E-means: outperforming k-means (in one-tenth the epochs)
Comments on “E-means: outperforming k-means (in one-tenth the epochs)”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Researchers at Meta's FAIR and École des Ponts present e-means, a mini-batch k-means using exponential-moving-average centroid updates, a cosine-decayed step and dead-cluster revival. Batch size and decay follow scaling laws fitted on one dataset (DEEP), so no sweep is needed. The authors report beating Lloyd's 25-epoch loss in 1.4 to 2.6 epochs, 10-45x GPU speedups, and 10 billion DINO vectors clustered into 16 million centroids in 2.5 epochs. These are the authors' own results, and the laws are less suited to low-dimensional or tiny datasets.
Why it ranked
No. 10 of 106 with a score of 0.68. The panel rated it highest on new information (0.94) and technical depth (0.94), lowest on audience breadth (0.51). Placed No. 10, No. 16 by score alone: the placement terms moved other stories past it.
- AI judgeJudge 2claude-haiku-5-5
- Original research with strong evidence, deep technical detail and scaling laws for k-means; clearly useful to data/ML builders, though niche in audience and of moderate broader significance.
- AI judgeJudge 3gpt-5.6-luna
- A detailed primary research paper introduces an apparently reproducible mini-batch k-means method, scaling laws, algorithms, ablations, and large-scale benchmarks. It is highly actionable for embedding, retrieval, and data practitioners, though relevance is concentrated in those niches.
Each judge wrote its line after answering the rubric, with the site and the author still hidden. Every answer behind this is in the table under Score.
Score
0.68total under rubric v2.3
- Significance0.67
- Reader interest0.73
- Audience breadth0.51
- Practical relevance0.74
- Substantiated0.91
- Technical depth0.94
- New information0.94
- Counts against: Promotional0.06
- Counts against: Overclaim0.31
Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).
| Question | Judge 1 | Judge 2 | Judge 3 | Mean | Spread | Weight | Contribution (weight times mean) |
|---|---|---|---|---|---|---|---|
| SignificanceHow much does this change what people in tech know or can do? | 0.69 | 0.55 | 0.78 | 0.67 | 0.23 | +0.24 | +0.162 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.60 | 0.70 | 0.90 | 0.73 | 0.30 | +0.26 | +0.191 |
| Audience breadthHow many people who build technology does it concern? | 0.54 | 0.45 | 0.55 | 0.51 | 0.10 | +0.13 | +0.067 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.70 | 0.70 | 0.82 | 0.74 | 0.12 | +0.22 | +0.163 |
| SubstantiatedHow well is it backed? Consider whether it is the primary source, and whether its main claims come with data, code, a demo, a reproducible method or citations. | 0.99 | 0.85 | 0.90 | 0.91 | 0.14 | +0.05 | +0.046 |
| Technical depthHow technically substantive is it? | 0.98 | 0.85 | 0.97 | 0.94 | 0.14 | +0.05 | +0.047 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.93 | 0.90 | 0.98 | 0.94 | 0.08 | +0.05 | +0.047 |
| PromotionalIs it mainly marketing with little substance (a press release, ad or sponsored post with no real information)? A launch post that carries real news is not promotional. | 0.08 | 0.05 | 0.04 | 0.06 | 0.04 | −0.19 | −0.011 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.46 | 0.25 | 0.22 | 0.31 | 0.24 | −0.11 | −0.034 |
| Sum of contributions | +0.676 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.676 | ||||||
Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.
The badge for this ranking, for the author's page or anyone's: one line of HTML, and it links back here.
<a href="https://rubric.hrih.si/editions/2026-11-12/10/"><img src="https://rubric.hrih.si/badge/2026-11-12/10.svg" alt="Ranked #10 on Rubric, 12 Nov 2026" width="245" height="28"></a>Claim check
Claim check: 8 no source given. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.
No source givenBenchmark result
On a wide array of embedding datasets, e-means beats the loss of Lloyd k-means at 25 epochs in 1.4 to 2.6 epochs, which translates to 10-45x GPU wall-clock speedups over state-of-the-art k-means implementations.
The article gives no link or citation for it.
The article says “it beats the loss of Lloyd k-means at 25 epochs in 1.4 to 2.6 epochs, which translates to 10-45 $\times$ GPU wall-clock speedups compared to state-of-the-art k-means implementations”
The source says No source to compare with.
No source givenNumber
e-means partitions DINO10B, a dataset of 10 billion DINO embeddings, into 16 million clusters in only 2.5 epochs, reaching a loss below Lloyd's at 25 epochs.
The article gives no link or citation for it.
The article says “e-means partitions DINO10B, a dataset of 10 billion DINO embeddings, into 16 million clusters in only 2.5 epochs, reaching a loss below Lloyd’s at 25 epochs”
The source says No source to compare with.
No source givenBenchmark result
Against Flash-KMeans, e-means needs 1.4 to 2.6 epochs to reach Flash-KMeans' MSE at 25 epochs, giving 9.5x to 45.5x wall-clock speedups.
The article gives no link or citation for it.
The article says “This number goes from 1.4 to 2.6 epochs, translating to 9.5x to 45.5x wall-clock speedups against Flash-KMeans, on top of reaching better loss at convergence”
The source says No source to compare with.
No source givenNumber
With IVF assignment, e-means is 4.3x cheaper at 1B and 34x cheaper at 10B than exact assignment, training 16.7M centroids in under one GPU-day at 10B.
The article gives no link or citation for it.
The article says “it is $4.3\times$ cheaper at 1B and $34\times$ cheaper at 10B, where it trains $16.7$ M centroids in less than one GPU-day”
The source says No source to compare with.
No source givenNumber
The batch-size scaling law incurs a median 0.19% excess MSE on held-out datasets.
The article gives no link or citation for it.
The article says “The obtained law incurs a median $0.19\%$ excess MSE when run on held-out datasets, close to the targeted $0.2\%$”
The source says No source to compare with.
No source givenBenchmark result
When Flash-KMeans is initialized with k-means++, e-means still has lower mean error after 5 epochs on all 8 configurations, with margins of 0.06% to 0.33%.
The article gives no link or citation for it.
The article says “e-means still has the lower mean error after 5 epochs on all 8 configurations, though the margin narrows to between $0.06\%$ and $0.33\%$”
The source says No source to compare with.
No source givenGeneralization
e-means reaches lower final loss than k-means++-initialized Lloyd, and reaches converged Lloyd loss about an order of magnitude faster on average.
The article gives no link or citation for it.
The article says “e-means reaches lower final convergence than a k-means++-initialized Lloyd algorithm, and reach the converged loss of state-of-the-art Lloyd implementations about an order of magnitude faster on average”
The source says No source to compare with.
No source givenGeneralization
The scaling laws fitted on DEEP alone transfer to unseen datasets and to settings orders of magnitude larger.
The article gives no link or citation for it.
The article says “These laws proved robust: they perform well on unseen datasets and settings that are several orders of magnitude higher than the set on which they have been determined”
The source says No source to compare with.
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN 2 pts, not on front page
Recorded for context only, checked 12 Nov 2026, 05:47 UTC. It played no part in the ranking.