Inside Baidu Search Innovation Contest: Winning AI Solutions Across Five Tracks

Contest Overview

The second Baidu Search Innovation Contest, themed “New Search·New Singularity,” ran for four months and drew more than 2,800 participants from 45 provinces and overseas cities. Over 81% of entrants were university students, with the majority being graduate students. The competition received more than 1,600 resumes, primarily targeting machine‑learning, deep‑learning, and AI product innovation roles.

During the event, Baidu organized nearly 20 online/offline training sessions, engaging over 50,000 students directly and reaching a developer community of one million. The contest aimed to open the largest AI application scenario—search—to young talent, fostering innovation and skill development.

Track 1 – Answer Organization

Out of 719 registrations, 220 submissions were received. The champion team from the Institute of Computing Technology, Chinese Academy of Sciences, refined a Lora‑fine‑tuned LLM pipeline, enriched training data with public Q&A, applied model distillation, and introduced NEFTune‑style noise injection to improve robustness. Their solution produced more user‑aligned answers, demonstrating deep technical analysis and solid experimental validation.

Track 2 – Vector‑Intersection Top‑K Search

With 549 participants and 113 submissions, the winning solution was created by a full‑time father and Wuhan University graduate. The team implemented multi‑threaded, multi‑stream parallelism, batch optimization, and a high‑efficiency bitset algorithm for vector intersection. They also devised a threshold‑iterative Top‑K method, achieving a 23× performance boost over the baseline.

Track 3 – AI‑Native Search Application

From 530 entrants and 83 projects, the champion team combined user research with AI capabilities to build an “AI Resume Assistant.” Their product excelled in user‑need discovery, delivering end‑to‑end recruitment workflow support and earning unanimous praise from judges for its innovation and execution.

Track 4 – GPU‑Accelerated Approximate Nearest Neighbor

Among 273 participants and 30 submissions, the winning team from Hangzhou Dianzi University’s Knowledge Graph Lab leveraged pipeline optimizations to surpass the baseline by 1.5× early on, then applied model‑index compression to double performance, ultimately achieving a 3× improvement.

Track 5 – Controllable Image Generation

Out of 390 registrations and 50 projects, the champion team from Beijing Institute of Technology used large‑scale data collection, cleaning, and multi‑Lora model fusion to enhance a diffusion‑based image generation system. Their approach improved the baseline model’s quality by five times, demonstrating the power of data‑driven Lora fine‑tuning.

Conclusion

The contest not only showcased cutting‑edge AI research applied to search but also provided a platform for young engineers to experiment, collaborate, and push the limits of current technology. Baidu emphasizes that such competitions are essential for discovering talent, accelerating innovation, and shaping the future of search.