Research map / Architectures and efficiency / Attention and transformers
MeSH: Memory-as-State-Highways for Recursive Transformers
Chengting Yu, Xiaobo Shu, Yadao Wang, Yizhen Zhang, Haoyi Wu, Jiaang Li, Rujiao Long, Ziheng Chen, Yuchi Xu, Wenbo Su, Bo 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 |
|---|---|---|---|
| add MeSH memory buffer and routers to recursive transformer | Pythia-1.4B recursive (4+8R2+4) · Task average 0-shot | average accuracy: 48.89 -> 50.56 | Improved |
| add MeSH to recursive transformer | Pythia-1.4B recursive (4+8R2+4) · Pile validation | perplexity: 7.63 -> 7.39 | Improved |
| add MeSH to recursive transformer | Pythia-1B recursive (3+5R2+3) · Pile validation | perplexity: 8.20 -> 7.90 | Improved |
| add MeSH to recursive transformer | Pythia-160M recursive (2+4R2+2) · Pile validation | perplexity: 11.79 -> 11.37 | Improved |
| add MeSH to recursive transformer | Pythia-410M recursive (3+6R3+3, 50% fewer params) · Task average 0-shot | average accuracy: 41.94 -> 43.53 | Improved |
| add MeSH to recursive transformer | Pythia-410M recursive (3+6R3+3, 50% fewer params) · Lambada Standard | perplexity: 81.75 -> 52.07 | Improved |
| add anchor connection to recursive transformer | Pythia-1.4B recursive · Task average 0-shot | average accuracy: 48.89 -> 49.39 | No clear effect |
| add residual connection to recursive transformer | Pythia-1B recursive · Task average 0-shot | average accuracy: 45.72 -> 46.19 | Improved |
| add residual connection to recursive transformer | Pythia-410M recursive (3+6R3+3) · Pile validation | perplexity: 9.65 -> 9.69 | Hurt |
| use MeSH recursive model instead of larger Vanilla | Pythia-1.4B · Task average 0-shot | average accuracy: 49.50 -> 50.56 | Improved |
| use MeSH recursive model instead of larger Vanilla | Pythia-160M · Task average 0-shot | average accuracy: 39.88 -> 39.41 | Hurt |
| add MeSH to recursive transformer | Pythia-1.4B recursive · Task average 5-shot | average accuracy: 50.99 -> 52.79 | Improved |
Closest papers by meaning
- Predictive Geometry of Hidden Trajectories in Transformers (NeurIPS 2026 · similarity 0.87)
- Exact Expressive Power of Transformers with Padding (NeurIPS 2025 · similarity 0.86)
- Stability and Generalization in Looped Transformers (NeurIPS 2026 · similarity 0.86)
- Separations in the Representational Capabilities of Transformers and Recurrent Architectures (NeurIPS 2024 · similarity 0.86)
- Mixture-of-Recursions: Learning Dynamic Recursive Depths for Adaptive Token-Level Computation (NeurIPS 2025 · similarity 0.85)
- MUDDFormer: Breaking Residual Bottlenecks in Transformers via Multiway Dynamic Dense Connections (ICML 2025 · similarity 0.85)
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.
