Research map / Vision-language and multimodal / Cross-modal learning
EchoingPixels: Aliasing-Resistant Joint Token Reduction for Audio-Visual LLMs
Chao Gong, Depeng Wang, Zhipeng Wei, Ya Guo, Huijia Zhu, Jingjing Chen
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 |
|---|---|---|---|
| use bidirectional attention in the CS2 encoder instead of causal (ablation) | CS2 encoder · Daily-Omni (20% budget ablation) | overall accuracy: 56.06 -> 60.65 | Improved |
| use joint cross-modal attention instead of intra-modal attention (ablation) | CS2 encoder · Daily-Omni (20% budget ablation) | overall accuracy: 57.73 -> 60.65 | Improved |
| add text instruction pre-fusion (ablation) | CS2 encoder · Daily-Omni (20% budget ablation) | overall accuracy: 57.39 -> 60.65 | Improved |
| replace random selection with learned scorer trained by STE (ablation) | token selection · Daily-Omni (20% budget ablation) | overall accuracy: 57.81 -> 60.65 | Improved |
| replace similarity-to-text heuristic with learned scorer (ablation) | token selection · Daily-Omni (20% budget ablation) | overall accuracy: 41.85 -> 60.65 | Improved |
| use STE instead of Gumbel-Softmax (ablation) | token selection · Daily-Omni (20% budget ablation) | overall accuracy: 59.06 -> 60.65 | Improved |
| use joint pool with dynamic budget instead of per-modality fixed ratio (ablation) | token selection · Daily-Omni (20% budget ablation) | overall accuracy: 58.23 -> 60.65 | Improved |
| replace TMRoPE with Sync-RoPE [18,18,18,10] (ablation) | positional encoding · Daily-Omni (20% budget ablation) | overall accuracy: 57.98 -> 60.65 | Improved |
| reallocate even 4 low-frequency channels to time ([20,20,20,4]) (ablation) | Sync-RoPE · Daily-Omni (20% budget ablation) | overall accuracy: 57.98 -> 59.65 | Improved |
| use [18,18,18,10] rather than [16,16,16,16] split (ablation) | Sync-RoPE · Daily-Omni (20% budget ablation) | overall accuracy: 59.31 -> 60.65 | Improved |
| use Sync-RoPE instead of interleaved TMRoPE-I (ablation) | positional encoding · Daily-Omni (20% budget ablation) | overall accuracy: 57.89 -> 60.65 | Improved |
| increase CS2 encoder depth from 3 to 4 layers (ablation) | CS2 encoder · Daily-Omni | accuracy: 51.04 -> 60.65 | Improved |
Closest papers by meaning
- Improving Visual Token Reduction via Rectifying Distortions for Efficient Multimodal LLM Inference (ICML 2026 · similarity 0.87)
- OmniSelect: Dynamic Modality-Aware Token Compression for Efficient Omni-modal Large Language Models (NeurIPS 2026 · similarity 0.87)
- ALMTokenizer: A Low-bitrate and Semantic-rich Audio Codec Tokenizer for Audio Language Modeling (ICML 2025 · similarity 0.86)
- AVOC: Enhancing Hour-Level Audio-Video Understanding in Omni-Modal LLMs via Retrieval-Inspired Token Compression (NeurIPS 2026 · similarity 0.86)
- EchoPrune: Interpreting Redundancy as Temporal Echoes for Efficient VideoLLMs (NeurIPS 2026 · similarity 0.86)
- Filter, Correlate, Compress: Training-Free Token Reduction for MLLM Acceleration (AAAI 2026 · 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.
