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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add UPI scaling to Mamba heads (with YaRN) (ablation) | Bamba-9B-v2 · PG-19 64k | perplexity: 98.01 -> 18.59 | Improved |
| add YaRN to Transformer layers (ablation) | Bamba-9B-v2 with UPI · PG-19 64k | perplexity: 25.79 -> 18.59 | Improved |
| restrict interpolation to selected top-K ERF heads rather than all Mamba heads (ablation) | UPI · PG-19 64k | perplexity: 65.92 -> 18.59 | Improved |
| raise the interpolated head fraction above 20-30% (ablation) | UPI · PG-19 | perplexity | Hurt |
| apply UPI + YaRN instead of base model | Bamba-9B-v2 · PG-19 64k | perplexity: 127.90 -> 18.59 | Improved |
| apply UPI instead of base model | Nemotron-H-8B · PG-19 64k | perplexity: 530.31 -> 44.01 | Improved |
| apply UPI instead of base model | Mamba2-2.7B · PG-19 64k | perplexity: 478.21 -> 22.24 | Improved |
| apply UPI + YaRN vs LongMamba + YaRN | Bamba-9B-v2 · PG-19 64k | perplexity: 53.14 -> 18.59 | Improved |
| apply UPI vs LongMamba | Mamba2-2.7B · LongBench-E | average score: 15.77 -> 14.52 | Hurt |
| apply UPI vs LongMamba | Nemotron-H-8B · LongBench-E | average score: 23.14 -> 25.15 | Improved |
| apply UPI + YaRN vs LongMamba + YaRN | Bamba-9B-v2 · LongBench-E | average score: 22.85 -> 23.68 | Improved |
| calibrate with ERF instead of searching hyperparameters | UPI vs LongMamba · calibration cost | time: ~10.5 hours -> under 3 minutes | Improved |
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.
- LongMamba — LongMamba: Enhancing Mamba's Long-Context Capabilities via Training-Free Receptive Field Enlargement (ICLR 2025) · better on 4 of 4 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
- MambaPEFT: Exploring Parameter-Efficient Fine-Tuning for Mamba (ICLR 2025 · similarity 0.89)
- Learning Rate Transfer for Hybrid Transformer-SSM Architectures (NeurIPS 2026 · similarity 0.89)
- Mamba Modulation: On the Length Generalization of Mamba Models (NeurIPS 2025 · similarity 0.88)
- Dynamical Properties of Tokens in Self-Attention and Effects of Positional Encoding (NeurIPS 2025 · similarity 0.88)
- MAP: Unleashing Hybrid Mamba-Transformer Vision Backbone's Potential with Masked Autoregressive Pretraining (CVPR 2025 · similarity 0.87)
- Transformers to SSMs: Distilling Quadratic Knowledge to Subquadratic Models (NeurIPS 2024 · similarity 0.86)
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.
