Solving Technical Challenges with Large AI Models at JD Retail: Reward Modeling, Query Expansion, and Model Pruning

JD Retail's technology team showcases how young algorithm engineers solve hard technical problems using large‑scale AI models.

Reward modeling for advertising image generation : The team replaced a single large reward model with multiple specialized small reward models that evaluate aspects such as product shape, placement, and color. They built a training‑inference framework where the generator model is fine‑tuned by reinforcement learning using multi‑dimensional quality signals. The resulting system achieves 98% usable image rate and a 30% recall improvement.

Query expansion for e‑commerce search : To bridge the gap between user queries and product descriptions, they designed a large‑language‑model (LLM) based query‑expansion framework. The pipeline consists of (1) pre‑training on consumer behavior and product data, (2) task‑driven fine‑tuning with high‑quality query‑expansion pairs, and (3) reinforcement‑learning alignment with a simulated search engine. Offline experiments and online A/B tests show a noticeable conversion‑rate boost.

Model pruning with FFT and RDP : To reduce the inference cost of text‑to‑image models, they applied Fast Fourier Transform (FFT) for frequency‑domain analysis and the Ramer‑Douglas‑Peucker (RDP) algorithm to locate redundant transformer blocks. The combined method increased training throughput by 40% without degrading performance. The core RDP implementation is shown below:

def rdp(points, epsilon):
    """Ramer‑Douglas‑Peucker algorithm for curve simplification."""
    def perpendicular_distance(pt, line_start, line_end):
        # compute perpendicular distance
        if np.array_equal(line_start, line_end):
            return np.linalg.norm(pt - line_start)
        else:
            return np.abs(np.cross(line_end - line_start, line_start - pt)) / np.linalg.norm(line_end - line_start)
    def rdp_recursion(points, epsilon):
        dmax = 0.0
        index = 0
        end = len(points)
        for i in range(1, end - 1):
            d = perpendicular_distance(points[i], points[0], points[-1])
            if d > dmax:
                index = i
                dmax = d
        if dmax > epsilon:
            results1 = rdp_recursion(points[:index+1], epsilon)
            results2 = rdp_recursion(points[index:], epsilon)
            return results1[:-1] + results2
        else:
            return [points[0], points[-1]]
    return rdp_recursion(points, epsilon)

They also built an Agent full‑link evaluation system that provides both local and end‑to‑end scores for different question types, enabling precise diagnosis of model failures.

Another contribution is a semantic‑driven smart‑assistant that maps user‑specified dimensions to structured product parameters using multi‑turn LLM reasoning and reinforcement learning, leading to higher user engagement and conversion.

Throughout these projects, the engineers emphasize iterative experimentation, cross‑domain knowledge transfer, and continuous learning from papers and open‑source communities.

The article concludes with a recruitment call for JD Retail’s algorithm team, inviting candidates to join a technically deep and commercially impactful environment.