AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

From Collapse to Control: Understanding and Extending Context Length in Emerging Hybrid Models via Universal Position Interpolation

Haochen Shen, Davis Wertheimer, Zheng Wang, Garrett Goon, Derrick Liu, Naigang Wang, Mudhakar Srivatsa, Raghu K. Ganti, Minjia Zhang

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: Sparse efficient transformers for long context

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 UPI scaling to Mamba heads (with YaRN) (ablation)Bamba-9B-v2 · PG-19 64kperplexity: 98.01 -> 18.59Improved
add YaRN to Transformer layers (ablation)Bamba-9B-v2 with UPI · PG-19 64kperplexity: 25.79 -> 18.59Improved
restrict interpolation to selected top-K ERF heads rather than all Mamba heads (ablation)UPI · PG-19 64kperplexity: 65.92 -> 18.59Improved
raise the interpolated head fraction above 20-30% (ablation)UPI · PG-19perplexityHurt
apply UPI + YaRN instead of base modelBamba-9B-v2 · PG-19 64kperplexity: 127.90 -> 18.59Improved
apply UPI instead of base modelNemotron-H-8B · PG-19 64kperplexity: 530.31 -> 44.01Improved
apply UPI instead of base modelMamba2-2.7B · PG-19 64kperplexity: 478.21 -> 22.24Improved
apply UPI + YaRN vs LongMamba + YaRNBamba-9B-v2 · PG-19 64kperplexity: 53.14 -> 18.59Improved
apply UPI vs LongMambaMamba2-2.7B · LongBench-Eaverage score: 15.77 -> 14.52Hurt
apply UPI vs LongMambaNemotron-H-8B · LongBench-Eaverage score: 23.14 -> 25.15Improved
apply UPI + YaRN vs LongMamba + YaRNBamba-9B-v2 · LongBench-Eaverage score: 22.85 -> 23.68Improved
calibrate with ERF instead of searching hyperparametersUPI vs LongMamba · calibration costtime: ~10.5 hours -> under 3 minutesImproved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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