This fan lamented that it was so difficult to get the ICML paper approved, and he carefully checked the probability that it felt like the imperial examination. After successfully graduating, being competed by major AI companies is really like emerging from a cocoon. The following is the information I checked. ICML is one of the "top two" in the field of machine learning and is also recognized as a hard currency in the industry. • Data for 2025: 12,107 articles were submitted, 3,260 were finally accepted, and the overall acceptance rate was
- 9%. Doesn’t it look low? But here are two cruel realities:
- The invisible elimination of "desk rejection": A large number of papers with incorrect formats and inconsistent directions have been screened out from the submission volume. For submissions that actually enter the review pool, the true acceptance rate will be further discounted.
- The competition for the top spot is fierce: High-quality papers in the Spotlight/oral report category account for only
- 6% of all submissions. Most of the papers that ordinary doctoral students can get are in the Poster category.
- Dimensionality reduction attack by the “King of Papers”: Ph.D. students from Dachang Research Institute (OpenAI, DeepMind, Domestic AI Lab) and top universities in North America account for nearly half of the admissions. These teams have top tutors, free GPU clusters, and mature pipeline scientific research models that ordinary students cannot compare with. :
- First year: finding direction and stepping into pitfalls ◦ Most people spend their first year making mistakes: changing instructors, changing direction, and the experimental results are all overturned. Those who can stably produce preliminary results are considered fast. ◦ It is more difficult for self-funded students: tutors usually give priority to allocating GPU resources and guidance to full scholarship students. You may have to save your own computing power and figure out your own direction, so your progress will naturally be slower.
- The second year: producing results and writing the first article ◦ This is the golden period for producing papers, but many people will get stuck on reproducing experiments and questioning reviewers. It is common for a paper to be rejected two or three times after submission. ◦ ICML is only held once a year. If you fail to apply, you will have to wait for a whole year. Many people submit their second papers and do not get accepted until the third year of their Ph.D.
- Third to fourth years: Double-wounded by "graduation pressure" and "peer progress" ◦ The rhythm of domestic direct doctoral students is: 5-year academic system, 3-4 years to produce results, and 5 years to graduate. ◦ The rhythm of overseas self-financed students is: there is no scholarship to support them, and staying for one more year will cost hundreds of thousands more. Moreover, there is no guarantee of graduation from the tutor, so they must rely on thesis to support graduation and find a job. ◦ When people around you are already preparing to graduate with 1-2 papers in hand, you are still getting rejected repeatedly for your first paper. The anxiety caused by this "poor progress" is engraved in your bones. The cost of studying for a PhD at your own expense is about 300,000 to 500,000 yuan per year. In the field of ML, this is really a matter of "hard work may not be rewarded": • Your direction may be overwhelmed by a large model, and after two years of small improvements, you are suddenly overturned by a SOTA paper from a major manufacturer; • Your reviewer may be a "big expert in the field" and reject it with just "insufficient innovation"; • Even if your tutor does not support submission, experimental data cannot be produced, and GPU cannot run, a paper may be aborted.