Learn · Intermediate

Null baselines and multiple comparisons: why an untrained model can look like it learned

A null baseline is the score your measurement produces when nothing actually happened, and in machine learning it is almost never zero. If you run a completely unchanged model through your evaluation pipeline, some problems it solved before will now fail and some it failed will now pass, purely from the machinery around it. Any metric that counts those flips will report a number. The multiple comparisons problem is the companion trap: test enough hypotheses and some will clear your significance threshold by luck alone. Together these two are responsible for a large share of results that fail to replicate.

Start with the medical version, because it is the one everyone already understands. A drug trial does not just give the drug to a hundred people and count who improved. It gives a sugar pill to another hundred and asks whether the drug group did better by more than the gap you would expect from chance. That second group is the null baseline made physical. Without it, "sixty-three patients improved" is not evidence of anything, because you have no idea how many would have improved anyway.

Machine learning evaluation has been slower to adopt the equivalent, partly because the noise sources are less obvious. Where does movement come from if the model did not change? Several places. Generation is sampled, so unless you fix every random seed the same prompt gives different answers. Even at temperature zero, batching changes the order in which floating-point numbers get added, and floating-point addition is not associative, so grouping the same numbers differently produces slightly different sums. Those tiny differences propagate through a long generation and occasionally flip a borderline answer. Serving stacks schedule work nondeterministically, mixture-of-experts models route differently under different batch compositions, and graders that use another model to judge correctness are themselves noisy. None of that is a bug. It is the floor.

The 2026 audit Phantom Gains measured that floor directly. The authors ran a frozen model, one that had learned nothing, through the same self-training evaluation pipeline as the real experiment, and it appeared to both acquire new capabilities and lose old ones. They also showed that the widely used "expansion" statistic, which counts problems newly solved, has a null far from zero, meaning the familiar claim that a model now solves something it never solved before is not by itself evidence of anything. When they replaced it with a per-problem exact test against a pooled baseline under false-discovery-rate control, the apparent gains disappeared on held-out data. This connects directly to how AI gets benchmarked and to ablation studies, which are the component-level version of the same discipline.

The multiple comparisons half is easier to state and just as damaging. A conventional significance threshold accepts a 5 percent chance of calling noise a result. Run one test and that is a reasonable risk. Run a hundred, which is roughly what happens when you sweep learning rates across several model sizes on several benchmarks, and you should expect about five false positives even in a world where nothing you tried works. If you then report only the configurations that looked good, you have published five findings and zero effects. The standard corrections are Bonferroni, which is blunt and divides your threshold by the number of tests, and false discovery rate control via the Benjamini-Hochberg procedure, which is less conservative and asks instead what fraction of your declared discoveries are likely false. FDR control is generally the right tool for machine learning, because you are usually screening many candidates and can tolerate a known share of false leads.

This is not a new complaint. Deep Reinforcement Learning that Matters showed in 2017 that reinforcement learning results swing wildly across random seeds and that many published comparisons were within seed variance. Show Your Work argued that a single reported number hides how much hyperparameter search bought the result. With Little Power Comes Great Responsibility found that many natural-language-processing experiments were statistically underpowered, meaning they could not have reliably detected the effects they claimed to find. The field keeps rediscovering this because the incentives run the other way: a corrected result is smaller and less publishable than an uncorrected one.

What to actually do about it is short. Run your unchanged model through the identical pipeline and report what it scored, including the same number of samples and the same grading path. Report variance across seeds, not just a mean. Say how many configurations you tried before the one you are showing. Use paired tests where the same problems are compared before and after, since that removes problem difficulty as a source of variance. And when screening many candidates, apply FDR control rather than eyeballing a threshold. A result that survives all of that is worth trusting, and one that does not was never there. The same logic underpins calibration, where the question is again whether a number means what it appears to mean, and it is the reason recursive self-improvement claims deserve unusual scrutiny: a loop that measures its own progress with an uncorrected metric will report progress forever.

Ground Truth. (2026, August 22). Null baselines and multiple comparisons: why an untrained model can look like it learned. Ground Truth. https://groundtruth.day/learn/null-baselines-and-multiple-comparisons.html

@misc{groundtruth:null-baselines-and-multiple-comparisons,
  title  = {Null baselines and multiple comparisons: why an untrained model can look like it learned},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {aug},
  url    = {https://groundtruth.day/learn/null-baselines-and-multiple-comparisons.html}
}

Topics: evaluation · statistics · research-methods · benchmarks · reproducibility