Research map / Architectures and efficiency / Attention and transformers
VCC: Scaling Transformers to 128K Tokens or More by Prioritizing Important Tokens
Zhanpeng Zeng, Cole Hawkins, Mingyi Hong, Aston Zhang, Nikolaos Pappas, Vikas Singh, Shuai Zheng
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 |
|---|---|---|---|
| scale input length from 16K to 128K (ablation) | VCC T5-base · NarrativeQA | EM: 5.9 -> 8.0 | Improved |
| scale input length from 16K to 128K (ablation) | VCC T5-base · NarrativeQA | F1: 16.6 -> 19.6 | Improved |
| use k=64 instead of k=16 at 64K length (ablation) | VCC T5-base · NarrativeQA | EM: 7.5 -> 7.2 | Hurt |
| use k=32 vs k=16 at 32K length (ablation) | VCC T5-base · NarrativeQA | EM: 6.4 -> 6.6 | No clear effect |
| larger pretraining run (VCC*) vs VCC (ablation) | VCC RoBERTa-base 4K · HotpotQA | F1: 74.6 -> 75.0 | Improved |
| scale VCC* from base to large | VCC* RoBERTa 4K · QuALITY | accuracy: 39.5 -> 41.8 | Improved |
| scale VCC from base to 3B | VCC T5 16K · QuALITY | accuracy: 36.5 -> 45.2 | Improved |
| VCC vs Longformer at 4K | RoBERTa-base · QuALITY | accuracy: 27.9 -> 39.6 | Improved |
| VCC vs RoBERTa 4K full attention | RoBERTa-base · WikiHop | runtime (ms): 414.1 -> 108.0 | Improved |
| VCC vs RoBERTa 4K full attention | RoBERTa-base · HotpotQA | F1: 76.1 -> 74.6 | Hurt |
| VCC vs LongT5 at 16K | T5-base · Arxiv | encoder-decoder runtime (ms): 1699.7 -> 672.8 | Improved |
| VCC vs T5 4K at same accuracy | T5-base · WikiHop | runtime (ms): 594.3 -> 181.7 | 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.
- Big Bird 4K — Big Bird: Transformers for Longer Sequences (NeurIPS 2020) · 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
- Vision-centric Token Compression in Large Language Model (NeurIPS 2025 · similarity 0.88)
- VQToken: Neural Discrete Token Representation Learning for Extreme Token Reduction in Video Large Language Models (NeurIPS 2025 · similarity 0.87)
- One Token per Highly Selective Frame: Towards Extreme Compression for Long Video Understanding (NeurIPS 2025 · similarity 0.87)
- Inducing High Energy-Latency of Large Vision-Language Models with Verbose Images (ICLR 2024 · similarity 0.87)
- Distilling Sequential Computation in Transformer Language Models (NeurIPS 2026 · similarity 0.87)
- On the Optimal Memorization Capacity of Transformers (ICLR 2025 · similarity 0.87)
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.
