Research map / Architectures and efficiency / Attention and transformers
Learned Meta-Tokens for Language Modeling
Alok Shah, Khush Gupta, Keshav Ramji, Pratik Chaudhari
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 meta-tokens and meta-attention in pretraining | GPT-2 124M with RoPE · PG19 | perplexity: 16.13 -> 14.79 | Improved |
| add meta-tokens and meta-attention | GPT-2 APE · List Recall (eval 512) | accuracy: 3.6 -> 98.7 | Improved |
| add meta-tokens and meta-attention | GPT-2 APE · Segment Counting (eval 512) | accuracy: 25.0 -> 80.9 | Improved |
| add meta-tokens and meta-attention | GPT-2 APE · Copying (eval 512) | accuracy: 7.8 -> 98.5 | Improved |
| use RoPE instead of APE with meta-tokens | Meta-attention model · Segment Counting (eval 1024) | accuracy: 33.3 -> 66.7 | Improved |
| increase fine-tune length from 4k to 8k (YaRN) | Meta-attention model · List Recall at 8k | accuracy: 3.5 -> 100.0 | Improved |
| zero positional encoding at meta-tokens (ablation) | Meta + APE (train 128) · List Recall | accuracy: 100.0 -> 99.3 | No clear effect |
| zero positional encoding at meta-tokens (ablation) | Meta + APE (train 512) · List Pointer extra-hard | accuracy: 11.1% -> 50.0% | Improved |
| zero positional encoding at meta-tokens (ablation) | Meta + APE (train 128) · Segment Counting | accuracy: 77.4 -> 63.1 | Hurt |
| zero positional encoding at meta-tokens (ablation) | Meta + APE (train 256) · Segment Counting | accuracy: 83.3 -> 88.1 | Improved |
| keep token embedding at meta-tokens (versus zeroing it) (ablation) | Meta + APE (train 128) · List Recall | accuracy: 17.4 -> 100.0 | Improved |
| keep token embedding at meta-tokens (versus zeroing it) (ablation) | Meta + RoPE (train 128) · Copying | accuracy: 6.9 -> 100.0 | Improved |
Closest papers by meaning
- The Role of Sparsity for Length Generalization in LLMs (ICML 2025 · similarity 0.89)
- A Training-Free Sub-quadratic Cost Transformer Model Serving Framework with Hierarchically Pruned Attention (ICLR 2025 · similarity 0.89)
- Beyond Next Token Prediction: Patch-Level Training for Large Language Models (ICLR 2025 · similarity 0.89)
- Unveiling Induction Heads: Provable Training Dynamics and Feature Learning in Transformers (NeurIPS 2024 · similarity 0.88)
- Compressed Context Memory for Online Language Model Interaction (ICLR 2024 · similarity 0.88)
- YaRN: Efficient Context Window Extension of Large Language Models (ICLR 2024 · similarity 0.88)
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.
