Research map / Architectures and efficiency / Attention and transformers
Avey-B
Devang Acharya, Mohammad Hammoud
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 unidirectional instead of bidirectional ranker (ablation) | Avey-B base · QA avg | F1: 36.51 -> 51.07 | Improved |
| use unidirectional instead of bidirectional ranker (ablation) | Avey-B base · IR avg | NDCG@10: 66.20 -> 74.82 | Improved |
| use unidirectional instead of bidirectional ranker (ablation) | Avey-B base · TC avg | F1: 93.29 -> 93.38 | No clear effect |
| interleave static then dynamic layers (S->D) instead of all-dynamic (ablation) | Avey-B base · SC avg | accuracy: 79.52 -> 87.39 | Improved |
| interleave static then dynamic layers (S->D) instead of D->S (ablation) | Avey-B base · QA avg | F1: 44.67 -> 51.07 | Improved |
| interleave S->D instead of all-static (ablation) | Avey-B base · IR avg | NDCG@10: 74.95 -> 74.82 | No clear effect |
| use divide-by-sum normalization instead of RMS norm (ablation) | Avey-B base · SC avg | accuracy: 80.22 -> 87.39 | Improved |
| use divide-by-sum normalization instead of softmax (ablation) | Avey-B base · IR avg | NDCG@10: 75.34 -> 74.82 (IR); 48.51 -> 51.07 (QA) | Mixed |
| use Avey-B instead of BERT at base scale | Avey-B · TC avg | F1: 89.82 -> 93.59 | Improved |
| use Avey-B instead of RoBERTa/ModernBERT at base scale on QA | Avey-B · QA avg | F1: 75.05 -> 62.45 | Hurt |
| add neural compression of top-k splits in ranker | Avey-B · throughput | speedup: 4.37x | Improved |
Closest papers by meaning
- MosaicBERT: A Bidirectional Encoder Optimized for Fast Pretraining (NeurIPS 2023 · similarity 0.86)
- Hierarchical Autoregressive Transformers: Combining Byte- and Word-Level Processing for Robust, Adaptable Language Models (ICLR 2025 · similarity 0.85)
- Progressive Residual Warmup for Language Model Pretraining (NeurIPS 2026 · similarity 0.84)
- How Transformers Learn Regular Language Recognition: A Theoretical Study on Training Dynamics and Implicit Bias (ICML 2025 · similarity 0.84)
- Mask-Enhanced Autoregressive Prediction: Pay Less Attention to Learn More (ICML 2025 · similarity 0.84)
- Fundamental Limits of Prompt Compression: A Rate-Distortion Framework for Black-Box Language Models (NeurIPS 2024 · similarity 0.84)
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.
