Search papers, labs, and topics across Lattice.
This paper introduces Multi-Head Attention Residuals (MHAR), an enhancement to the traditional Transformer architecture that allows each sublayer to attend to depth history through multiple learned softmax distributions, rather than a single shared query. By reshaping the routing query into multiple heads per subspace, MHAR effectively mitigates the information loss that occurs when subspaces disagree on which layers to read, leading to improved validation loss across various model sizes. The results demonstrate that MHAR consistently outperforms standard Transformers, with the most significant gains observed at larger scales, suggesting that the head count is a critical design parameter for optimizing performance.
Multi-Head Attention Residuals achieve superior validation loss by allowing Transformers to leverage multiple attention heads, revealing that subspace disagreement is a key factor in model performance.
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.
