mlsanityData to 7 Oct 2026

Aperture: Merge-Consistent Rotary States for Compressed Tokens

arXiv 2609.36781 ↗First posted 29 Sep 2026cs.AIcited by 0

On the map: Efficient LLM inference and serving → Prompt and context compression also fits: Audio and omni-modal token compression

Abstract (arXiv)

Token compression combines content from several positions, yet rotary position embeddings usually assign the merged token one coordinate. We ask what positional information must survive later merges. Aperture stores Fourier moments of the token's weighted support at the model's rotary frequencies. We prove that these moments have minimal real dimension among continuous states sufficient for the selected expected rotary interactions. Represented mass makes updates additive; attention normalisation remains a separate readout choice. Uniform intervals give a centre rotation times a sinc gain. We characterise when centres determine interval widths and construct matched examples where they do not. Numerical checks verify the weighted-support implementation. In trained temporal readers, compression transfer varies with gain calibration and feature placement. In a prespecified native video question-answering comparison, stored support reaches 65.63% accuracy versus 67.12% for the deployed merging rule. These results separate exact positional preservation under compression from downstream benefit.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in Prompt and context compression, one of 12 directions in Efficient LLM inference and serving. Builds on 30 of its 35 references that are in the map. First posted on arXiv 29 Sep 2026.

Builds on

30 of 35 references.

References · 30

  1. 2025Qwen2.5-VL Technical Reportcited by 5,418
  2. 2021RoFormer: Enhanced Transformer with Rotary Position Embeddingcited by 4,581
  3. 2024Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolutioncited by 4,151
  4. 2022Robust Speech Recognition via Large-Scale Weak Supervisioncited by 3,972
  5. 2023Efficient Streaming Language Models with Attention Sinkscited by 2,115
  6. 2021Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fieldscited by 1,957
  7. 2024Video-MME: The First-Ever Comprehensive Evaluation Benchmark of Multi-modal LLMs in Video Analysiscited by 1,375
  8. 2025SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Featurescited by 1,127
  9. 2023MVBench: A Comprehensive Multi-modal Video Understanding Benchmarkcited by 1,086
  10. 2022Token Merging: Your ViT But Fastercited by 995
  11. 2024SnapKV: LLM Knows What You are Looking for Before Generationcited by 848
  12. 2023Extending Context Window of Large Language Models via Positional Interpolationcited by 721
18 more
  1. 2025Qwen2.5-Omni Technical Reportcited by 704
  2. 2016NTU RGB+D: A Large Scale Dataset for 3D Human Activity Analysiscited by 683
  3. 2023YaRN: Efficient Context Window Extension of Large Language Modelscited by 584
  4. 2020Rescaling Egocentric Visioncited by 529
  5. 2024LLaVA-PruMerge: Adaptive Token Reduction for Efficient Large Multimodal Modelscited by 394
  6. 2016Kernel Mean Embedding of Distributions: A Review and Beyondcited by 360
  7. 2019Time2Vec: Learning a Vector Representation of Timecited by 208
  8. 2024StreamingBench: Assessing the Gap for MLLMs to Achieve Streaming Video Understandingcited by 135
  9. 2025OVO-Bench: How Far is Your Video-LLMs from Real-World Online Video Understanding?cited by 134
  10. 2024Apollo: An Exploration of Video Understanding in Large Multimodal Modelscited by 87
  11. 2025VideoRoPE: What Makes for Good Video Rotary Position Embedding?cited by 63
  12. 2019Self-attention with Functional Time Representation Learningcited by 52
  13. 2022Exact-NeRF: An Exploration of a Precise Volumetric Parameterization for Neural Radiance Fieldscited by 15
  14. 2021Enhancing Multi-Scale Implicit Learning in Image Super-Resolution with Integrated Positional Encodingcited by 12
  15. 2025PPE: Positional Preservation Embedding for Token Compression in Multimodal Large Language Modelscited by 4
  16. 2025EchoingPixels: Aliasing-Resistant Joint Token Reduction for Audio-Visual LLMscited by 4
  17. 2026StreamingEval: A Unified Evaluation Protocol towards Realistic Streaming Video Understandingcited by 3
  18. 2026TIE: Time Interval Encoding for Video Generation over Eventscited by 2

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026Unified Spatiotemporal Token Compression for Video-LLMs at Ultra-Low Retention
  2. 2026One Token per Highly Selective Frame: Towards Extreme Compression for Long Video Understanding
  3. 2025Cramming 1568 Tokens into a Single Vector and Back Again: Exploring the Limits of Embedding Space Capacity
  4. 2026Compressing Sequences in the Latent Embedding Space: K-Token Merging for Large Language Models
  5. 2025LLaVA-Scissor: Token Compression with Semantic Connected Components for Video LLMs
  6. 2024Position IDs Matter: An Enhanced Position Layout for Efficient Context Compression in Large Language Models
  7. 2026Beyond Selection: Token Parameterization for Extreme Visual Token Compression
  8. 2025Representation Shift: Unifying Token Compression with FlashAttention