AI Paper Atlas

Does it actually help? 37 common changes, by what papers report

For each change, the share of papers at ICML, NeurIPS, ICLR, CVPR, AAAI, ICCV, ECCV, TPAMI, IJCV and JMLR (2016–2026) that report a gain or a loss after making it. From 1,254,369 results read from 124,803 papers.

Adding a component: 82% of papers report a gain. Turning a knob: 65%.

Each paper counts once per change, with the outcome it reports most often.

Knobs: a setting turned up or down

ChangePapersReport a gainReport a loss
More training data1,14778%6%
More layers / deeper model4,15369%17%
Larger beam / more samples (best-of-N)31468%11%
More in-context examples28768%9%
Larger model or backbone51067%13%
Longer context window43262%25%
Larger embedding / hidden size1,18561%18%
More diffusion sampling steps44061%17%
Larger batch size77460%20%
Higher weight decay8859%26%
More attention heads10658%18%
More retrieved passages (RAG)5157%25%
Train longer36955%21%
Larger LoRA rank23346%20%
Larger learning rate44744%38%
Classifier-free guidance scale up15035%36%
Higher sampling temperature12234%42%

Components: a part added to a method

ChangePapersReport a gainReport a loss
Add a contrastive loss1,26595%2%
Add attention module46790%5%
Knowledge distillation2,02689%7%
Add a reconstruction / auxiliary loss95089%5%
Test-time augmentation8089%8%
Pretrained initialization8686%5%
Add learning-rate warmup34584%9%
Ensembling1,04783%9%
Add residual connections53382%9%
Add data augmentation66981%9%
Add positional encoding62981%8%
Add EMA of weights36980%9%
Self-consistency / majority vote56178%10%
Add mixup or CutMix53076%17%
Cosine learning-rate schedule6673%14%
Add batch / layer normalization42772%15%
Chain-of-thought prompting1,10169%18%
Add dropout78369%19%
Add label smoothing18867%20%
Add gradient clipping8864%12%

Check a change for your own setup

The results were read by AI from each paper's full text and matched by wording. These are what papers chose to report, and published results lean toward gains.

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy