Are modern ML PhDs becoming too incremental, or is this just what research looks like now? [D]
A Reddit discussion questions if most ML PhDs are just polished benchmark papers
A candid Reddit post from an active ML PhD student has sparked widespread discussion across the research community. The author describes a troubling pattern they observe in their own work and across top-tier conferences: most papers follow a predictable formula—take an existing idea, combine it with another, apply it in a slightly different context, tune benchmarks, and claim 'state-of-the-art.' While acknowledging that incremental work has value, the author argues that many contributions are closer to extended master’s theses than true PhD-level science, relying more on compute and refined writing than on fundamental understanding.
The post highlights the tension between field incentives and genuine progress. With conferences rewarding empirical wins and clean narratives, researchers may optimize for publishable deltas rather than reusable methods, failure mode analyses, or deeper mechanistic insights. The author asks whether ML PhDs are declining in quality relative to other disciplines, or if this is normal for a fast-moving, empirical field. The discussion touches on the sustainability of benchmark-driven research and what separates a strong incremental PhD from a mere collection of polished papers.
- Many ML PhDs follow a formula: combine existing ideas, tune hyperparameters, claim SOTA on benchmarks
- Authors worry incentives reward temporary leaderboard wins over reusable methods or mechanistic understanding
- Field debates whether this pattern is normal cumulative research or a decline in scientific depth
Why It Matters
Questions the long-term value of ML research output if most PhDs produce only temporary benchmark improvements.