# Position-Agnostic Pre-Projection for Transformer Attention: Nonlinear Feature Construction and Content Skip Before Q/K/V

arXiv: https://arxiv.org/abs/2604.10791 · HTML: https://mlsanity.com/paper/2604.10791/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 7 of its 7 references that are in the map.
- First posted on arXiv 12 Apr 2026.

## Builds on (most cited first, 7 of 7)

- [LLaMA: Open and Efficient Foundation Language Models](https://mlsanity.com/paper/2302.13971/) (2023)
- [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)
- [Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling](https://mlsanity.com/paper/2304.01373/) (2023)
- [BERT Rediscovers the Classical NLP Pipeline](https://mlsanity.com/paper/1905.05950/) (2019)
- [The LAMBADA dataset: Word prediction requiring a broad discourse context](https://mlsanity.com/paper/1606.06031/) (2016)
- [Talking-Heads Attention](https://mlsanity.com/paper/2003.02436/) (2020)

## Abstract (arXiv)

We propose two complementary modifications to transformer attention blocks. First, a non-linear pre-projection MLP is inserted between layer norm and Q/K/V projections, constructing richer features in a position-agnostic manner before any positional encoding is applied. Second, a content skip connection routes the pre-projection's features around the attention mechanism, allowing content information to bypass position-aware attention where beneficial. In frozen-probe experiments on Pythia-160M and 410M, the combined approach achieves the strongest results across methods: +40.6% LAMBADA accuracy and -39% perplexity at 160M scale. Learned skip connection weights reveal a consistent pattern across model sizes: later transformer layers activate the content bypass more strongly than earlier layers, suggesting that deeper layers benefit from content information that does not pass through positional attention. All modifications add no K/V cache overhead.