AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Frayed RoPE and Long Inputs: A Geometric Perspective

Davis Wertheimer, Aozhong Zhang, Derrick Liu, Penghang Yin, Naigang Wang

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: Attention head mechanisms

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add RoPE-ID (high-frequency RoPE on half the channels) with length-based temperature scalingRoPE-ID on Llama-1B · RULER average, length 16kaccuracy: 0.03 -> 30.83Improved
add RoPE-ID (scaling) versus tuning-free YaRN at 1BRoPE-ID on Llama-1B · RULER average, length 16kaccuracy: 30.25 -> 30.83Improved
RoPE-ID (scaling) versus YaRN at 3B, 8kRoPE-ID on Llama-3B · RULER average, length 8kaccuracy: 45.09 -> 43.39Hurt
add input-length temperature scaling to RoPE-ID (ablation)RoPE-ID on Llama-1B · RULER average, length 16kaccuracy: 14.29 -> 30.83Improved
vary temperature scaling exponent from 2.0 (ablation)RoPE-ID on Llama-1B · RULER average, length 16kaccuracy: 30.83 -> 31.15 (2.5)No clear effect
apply RoPE to half the channels without raising frequency (HalfRoPE)RoPE-ID on Llama-1B · RULER average, length 8kaccuracy: 43.07 (4k) -> 0.14 (8k)Hurt
raise RoPE frequency on all channels (High frequency)RoPE-ID on Llama-1B · RULER average, length 4kaccuracy: 39.72 -> 16.04Hurt
raise lowest frequency so wavelength is 2048 instead of 4096 (ablation)RoPE-ID on Llama-1B · RULER average, length 16kaccuracy: 16.07 -> 30.83Improved
reduce fraction of RoPE channels from 1/2 to 1/8 (ablation)RoPE-ID on Llama-1B · RULER average, length 16kaccuracy: 30.83 -> 2.18Hurt
reduce fraction of RoPE channels from 1/2 to 1/8 at short context (ablation)RoPE-ID on Llama-1B · RULER average, length 4kaccuracy: 39.15 -> 46.30Improved
highest-frequency wavelength 64 instead of 32 (ablation)RoPE-ID on Llama-1B · RULER average, length 16kaccuracy: 30.84 -> 9.46Hurt
RoPE-ID versus RoPE on short commonsense tasksRoPE-ID on Llama-3B · commonsense averageaccuracy: 51.96 -> 52.08No clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy