Attention Residuals

TL;DR

This paper introduces Attention Residuals (AttnRes), a novel method replacing fixed residual accumulation with input-dependent attention, improving model depth handling and performance in large language models.

Contribution

It proposes AttnRes and Block AttnRes, enabling selective, content-dependent aggregation of layer outputs, reducing dilution and overhead in large-scale models.

Findings

AttnRes improves uniformity of output magnitudes and gradients.

Block AttnRes reduces memory overhead while maintaining performance gains.

Pre-training with AttnRes enhances downstream task performance.

Abstract

Residual connections with PreNorm are standard in modern LLMs, yet they accumulate all layer outputs with fixed unit weights. This uniform aggregation causes uncontrolled hidden-state growth with depth, progressively diluting each layer's contribution. We propose Attention Residuals (AttnRes), which replaces this fixed accumulation with softmax attention over preceding layer outputs, allowing each layer to selectively aggregate earlier representations with learned, input-dependent weights. To address the memory and communication overhead of attending over all preceding layer outputs for large-scale model training, we introduce Block AttnRes, which partitions layers into blocks and attends over block-level representations, reducing the memory footprint while preserving most of the gains of full AttnRes. Combined with cache-based pipeline communication and a two-phase computation strategy, Block AttnRes becomes a practical drop-in replacement for standard residual connections with minimal overhead. Scaling law experiments confirm that the improvement is consistent across model sizes, and ablations validate the benefit of content-dependent depth-wise selection. We further integrate AttnRes into the Kimi Linear architecture (48B total / 3B activated parameters) and pre-train on 1.4T tokens, where AttnRes mitigates PreNorm dilution, yielding more uniform output magnitudes and gradient distribution across depth, and improves downstream performance across all evaluated tasks.