Research map / Architectures and efficiency / Attention and transformers
Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing
Zihang Dai, Guokun Lai, Yiming Yang, Quoc V. Le
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 |
|---|---|---|---|
| trade sequence length for additional encoder depth | Funnel-Transformer versus L12H768 · GLUE-AVG dev, base-scale MLM | score: 84.4 -> 85.3 | Improved |
| use Funnel-Transformer compression | Funnel-Transformer versus MPNet Base · GLUE-AVG dev, large-scale ELECTRA | score: 87.7 -> 88.3 | Improved |
| use Funnel-Transformer compression | Funnel-Transformer versus MPNet Base · RACE test | accuracy: 72.0 -> 79.7 | Improved |
| use Funnel-Transformer with decoder | Funnel-Transformer versus ELECTRA Large · SQuAD2.0 dev | EM: 88.0 -> 87.6 | Hurt |
| remove the separate, unpooled [cls] state (ablation) | Funnel-Transformer · GLUE-AVG dev | score: 83.5 -> 82.9 | Hurt |
| remove pool-query-only attention by using pooled states for keys and values too (ablation) | Funnel-Transformer · GLUE-AVG dev | score: 83.5 -> 83.0 | Hurt |
| replace relative attention with absolute position embeddings (ablation) | Funnel-Transformer · GLUE-AVG dev | score: 83.5 -> 81.4 | Hurt |
| replace mean pooling with max pooling (ablation) | Funnel-Transformer · GLUE-AVG dev | score: 83.5 -> 83.4 | No clear effect |
| replace mean pooling with Top-Attn pooling (ablation) | Funnel-Transformer · GLUE-AVG dev | score: 83.5 -> 75.8 | Hurt |
| use a 2-block layout instead of 3 blocks (ablation) | Funnel-Transformer · GLUE-AVG dev | score: 83.5 -> 83.4 | Hurt |
| use a 4-block layout instead of 3 blocks (ablation) | Funnel-Transformer · GLUE-AVG dev | score: 83.5 -> 82.9 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- ELECTRA Large — ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators (ICLR 2020) · better on 0 of 1 shared result
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
- Distilling Sequential Computation in Transformer Language Models (NeurIPS 2026 · similarity 0.88)
- Beyond Attention: Breaking the Limits of Transformer Context Length with Recurrent Memory (AAAI 2024 · similarity 0.88)
- Transformers Get Stable: An End-to-End Signal Propagation Theory for Language Models (ICML 2024 · similarity 0.88)
- Compress, Gather, and Recompute: REFORMing Long-Context Processing in Transformers (NeurIPS 2025 · similarity 0.87)
- Curse of High Dimensionality Issue in Transformer for Long Context Modeling (ICML 2025 · similarity 0.87)
- Bigram Subnetworks: Mapping to Next Tokens in Transformer Language Models (NeurIPS 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.
