A Better Spur Should Start From Each Objective

In the digital marketplace, a product page is often a dense wall of text, packed with technical specifications and marketing jargon that can overwhelm a shopper. To bridge the gap between complex information and human understanding, artificial intelligence is increasingly tasked with a dual job: identifying the most important words in a product description and then writing a short, clear explanation for them. This is not merely a text-editing exercise; it is a balancing act. The system must satisfy strict offline rules, such as ensuring the extracted words are accurate and cover the full range of product features, while simultaneously chasing online goals like maximizing user clicks and minimizing dislikes. These goals often pull in opposite directions. A feature that generates many clicks might be too generic to be accurate, while a highly accurate technical term might be too obscure to interest a casual browser. When computer programs try to optimize for all these conflicting goals at once, they often stumble, oscillating wildly or getting stuck in a loop where improving one metric causes another to collapse.

Researchers at Harbin Institute of Technology, JD.com, and the Chinese Academy of Sciences have developed a new method to solve this specific problem of conflicting goals. They call their approach Multi-Marginal Preference Optimization, a system designed to stop the computer from treating all objectives as a single, messy pile of instructions. Instead of forcing the model to find a single "best" answer that compromises on everything, the new framework treats each goal as a separate lane on a highway. It first cleans up the data it receives, smoothing out the noisy and sparse signals that come from real user behavior, such as clicks and likes, which can be misleading if taken at face value. By adjusting how the computer sees these rewards, the system ensures that rare but valuable information is not ignored simply because it appears less often than common terms.

The core innovation lies in how the system handles the learning process itself. When the computer tries to improve its performance, it generates mathematical updates based on its goals. In older methods, these updates were simply added together, which often caused the instructions for one goal to cancel out or distort the instructions for another. The new method separates these updates before they are applied. It identifies which parts of the learning signal are essential for basic accuracy and which parts are specific to user preferences, then projects the preference signals into a space where they do not interfere with the core rules. This allows the model to learn to be more interesting to users without accidentally forgetting how to be accurate.

Furthermore, the researchers added a safety mechanism to prevent the system from becoming too aggressive late in its training. As models improve, they sometimes fixate on one goal so intensely that they neglect the others, leading to a sudden drop in overall quality. The new system uses the model's own ability to follow instructions to set a boundary. It asks the model to generate examples of perfect behavior under ideal conditions and uses those examples to define a safe zone for future updates. If the model tries to push too hard in one direction, this internal guide gently pulls it back, ensuring that progress remains steady and balanced across all metrics.

The results of this approach were tested on a dataset of real-world e-commerce products, involving 65,232 training products and 8,879 evaluation products, along with user behavior data collected over the past three months. The system was compared against several other advanced methods and found to be significantly more stable and effective. In offline tests, it achieved a completeness score of 94.11 percent and an accuracy of 73.43 percent, outperforming previous methods by wide margins. Crucially, it managed to hit a target coverage rate of 22.82 percent, which is nearly perfect, whereas other methods either fell short or overshot the mark. These improvements were not just theoretical; when deployed in a live online test with real shoppers, the system driven by this new method increased the number of orders placed by 3.14 percent and the total value of those orders by 5.07 percent compared to the previous standard.

The success of this framework extends beyond online shopping. The researchers also applied it to tasks involving tool usage and computer code generation, areas where multiple constraints must be met simultaneously. In these tests, the method consistently produced better results than standard approaches, particularly in avoiding the "mode collapse" where a model finds a single easy solution and repeats it endlessly. Instead, the new system maintained a diverse range of high-quality outputs while adhering to strict rules. By decoupling conflicting goals and providing a structured way to balance them, this work offers a practical solution for making artificial intelligence more reliable in complex, real-world environments where multiple, competing objectives are the norm.