,000 architectures via random, stratified, and Latin Hypercube strategies, train them on COCO-mini, and build a LightGBM surrogate predictor. To sharpen the predictor in the high-performance regime most relevant to NAS, we propose a Self-Evolving Mechanism that progressively aligns the predictor's training distribution with the high-performance frontier, by using the predictor itself to discover and evaluate informative architectures in each iteration. This method grows the pool to 1,500 architectures and raises the ensemble predictor's R2 from 0.770 to 0.815 and Sparse Kendall Tau from 0.694 to 0.752, demonstrating strong predictive accuracy and ranking consistency. Using the final predictor as the fitness function for evolutionary search, we discover architectures that surpass all official YOLOv8-YOLO12 baselines at comparable latency on COCO-mini, confirming the predictor's discriminative power for top-performing detection architectures. The code is available at https://github.com/VDIGPKU/YOLO-NAS-Bench.
Figures & tables
Figure 1: Latency vs. mAP on COCO-mini. Architectures discovered by our predictor-guided EA search consistently Pareto-dominate all official YOLO baselines (v8–v12) across the full latency spectrum, demonstrating the strong discriminative power of the YOLO-NAS-Bench surrogate predictor.
Figure 2: Overview of YOLO-NAS-Bench. (1) A YOLO-style search space spanning channel, depth, and operator dimensions across both backbone and neck is defined. (2) 1,000 architectures are sampled via three complementary strategies and trained on COCO-mini. (3) A LightGBM predictor is trained on the resulting {architecture, mAP} pairs. (4) The Self-Evolving Predictor iteratively expands the pool with high-value architectures discovered by evolutionary search, and retrains the predictor over 10 rounds, yielding an ensemble of 10 LightGBM models over 1,500 architectures.
Dimension
Stage
Candidates
#
Channel
S2
128, 192, 256
3
S3
256, 320, 384, 448, 512
5
S4 1 1 1 S4 channels must be at least S3 to avoid degenerate configurations.
384, 512, 640, 768
4
S5
768, 1024, 1280
3
Depth
S2
1, 2
2
S3
2, 3, 4
3
Table 1: Search space of YOLO-NAS-Bench. Channel, depth, and operator choices for each stage.
Figure 3: Self-Evolving Predictor. Starting from 1,000 architectures, the loop partitions latency into 10 buckets. For each bucket, EA search selects the top 5 architectures using predicted mAP as fitness and real latency as constraint. In each round, these 50 new architectures are trained on COCO-mini, merged into the pool, and the predictor is retrained. After 10 rounds the pool grows to 1,500 architectures, yielding an enhanced benchmark enriched in the high-performance regime most relevant to NAS.
Parameter
Value
Epochs
120
Batch size
128
Image size
640×640
Learning rate ( lr0 )
0.01
LR schedule
Step (decay at epoch 100)
Mosaic
1.0
Table 2: Unified training configuration for all architectures in YOLO-NAS-Bench.
Figure 4: Predicted vs. ground-truth mAP on the full 1,500-architecture pool. Each point is an architecture colored by its sampling source. Points cluster closely around the y=x diagonal, confirming strong agreement between the ensemble predictor and ground-truth performance.
Setting
#Archs
R2
sKT
Before Self-Evolving
1,000
0.770
0.694
After Self-Evolving
1,500
0.815
0.752
Table 3: Predictor quality before and after Self-Evolving. Metrics are reported on the validation split (20%). Ensemble: 10 homogeneous LightGBM models.
Small (S)
Medium (M)
Large (L)
Extra-Large (X)
Model
mAP (%)
Lat. (ms)
mAP (%)
Lat. (ms)
mAP (%)
Lat. (ms)
mAP (%)
Lat. (ms)
Ours
31.9
16.85
32.7
21.00
33.4
27.09
33.6
35.00
YOLO12
27.2
20.40
30.4
22.17
30.8
42.51
32.9
54.13
YOLO11
27.7
15.30
30.8
20.03
32.1
29.84
32.8
37.73
YOLOv10
25.8
15.89
29.7
20.01
30.1
25.83
30.5
36.84
YOLOv9
27.1
19.25
31.6
25.59
31.6
26.85
32.3
53.07
Table 4: Predictor-guided EA search results vs. official YOLO baselines on COCO-mini. All models are trained from scratch under the same protocol. Latency is measured on a single P40 GPU. The top block shows architectures discovered by our predictor-guided EA; subsequent blocks show official baselines.
Predictor
R2
sKT
LightGBM
0.768
0.699
XGBoost
0.758
0.696
NGBoost
0.755
0.704
Random Forest
0.744
0.678
MLP
0.053
0.440
Table 5: Predictor type comparison on initial 1,000 architectures (no ensemble, same data split).
Expansion strategy
#Archs
R2
sKT
Initial (no expansion)
1,000
0.770
0.694
Random +200
1,200
0.776
0.701
Self-Evolving +200
1,200
0.798
0.738
Table 6: Self-Evolving vs. random pool expansion. Same pool size (1,200); metrics on validation split (20%).