AdCoEvo: A Co-Evolving Multi-Agent Advertising Delivery System
Authors: Tianzhu Hu, Hang Yu, Zeyuan Zhao, Zeyu Pan, Huan Luo, Chun Gan, Mingzhe Cao, Liubin Wang, Mingming Li, Jiao Dai, Jizhong Han, Songlin Hu
Organizations: Institute of Information Engineering, Chinese Academy of Sciences · School of Cyberspace Security, University of Chinese Academy of Sciences · The University of Sydney
Digital advertising is an important means for platforms to reach potential users and drive business growth. Advertising delivery is the core link between marketing objectives and actual returns. Existing methods have achieved partial automation in individual tasks such as bidding and budget allocation, but they still rely on human operators for overall coordination, making it difficult to support large-scale operations. They also often fail to achieve optimal performance, and individual experience remains difficult to turn into reusable operational capabilities. In addition, as advertising markets change, advertising delivery systems need timely strategy updates to remain effective in dynamic markets. To automate the entire advertising operations process and improve adaptability to market changes, we propose AdCoEvo, a co-evolving multi-agent advertising delivery system. Specifically, the system connects delivery decisions across different timescales and operational granularities through role specialization. Diagnostic results guide account planning and strategy adjustment in the next period. The multi-agent system updates its model parameters based on actual delivery outcomes and continually improves delivery strategies through mutual feedback among planning, adjustment, and diagnosis, thereby achieving co-evolution. Experiments show that, relative to the rule-based operational baseline (Rule), AdCoEvo increases advertiser value by 14.02% while reducing cost per acquisition by 7.07%. Compared with a multi-agent system with fixed parameters, AdCoEvo increases advertiser value by 3.96% and reduces cost per acquisition by 1.24%, indicating that co-evolution can yield greater delivery returns.
Figures & tables
Figure 1: From manual to multi-agent advertising delivery.
Figure 2: Role collaboration and continuous co-evolution in AdCoEvo. The upper panel shows delivery execution, and the lower panel shows role-policy updates.
Method
Sequence 1
Sequence 2
Sequence 3
Sequence 4
Sequence 5
Overall
ADV ↑
CPA ↓
ADV ↑
CPA ↓
ADV ↑
CPA ↓
ADV ↑
CPA ↓
ADV ↑
CPA ↓
ADV ↑
CPA ↓
Score ↑
AuctionNet (Sparse)
No Operation
26,440.6
115.575
26,372.9
115.676
26,175.7
117.539
26,598.1
116.739
26,286.4
115.531
26,374.7
116.207
1.0000
Rule
28,680.3
104.124
28,333.9
103.629
28,376.6
105.553
28,537.4
104.623
28,195.3
103.658
28,424.7
104.315
1.2006
AdPilot *
28,090.3
118.479
27,012.3
118.847
27,103.3
120.397
27,046.0
119.133
26,876.5
118.358
27,225.7
119.040
1.0077
AdCoEvo (Fixed)
31,720.6
98.368
30,709.7
98.583
30,615.4
100.112
31,675.0
99.104
31,154.9
97.600
31,175.1
98.150
1.3995
Table 1: Delivery performance across market sequences. Overall aggregates results from five 50-period sequences.
Business Performance
Estimated API Cost
Latency
Model
ADV ↑
CPA ↓
IR (%) ↓
Score ↑
CNY/Period ↓
CNY/10K ADV ↓
P50(s) ↓
P95(s) ↓
Period(s) ↓
Closed-Source
GPT-5.6-Terra
31,243.6
100.176
0.00
1.3742
5.733
1.8348
3.84
5.27
710.8
Claude-Sonnet-5
31,024.5
100.027
3.22
1.3344
8.152
2.6275
9.18
15.30
1,340.0
Gemini-3.1-Flash-Lite
30,291.8
101.393
8.16
1.2348
0.683
0.2255
1.40
2.23
470.8
Table 2: Delivery performance, API costs, and inference latency of AdCoEvo and foundation models in Sparse. Bold values indicate the best results.
Table 5
O gains
P+O gains
AdCoEvo gains
Seq.
ADV ↑
Score ↑
ADV ↑
Score ↑
ADV ↑
Score ↑
1
0.62
2.25
0.93
2.56
1.28
2.91
2
0.74
3.30
1.03
3.58
5.34
7.93
3
0.64
2.84
1.01
3.11
6.89
9.10
4
0.56
1.67
0.94
1.98
2.10
3.02
5
0.59
2.11
1.33
2.17
4.32
5.66
Table 5: ADV and Score gains (%) over AdCoEvo (Fixed) under different role-update settings in Sparse. All settings retain the complete role composition.
Figure 3: ADV uplift and CPA change relative to AdCoEvo (Fixed) under market switches, averaged across five market sequences using five-period windows.
Figure 4: ADV (solid bars, left axis) and CPA (hatched bars, right axis) under different initial budgets in Sparse.
Parameter
ADV ↑
CPA ↓
Score ↑
Observation length
W=6
29,547.0
97.742
1.3319
W=12 (default)
32,408.7
96.936
1.4731
W=18
30,649.6
107.914
1.2514
Adjustment limit
Table 6: Effects of observation length W and adjustment limit K on delivery performance.
Online advertising studies how digital ads are targeted, ranked, priced, measured, and delivered while balancing budgets, business goals, and user interests.