# Multi-Head Attention Residuals

arXiv: https://arxiv.org/abs/2607.27230 · HTML: https://mlsanity.com/paper/2607.27230/

## Key facts

- Cited by 2 papers in the map (data through 14 Oct 2026).
- Builds on 34 of its 42 references that are in the map.
- First posted on arXiv 22 Jul 2026.

## Builds on (most cited first, 30 of 34)

- [Attention Is All You Need](https://mlsanity.com/paper/1706.03762/) (2017)
- [Deep Residual Learning for Image Recognition](https://mlsanity.com/paper/1512.03385/) (2015)
- [Training Verifiers to Solve Math Word Problems](https://mlsanity.com/paper/2110.14168/) (2021)
- [Qwen3 Technical Report](https://mlsanity.com/paper/2505.09388/) (2025)
- [Evaluating Large Language Models Trained on Code](https://mlsanity.com/paper/2107.03374/) (2021)
- [Measuring Massive Multitask Language Understanding](https://mlsanity.com/paper/2009.03300/) (2020)
- [Densely Connected Convolutional Networks](https://mlsanity.com/paper/1608.06993/) (2016)
- [Measuring Mathematical Problem Solving With the MATH Dataset](https://mlsanity.com/paper/2103.03874/) (2021)
- [HellaSwag: Can a Machine Really Finish Your Sentence?](https://mlsanity.com/paper/1905.07830/) (2019)
- [FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness](https://mlsanity.com/paper/2205.14135/) (2022)
- [Program Synthesis with Large Language Models](https://mlsanity.com/paper/2108.07732/) (2021)
- [Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer](https://mlsanity.com/paper/1701.06538/) (2017)
- [Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity](https://mlsanity.com/paper/2101.03961/) (2021)
- [GPQA: A Graduate-Level Google-Proof Q&A Benchmark](https://mlsanity.com/paper/2311.12022/) (2023)
- [Solving Quantitative Reasoning Problems with Language Models](https://mlsanity.com/paper/2206.14858/) (2022)
- [GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints](https://mlsanity.com/paper/2305.13245/) (2023)
- [The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale](https://mlsanity.com/paper/2406.17557/) (2024)
- [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)
- [Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free](https://mlsanity.com/paper/2505.06708/) (2025)
- [ReZero is All You Need: Fast Convergence at Large Depth](https://mlsanity.com/paper/2003.04887/) (2020)
- [Mixture-of-Depths: Dynamically allocating compute in transformer-based language models](https://mlsanity.com/paper/2404.02258/) (2024)
- [DeepNet: Scaling Transformers to 1,000 Layers](https://mlsanity.com/paper/2203.00555/) (2022)
- [Training Very Deep Networks](https://mlsanity.com/paper/1507.06228/) (2015)
- [Nemotron-CC: Transforming Common Crawl into a Refined Long-Horizon Pretraining Dataset](https://mlsanity.com/paper/2412.02595/) (2024)
- [Attention Residuals](https://mlsanity.com/paper/2603.15031/) (2026)
- [RealFormer: Transformer Likes Residual Attention](https://mlsanity.com/paper/2012.11747/) (2020)
- [DenseFormer: Enhancing Information Flow in Transformers via Depth Weighted Averaging](https://mlsanity.com/paper/2402.02622/) (2024)
- [MUDDFormer: Breaking Residual Bottlenecks in Transformers via Multiway Dynamic Dense Connections](https://mlsanity.com/paper/2502.12170/) (2025)
- [AI capabilities can be significantly improved without expensive retraining](https://mlsanity.com/paper/2312.07413/) (2023)

## Built on by (most cited first, 2 of 2)

- [Role-Decoupled Attention Residuals: Separating Matching and Content Retrieval Across Depth](https://mlsanity.com/paper/2608.01075/) (2026)
- [T-Router: Learning Thalamic Routing for Reasoning with Parameter-Efficient Reinforcement Learning](https://mlsanity.com/paper/2609.39109/) (2026)

## Abstract (arXiv)

Transformers propagate information across depth through a single additive residual stream: every sublayer reads only the most recent state. Attention residuals relax this by letting each sublayer attend, through a learned softmax. However, that read uses a single query shared across the entire width, so every feature subspace must read the depth history through one distribution. The cost of this forced compromise grows with how much the subspaces disagree about which layers to read, and disagreement grows with model width. We introduce Multi-Head Attention Residuals (MHAR): the routing query is reshaped into H per-subspace heads, each with its own softmax over the depth history. The read becomes block-diagonal, the reshape adds zero parameters and negligible compute, and H = 1 recovers attention residuals exactly. Trained from scratch on a deduplicated Nemotron-based anneal corpus that is quality-filtered and STEM- and code-heavy, MHAR improves validation loss over a standard Transformer at 100M, 350M, and 1B (-0.061, -0.149, and -0.140). It achieves the best result among four methods in every setting, with the gain increasing from 100M to the larger scales. The head count is a real design axis rather than a free knob: validation loss is U-shaped with respect to H, with a flat optimum at H = 4 or H = 8 across scales. We adopt H = 8 for large-scale models; over-splitting beyond this point (H = 16) consistently gives back part of the gain. A direct probe of the trained queries confirms that learned subspace disagreement is the underlying driver. Fused Triton routing kernels increase attention-residual training throughput from 0.2-0.5x to 0.55-0.88x of the baseline while maintaining near-baseline peak memory. An identity-preserving conversion using delta attention residuals supports 8B mid-training, yielding improvements of +3.2 on GSM8K and +3.1 on GPQA.