Residual connections give each Transformer block access to all earlier layers only through one fixed sum, broadcast to all of the block's inputs. This broadcast bottleneck imposes three constraints: the query, key, value and residual inputs of a block, which we call its roles, read the same vector; all hidden dimensions within a role share one set of weights over earlier layers; and every earlier layer has the same fixed weight for every token. We propose the Multi-Role Multi-Group Mixing Residual (M3), which removes all three: every role, and every group of hidden dimensions within a role, reads its own token-dependent mixture of all earlier layers. M3 starts as an exact residual connection and adds as little as 1.4% parameters and 1.8% FLOPs. From 100M to 1.4B parameters, M3Former outperforms the Transformer and coarse-grained cross-layer connections at every size, matching a Transformer 1.19-1.34x the size of the one it modifies, trained with 1.4-1.8x the compute. Relaxing only some of the constraints recovers only part of this gain. M3Former also improves sparse mixture-of-experts models of up to 4.6B parameters and trains more stably in deeper networks. Code: https://github.com/Kingchet10/m3former
Xingrun Xing, Zheng Liu, Shitao Xiao Beijing Academy of Artificial Intelligence, Boyan Gao, Yiming Liang, Haokun Lin, Xianlin Zeng, Guoqi Li Institute of automation, Chinese academy of science, Chinese Academy of Sciences 等 9 位作者
Modern large language models (LLMs) driven by scaling laws achieve emergent intelligence in large model sizes. Recently, the increasing concerns about cloud costs, latency and privacy make it an urgent requirement to develop compact edge language models. Distinguished from direct pretraining that bounded by parameter scaling law, this work proposes the unified pruning-aware pretraining, focusing on pretraining compact models while preserving performance of much larger source models, termed EfficientLLM. It features following characteristics: 1) Pruning in Pretraining Corpus: we introduce minimal parameter groups to decouple LLMs and continuously optimize model architecture with classic pruning methods like LLM-Pruner and SparseGPT during pretraining. We reveal that it achieves top-quality compact language models to scale up LLM pruning to large scale pretraining. 2) Auto-Designed Architecture: the LLM architecture is auto-designed during saliency-driven pruning, unifying pretraining, architectural design, and parameter pruning into a single process. Based on these, EfficientLLM significantly outperforms directly pretrained baselines with 100M \sim 1B parameters, such as MobileLLM, SmolLM, Qwen2.5-0.5B, OLMo-1B, Llama3.2-1B in commen sense benchmarks, which bridges the performance gap between traditional LLM compression and direct pretraining. We open source on https://github.com/Xingrun-Xing2/EfficientLLM.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表