Latentic LabsPublished by Latentic Labs with scime atlasscime
The scime Dispatch

Special edition · one baseline, many numbers · 1,342,992 baseline numbers from 106,622 papers

ICML · NeurIPS · ICLR · CVPR · AAAI · ICCV · ECCV · journals, 2016–2026November 2026Free · atlas.scime.ai/newspaper

← All editions

For 55% of the 600 most-reported baselines, papers' numbers are more than 5% apart

3DGS on Mip-NeRF360 is reported by 56 papers, from 16.06 to 29.04 PSNR. "We beat 3DGS by 2 points" depends on whose number you started from.

Results tables compare a new method with baselines, and the baseline numbers are usually copied from another paper or rerun by the authors. scime atlas read the results tables of 106,622 papers (ICML, NeurIPS, ICLR, CVPR, AAAI, ICCV, ECCV, TPAMI, IJCV, JMLR, 2016–2026) and found 1,342,992 baseline numbers. Grouped by method, benchmark and metric, 18,843 baselines appear in two or more papers; 600 appear in 5 or more.

For 328 of those 600 the highest and lowest reported numbers are more than 5% apart (measured against the median), and for 175 even the middle half of the papers is that far apart. Some of it is real: different training budgets, splits, resolutions or reimplementations. Some is copying from different sources, and a little is misreading. Either way, the number a paper calls "the baseline" is a choice.

3DGS on Mip-NeRF360 (PSNR)

56 papers report 16.06 to 29.04, median 27.46

3DGS on Deep Blending (PSNR)

43 papers report 28.04 to 30.07, median 29.46

GCN on Cora (accuracy)

42 papers report 74.82 to 87.78, median 82.03

GCN on Pubmed (accuracy)

38 papers report 75.61 to 88.9, median 79

3DGS on Tanks&Temples (PSNR)

37 papers report 22.22 to 25.54, median 23.67

GCN on Citeseer (accuracy)

30 papers report 67.42 to 81.39, median 70.3

Each dot is one paper's number. Every line runs from 25% below to 25% above the median (the tick in the middle); numbers further out sit at the ends. The shaded band holds the middle half of the papers.

AI-assisted. The numbers were read by AI from each paper's results tables and not checked by hand; a few are misread, and a benchmark name can hide different splits or settings. Open the papers before you quote a number.

In numbers

Papers with results tables read106,622
Baseline numbers1,342,992
Baselines in 5+ papers600
Of them, more than 5% apart55%
Middle half more than 5% apart29%
Approach pages with a disagreeing baseline405 of 1,260

If you compare with a baseline

  • Say where the number comes from: your own run (with its budget) or a paper (cite it).
  • Where the range is wide, rerun the baseline under your setup, or report the spread.
  • Check the split and the setting, not only the benchmark name.

Per topic

Every approach page on the research map lists the baseline numbers that disagree inside it, with the papers on both ends.

B

The most reported baselines

the 40 reported by the most papers

The baselines that appear in the most results tables, each with every paper's number. Some agree to the decimal (the Swin and DeiT numbers on ImageNet are mostly copied from the original papers); others spread over many points.

Lines run from 25% below to 25% above each median; numbers further out sit at the ends.

