Files
myWiki/papers/tapered-language-models.md

60 lines
2.5 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

---
title: "Tapered Language Models"
created: 2026-06-29
updated: 2026-06-29
type: paper
tags: [language-model, architecture, transformer, mlp, efficiency, depth-aware]
sources: [https://arxiv.org/abs/2606.23670]
authors: ["Reza Bayat (Mila)", "Ali Behrouz (Cornell)", "Aaron Courville (Mila / UdeM / CIFAR)"]
venue: arXiv:2606.23670v1
year: 2026
---
# Tapered Language Models (TLMs)
> arXiv: [2606.23670v1](https://arxiv.org/abs/2606.23670), cs.LG, June 2026
## 一句话
现代 LLM 对所有层**均等分配参数**——这是从原始 Transformer 继承的默认设置,从未被质疑。本文发现**早期层需要更多容量、后期层只需精化残差流**,提出 [[depth-aware-capacity-allocation|深度感知容量分配]]:在固定总参数量下,将 MLP 宽度从前向后单调递减taper零额外成本提升 perplexity 和下游性能。
## 核心发现
### 不对称性证据
后期层对输出的贡献是**精化refine残差流**,而非像早期层那样进行大幅度变换。因此:
- **更多容量给早期层** → perplexity 改善
- **更多容量给后期层** → 反而**损害** perplexity
### Tapered Language Model (TLM)
在固定总参数预算下将某一参数承载组件MLP 宽度)沿深度**单调递减**
- MLP 是自然的 taper 目标:支配所有 LM 家族的参数量,宽度是单一、干净的调节轴
- 推荐使用 [[cosine-taper-schedule|余弦衰减调度]]
## 实验结果
| 规模 | 架构 | 结果 |
|------|------|------|
| 440M / 1B / 3B | Transformer | Cosine taper 一致优于 uniform baseline |
| 440M | Gated Attention | 同上 |
| 440M | Hope-attention | 同上 |
| 440M | Titans | 同上 |
- **零额外参数、零额外计算**
- 440M Transformeruniform 16.28 → cosine taper **14.44**(改善 1.84 perplexity
- 最优 taper 范围1.50× → 0.50× baseline FF width
- U 形曲线:过强或过弱的 taper 均不如中间值
## 架构无关性
TLM 原则适用于Transformer、Gated Attention、Hope-attention、Titans 四种异构架构——说明深度感知容量分配是**跨架构的通用设计轴**,一个"藏在眼皮底下的免费杠杆"。
## 相关概念
- [[depth-aware-capacity-allocation|深度感知容量分配]]
- [[mlp-width-tapering|MLP 宽度渐缩]]
- [[cosine-taper-schedule|余弦衰减调度]]
- [[subquadratic-transformer-alternatives|次二次方 Transformer 替代]]
- [[recurrent-transformer-architectures|循环 Transformer 架构]]