AdCoEvo: A Co-Evolving Multi-Agent Advertising Delivery System
Advertising delivery still relies heavily on manual coordination across budgeting, bidding, and diagnosis, limiting scale and performance. AdCoEvo automates this pipeline by assigning Planner, Optimizer, and Inspector roles to an 8B-parameter language model and updating their LoRA adapters in rotation using settled business feedback, so the agents mutually adapt during continuous simulated delivery. The authors report gains of roughly 14% in advertiser value and 7% lower cost per acquisition over rule-based baselines in the AuctionNet simulator.
The paper introduces a sensible operational decomposition and a practical offline-to-online training pipeline, but the empirical evidence is thinner than the framing suggests. The reported co-evolution gains are modest (3.96% ADV, 1.24% CPA over the fixed multi-agent baseline) and rest on just five 50-period market sequences without variance estimates or significance tests. The co-evolution mechanism is effectively alternating fine-tuning of one role at a time; no theoretical or equilibrium analysis is provided to justify why this should converge or remain stable.
The strongest aspect is the task-specific role decomposition. By distinguishing account-level budget planning, ad-unit bid adjustment, and period-end diagnosis, the authors align agent responsibilities with real advertising workflows and construct rewards that reflect each decision’s scope and feedback timing ($r_a^P = Q_a + \omega Q_{all}$, $r_{a,t}^O = Q_{a,t} + \omega Q_{all,t}$, $r_{e-1}^I = \mathbf{w}^\top(\mathbf{y}_e - \mathbf{y}_{e-1})$). The leave-one-out advantage estimation for offline training and the lightweight PrefixValueModel for online baselines are both reasonable design choices.
First, the evaluation lacks statistical rigor. Only five independent market sequences are reported, and no confidence intervals or standard errors accompany the headline metrics. A 3.96% improvement in ADV over the fixed baseline is impossible to interpret without knowing sequence-level variance. Second, the Inspector’s reward relies on a hand-tuned weight vector $\mathbf{w} = [1.0, 0.2, 0.5, 0.1]^\top$ with no sensitivity analysis, yet Inspector updates are critical to the co-evolution claim. Third, the market-switch experiment in Figure 3 is limited to a single excursion and return; the CPA advantage disappears after the first switch and only gradually recovers, suggesting the system may track recent performance rather than robustly adapt.
The evidence supports the claim that task-specific training outperforms generic rules and zero-shot foundation models, but it does not isolate co-evolution as the decisive factor. Table 2 compares an 8B model with SFT, RL, and LoRA tuning against much larger foundation models used without domain-specific adaptation; this mostly demonstrates the value of task-specific training rather than multi-agent co-evolution. The comparison with AdPilot* is more relevant, yet AdPilot* is held fixed during evaluation while AdCoEvo continues updating, confounding architecture differences with online learning.
Partial reproducibility is supported by open-source code and the public AuctionNet simulator, but several gaps remain. The SFT demonstrations are generated by an undisclosed DeepSeek V4 Flash prompting pipeline, and exact prompt templates are not provided. Co-evolution relies on heuristic rollbacks (KL > 0.01 or retained-token fraction < 0.4 triggers parameter restoration; Appendix C.2) whose sensitivity is unexplored. The General environment calibration uses proprietary dataset configurations, so independent researchers may not replicate that setting exactly.
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.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.