What Makes Good Research?

I am delighted to see more and more people joining the AI research community. For instance, the number of submissions to ICML has surged from 1,767 in 2017 to 9,653 in 2024. However, during recent discussions with peers, I noticed many are confused about a fundamental question: Am I working on good research—or even research at all? I humbly wish to share my thoughts on this matter.

Three Levels of Research I categorize research into three levels. However, let me be frank—many accepted papers fail to reach even the lowest of these levels, as they can scarcely be considered research. Let’s explore these levels from the lowest to the highest:

Level 3: Predictable Outcomes with Engineering Effort

At this level, the results are easy to accept once the proposed method is understood. The novelty often comes from new tasks, experimental settings, or benchmarks. However, the solutions themselves are typically straightforward or based on common sense. Despite their limited fundamental contributions, such works deserve respect for their engineering effort and ability to produce strong results. Practical impact and reproducibility often characterize this level.

Level 2: Unpredictable Outcomes with Innovative Methods

At this level, the results are surprising and counterintuitive—people often find it hard to believe the proposed method could achieve such outcomes. These works uncover insights or phenomena that no one had realized before. They are inspiring, often opening new directions for further exploration. Level 2 research demonstrates creativity, deep understanding, and a willingness to challenge existing paradigms. This is where groundbreaking methods emerge and where researchers most often push the boundaries of knowledge.

Level 1: Field-Defining Breakthroughs

Level 1 research is rare and cannot be directly pursued—it depends on timing, opportunity, and cumulative progress in a field. Such breakthroughs emerge when the accumulation of prior work reaches a tipping point, creating the conditions for a fundamental shift. Achieving this level requires not only deep expertise but also a stroke of luck. These works redefine entire fields, inspire new disciplines, and leave a lasting impact on the scientific community.

This framework is not about labeling research as “good” or “bad” but about understanding the roles different kinds of contributions play in advancing knowledge. All levels have value—engineering efforts drive practical progress, innovative methods inspire new directions, and breakthroughs reshape entire landscapes. I hope this perspective helps researchers, especially newcomers, reflect on their work and find motivation to pursue contributions that matter to them—and to the field.