AI 时代，PhD 路上的 10 个关键问题｜对谈李磊：港大 PhD、小米 MiMo 研究员

Summary

This 十字路口Crossing interview follows 李磊, a University of Hong Kong NLP doctoral student and Xiaomi large-model post-training researcher, across research method, advisor choice, academic communication, industry transition, and AI-era career anxiety. Its central synthesis is Doctoral Research Operating Method / 博士研究操作法: choose consequential problems, isolate the smallest decisive assumption, test it cheaply, seek feedback early, and only then accelerate execution with AI and parallel work.

The episode also develops Research Impact Beyond Acceptance / 论文录用之外的研究影响力, arguing that publication is a knowledge-sharing mechanism rather than the final reward, and Benchmark Adoption Design / Benchmark 采用设计, which joins a meaningful new capability to realistic tasks, reproducibility, low-friction tooling, and ecosystem compatibility. Its later discussion links Recursive Self-Improvement, Research Taste, large-scale reinforcement learning, and high-quality data: models can increasingly execute defined searches and improvement loops, while problem choice and the definition of valuable progress remain major human bottlenecks.

Key Claims

• Doctoral Research Operating Method / 博士研究操作法 combines strong fundamentals, important problem selection, rapid minimal-assumption tests, early feedback, and accelerated execution after the direction survives validation.
• Difficult and important problems can reward a small but rigorous advance, whereas easy crowded tasks may be copied or surpassed quickly; difficulty alone is not sufficient without consequence and testability.
• Doctoral Training and Career Alignment / 博士培养与职业对齐 makes a PhD most defensible as protected time for deep curiosity and research formation, not as a required ticket to a large-company job.
• Advisor fit matters more than nearby school-ranking differences; applicants should inspect graduate outcomes and wellbeing, speak privately with current and former students, and ideally test the relationship through research-assistant work.
• Research Impact Beyond Acceptance / 论文录用之外的研究影响力 treats code, data, maintenance, real adoption, and community feedback as evidence beyond paper counts, while recognizing that download and attention metrics can also be manipulated.
• Academic social media can shorten feedback cycles and help work cross an overloaded attention system, but it should amplify rather than replace rigorous research.
• Benchmark Adoption Design / Benchmark 采用设计 requires a consequential capability, realistic use cases, reproducibility, usability, and compatibility with common evaluation tooling; Video-MME is presented as a long-video example.
• Model-assisted data generation, parameter search, and algorithm optimization are partial forms of Recursive Self-Improvement, but current loops usually retain human control over objectives and research direction.
• High-quality data is model-relative: examples that are useful for one capability level can be too easy, too hard, or economically unhelpful for another.

Key Quotes

“只要一直 build，就有好事发生” - the episode’s shorthand for creating a persistent public trail of useful work rather than waiting for one outreach message to succeed.

“论文迟早会中，饭要吃、觉要睡” - the closing reminder that publication anxiety does not suspend basic care.

Connections

• 李磊, Xiaomi, and 十字路口Crossing - guest, employer context, and interview venue.
• Doctoral Research Operating Method / 博士研究操作法, Doctoral Training and Career Alignment / 博士培养与职业对齐, and Research Taste - problem choice, validation, formation, and judgment in doctoral research.
• Research Impact Beyond Acceptance / 论文录用之外的研究影响力 and Academic AI Research Role - boundaries between publication, public communication, adoption, and durable research value.
• Benchmark Adoption Design / Benchmark 采用设计, Video-MME, and Agent Evaluation Benchmarks - scientific and operational conditions for useful evaluation.
• Recursive Self-Improvement, Reinforcement Learning AGI Path, and Data-First Post-Training / 数据优先后训 - later technical discussion of feedback, training, and model improvement.

Contradictions

• No settled contradiction found. The episode’s claim that papers are becoming an employment “bonus” rather than a gold standard is a practitioner hiring judgment, not evidence that publications have lost importance across institutions or disciplines.
• The research method is grounded mainly in AI and NLP, where experiments and feedback can be comparatively fast; it may transfer poorly to wet-lab, clinical, field, hardware, or other long-cycle research.
• Comparisons among mainland China, Hong Kong, and the United States are explicitly experience-limited, especially for U.S. visas, internships, and doctoral life.
• Biographical achievements, reviewer workload, model capability, training transparency, data-industry economics, labor-market change, and enrollment trends remain episode-attributed rather than independently audited here.
• Download counts, social reach, paper acceptance, and benchmark leadership are all partial and potentially gameable signals; none independently proves scientific importance or real-world value.