Baseline · benchmark · metricPapersEach paper's numberReportedTopic
3DGS
Mip-NeRF360 · PSNR · higher is better
5616.06 – 29.04
median 27.46
Gaussian splatting
3DGS
Deep Blending · PSNR · higher is better
4328.04 – 30.07
median 29.46
Gaussian splatting
GCN
Cora · accuracy · higher is better
4274.82 – 87.78
median 82.03
GNN architectures
GCN
Pubmed · accuracy · higher is better
3875.61 – 88.9
median 79
GNN architectures
3DGS
Tanks&Temples · PSNR · higher is better
3722.22 – 25.54
median 23.67
Gaussian splatting
DeiT-S
ImageNet-1k val · top-1 accuracy (%) · higher is better
3579.8 – 81.2
median 79.8
Attention and transformers
Swin-T
ImageNet-1K val · top-1 accuracy · higher is better
3481.18 – 81.3
median 81.3
Attention and transformers
GCN
Citeseer · accuracy · higher is better
3067.42 – 81.39
median 70.3
GNN architectures
DDPM
CIFAR-10 · FID · lower is better
291.97 – 7.76
median 3.17
Fast diffusion sampling
GAT
Cora · accuracy · higher is better
2876.24 – 88.3
median 83
GNN architectures
DeiT-B
ImageNet-1k val · top-1 accuracy · higher is better
2681.8 – 83.4
median 81.8
Attention and transformers
GAT
Citeseer · accuracy · higher is better
2668.4 – 80.52
median 72.5
GNN architectures
MAML
miniImageNet 5-way 5-shot · accuracy · higher is better
2651.95 – 70.02
median 63.11
Zero-/few-shot and meta-learning
3DGS
Mip-NeRF360 · LPIPS · lower is better
230.182 – 0.389
median 0.216
Gaussian splatting
Qwen2.5-VL-7B
MMStar · accuracy · higher is better
2148.53 – 67.1
median 62.1
Multimodal LLMs
ERM
Waterbirds · worst-group accuracy · higher is better
2038.9 – 85.78
median 69.1
Recognition and long-tail
MAE
ADE20K · mIoU · higher is better
2026.54 – 51.5
median 48.1
Self-supervised and contrastive
Swin-S
ImageNet-1K val · top-1 accuracy · higher is better
2083 – 83.3
median 83
Attention and transformers
MAML
miniImageNet 5-way 1-shot · accuracy · higher is better
2032.93 – 54.73
median 48.7
Zero-/few-shot and meta-learning
GroupDRO
Waterbirds · worst-group accuracy · higher is better
1963.6 – 93.23
median 90.7
Recognition and long-tail
Swin-B
ImageNet-1K val · top-1 accuracy (%) · higher is better
1983.3 – 83.5
median 83.5
Attention and transformers
WL
NCI1 · accuracy · higher is better
1979.98 – 85.8
median 84.46
GNN architectures
LLaVA-1.5
MME · score · higher is better
181434 – 1859
median 1511
Multimodal LLMs
GIN
NCI1 · accuracy · higher is better
1875.04 – 82.75
median 80
GNN architectures
CenterPoint
nuScenes test · mAP · higher is better
1849.54 – 60.3
median 58
3D detection and driving
WGAN-GP
CIFAR-10 · FID · lower is better
1818.21 – 72.23
median 32.75
GANs
OpenVLA-OFT
LIBERO-Long · success rate (%) · higher is better
1890.7 – 95.8
median 94.5
Robot action and navigation
GAT
Pubmed · accuracy · higher is better
1777.2 – 87.39
median 79
GNN architectures
T2M-GPT
HumanML3D test · FID · lower is better
170.116 – 0.141
median 0.116
Video generation
Qwen2.5-VL-7B
MathVista · accuracy · higher is better
1761.4 – 70.2
median 68.2
Multimodal LLMs
KPConv
S3DIS Area 5 · mIoU · higher is better
1764.6 – 67.1
median 67.1
Point cloud classification
ResNet-50
ImageNet · top-1 accuracy · higher is better
1767.3 – 78.6
median 76.13
RNNs and normalization
DiffusionDrive
NAVSIM navtest · PDMS · higher is better
1688.1 – 88.1
median 88.1
Robot action and navigation
Scaffold-GS
Deep Blending · PSNR · higher is better
1629.98 – 30.37
median 30.21
Gaussian splatting
DeiT-B
ImageNet-1K · Top-1 accuracy · higher is better
1681.73 – 83.36
median 81.8
Attention and transformers
DUET
R2R Val Unseen · SR · higher is better
1670.2 – 72
median 72
Robot action and navigation
Qwen2.5-VL-7B
MathVision · accuracy · higher is better
1624 – 26.4
median 25.15
Multimodal LLMs
ResNet-50
ImageNet-1k val · top-1 accuracy · higher is better
1675.3 – 80.83
median 76.2
CNN features
VAE
CelebA · FID · lower is better
1638.76 – 132.8
median 63
Variational and generative sampling
GIN
MUTAG · accuracy · higher is better
1674 – 93.5
median 89.4
GNN architectures

Free for signed-in readers

Read the rest of this edition

Two more pages: the 30 baselines whose numbers are furthest apart, with the papers at both ends, and every research area side by side. Sign in with a free scime account to keep reading.

  • C The widest ranges
  • D Area by area

Sign in free

Already signed in? Reload this page. Accounts are free; we only use your email to sign you in and, if you agree, to tell you about new editions.

E

About the data

how this edition was made

Numbers. For 106,622 papers an AI model read the full text and recorded each row of the results tables: setting (benchmark and split), metric, method, whether it is the paper's own method or a baseline, and the number. These were not checked by hand.

Grouping. Baseline rows are grouped by method, setting and metric with letters, digits and case folded (as on the topic pages), so "Mip-NeRF 360" and "MipNeRF360" are one benchmark but different splits named the same way are not told apart. A paper that reports the same baseline twice with different numbers is left out of that group; numbers on another scale (a fraction among percentages) and far outliers are dropped, as on the topic pages. "More than 5% apart": highest minus lowest, divided by the median.

Corrections: contact@latenticlabs.com. Not affiliated with or endorsed by any of the conferences or journals.

Get the next scime Dispatch and a monthly field update

One email when a new edition is out (conference editions and special editions like this one), plus a short monthly note on what moved in machine learning research.

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

← All editions · © 2026 Latentic Labs