# Most Transformer Modifications Still Do Not Transfer at 1-3B: A 2020-2026 Update to Narang et al. (2021) with Downstream Evaluation and a Noise Floor

arXiv: https://arxiv.org/abs/2605.20798 · HTML: https://mlsanity.com/paper/2605.20798/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Builds on 23 of its 43 references that are in the map.
- First posted on arXiv 20 May 2026.

## Builds on (most cited first, 23 of 23)

- [Measuring Massive Multitask Language Understanding](https://mlsanity.com/paper/2009.03300/) (2020)
- [Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge](https://mlsanity.com/paper/1803.05457/) (2018)
- [HellaSwag: Can a Machine Really Finish Your Sentence?](https://mlsanity.com/paper/1905.07830/) (2019)
- [TruthfulQA: Measuring How Models Mimic Human Falsehoods](https://mlsanity.com/paper/2109.07958/) (2021)
- [PIQA: Reasoning about Physical Commonsense in Natural Language](https://mlsanity.com/paper/1911.11641/) (2019)
- [BoolQ: Exploring the Surprising Difficulty of Natural Yes/No Questions](https://mlsanity.com/paper/1905.10044/) (2019)
- [Can a Suit of Armor Conduct Electricity? A New Dataset for Open Book Question Answering](https://mlsanity.com/paper/1809.02789/) (2018)
- [GLU Variants Improve Transformer](https://mlsanity.com/paper/2002.05202/) (2020)
- [The LAMBADA dataset: Word prediction requiring a broad discourse context](https://mlsanity.com/paper/1606.06031/) (2016)
- [On Layer Normalization in the Transformer Architecture](https://mlsanity.com/paper/2002.04745/) (2020)
- [RACE: Large-scale ReAding Comprehension Dataset From Examinations](https://mlsanity.com/paper/1704.04683/) (2017)
- [Going deeper with Image Transformers](https://mlsanity.com/paper/2103.17239/) (2021)
- [CogView: Mastering Text-to-Image Generation via Transformers](https://mlsanity.com/paper/2105.13290/) (2021)
- [Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free](https://mlsanity.com/paper/2505.06708/) (2025)
- [Query-Key Normalization for Transformers](https://mlsanity.com/paper/2010.04245/) (2020)
- [Primer: Searching for Efficient Transformers for Language Modeling](https://mlsanity.com/paper/2109.08668/) (2021)
- [Same Pre-training Loss, Better Downstream: Implicit Bias Matters for Language Models](https://mlsanity.com/paper/2210.14199/) (2022)
- [Nemotron-CLIMB: CLustering-based Iterative Data Mixture Bootstrapping for Language Model Pre-training](https://mlsanity.com/paper/2504.13161/) (2025)
- [Theory, Analysis, and Best Practices for Sigmoid Self-Attention](https://mlsanity.com/paper/2409.04431/) (2024)
- [DenseFormer: Enhancing Information Flow in Transformers via Depth Weighted Averaging](https://mlsanity.com/paper/2402.02622/) (2024)
- [Scalable-Softmax Is Superior for Attention](https://mlsanity.com/paper/2501.19399/) (2025)
- [Scaling Laws Are Unreliable for Downstream Tasks: A Reality Check](https://mlsanity.com/paper/2507.00885/) (2025)
- [HybridNorm: Towards Stable and Efficient Transformer Training via Hybrid Normalization](https://mlsanity.com/paper/2503.04598/) (2025)

## Built on by (most cited first, 1 of 1)

- [Abstention and Noise Filtering: Two Missing Primitives of Softmax Attention](https://mlsanity.com/paper/2609.22005/) (2026)

## Abstract (arXiv)

Narang et al. (2021) evaluated 40+ Transformer modifications at T5-base scale and concluded that most did not transfer. Five years later, the typical working regime has moved to 1-3B parameters, downstream evaluation has replaced pretraining perplexity, and a substantially different catalogue of modifications has emerged. We revisit their question by testing 20 post-2021 Transformer modifications at 1.2B and 3B under strict iso-data, iso-compute, iso-recipe control, with a multi-seed baseline noise floor and CLIMB-12 downstream evaluation as the primary metric. The central finding reproduces theirs at this curated set: most modifications do not transfer. Of the 20 modifications, only two clear Bonferroni correction at 1.2B; one of those two further fails to train stably at 3B under the shared recipe. We also find that the loss-downstream gap reported by Tay et al. (2023) enlarges several-fold for attention-output modifications: two significant failures converge to within 2-3% of baseline validation loss yet drop 6-16 CLIMB-points. We conclude that noise-floor reporting, downstream evaluation, and cross-scale stability testing are now prerequisites for architecture comparisons at 1-3B.