← 研踪 PaperTrailarXiv:2609.02986 · v3
Two Principles for Autoregressive Language Modeling
首次提交 2026-09-02 22:49
版本更新 2026-11-18 22:40
雷达观测 2026-11-19 15:04
摘要
We investigate the functional differentiation learned by RoPE Transformers to derive principles for autoregressive language modeling. We introduce two intervention-based metrics: RoPE Frequency Importance Score (RFIS) measures how removing a frequency contribution changes a head's full attention distribution, and RoPE Positional Dependence (RPD) isolates the effect of rotary modulation. In Qwen3 and Llama3.1, these metrics distinguish retrieval from positional modeling, while individual heads can exhibit both functions. A mid-low-frequency Global Positional Band (GPBand) separates their frequency dependencies and tracks the training sequence length in controlled experiments, suggesting that global positional fitting can hinder zero-shot length extrapolation. These findings motivate two principles. (1) Functional Organization: retrieval and positional functions should be assigned separately and mixed within layers with layer-specific allocation; and (2) Positional Dependence: positional modeling should operate only locally, while global access should rely on position-independent retrieval. Matched pretraining from scratch and targeted architectural ablations at 380M and 1.4B parameters support both principles and highlight the importance of mechanism choice in realizing each function. Together, these findings suggest a promising route toward architectural innovation grounded in the functional requirements of autoregressive language modeling.
作者备注:29 pages, 16 figures, 10 tables
观测到的版本
相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。
v4 · 2026-11-20 00:38 · 最新观测版本
Two Principles for Autoregressive Language Modeling
We investigate the functional differentiation learned by RoPE Transformers to derive principles for autoregressive language modeling. We introduce two intervention-based metrics: RoPE Frequency Importance Score (RFIS) measures how removing a frequency contribution changes a head's full attention distribution, and RoPE Positional Dependence (RPD) isolates the effect of rotary modulation. In Qwen3 and Llama3.1, these metrics distinguish retrieval from positional modeling, while individual heads can exhibit both functions. A mid-low-frequency Global Positional Band (GPBand) separates their frequency dependencies and tracks the training sequence length in controlled experiments, suggesting that global positional fitting can hinder zero-shot length extrapolation. These findings motivate two principles. (1) Functional Organization: retrieval and positional functions should be assigned separately and mixed within layers with layer-specific allocation; and (2) Positional Dependence: positional modeling should operate only locally, while global access should rely on position-independent retrieval. Matched pretraining from scratch and targeted architectural ablations at 380M and 1.4B parameters support both principles and highlight the importance of mechanism choice in realizing each function. Together, these findings suggest a promising route toward architectural innovation grounded in the functional requirements of autoregressive language modeling.
查看此版本解读 →查看此版本原文 ↗v3 · 2026-11-18 22:40
Two Principles for Autoregressive Language Modeling
We investigate the functional differentiation learned by RoPE Transformers to derive principles for autoregressive language modeling. We introduce two intervention-based metrics: RoPE Frequency Importance Score (RFIS) measures how removing a frequency contribution changes a head's full attention distribution, and RoPE Positional Dependence (RPD) isolates the effect of rotary modulation. In Qwen3 and Llama3.1, these metrics distinguish retrieval from positional modeling, while individual heads can exhibit both functions. A mid-low-frequency Global Positional Band (GPBand) separates their frequency dependencies and tracks the training sequence length in controlled experiments, suggesting that global positional fitting can hinder zero-shot length extrapolation. These findings motivate two principles. (1) Functional Organization: retrieval and positional functions should be assigned separately and mixed within layers with layer-specific allocation; and (2) Positional Dependence: positional modeling should operate only locally, while global access should rely on position-independent retrieval. Matched pretraining from scratch and targeted architectural ablations at 380M and 1.4B parameters support both principles and highlight the importance of mechanism choice in realizing each function. Together, these findings suggest a promising route toward architectural innovation grounded in the functional requirements of autoregressive language modeling.
查看此版本解读 →查看此版本原文 ↗