Research map / Large language models / Language, safety and interpretability
Lexinvariant Language Models
Qian Huang, Eric Zelikman, Sarah Li Chen, Yuhuai Wu, Gregory Valiant, Percy Liang
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 |
|---|---|---|---|
| replace stable embeddings with per-sequence random Gaussian embeddings | Transformer LM · LookUp, Pile char | accuracy: 48.50 -> 91.80 | Improved |
| random embeddings | Transformer LM · Permutation, Pile char | accuracy: 27.66 -> 59.35 | Improved |
| random embeddings | Transformer LM · LookUp, Github 32k | accuracy: 4.25 -> 80.20 | Improved |
| random embeddings | Transformer LM · LookUp, Wiki-40B 32k | accuracy: 8.75 -> 59.35 | Improved |
| random embeddings, symbols sampled by token frequency | Transformer LM · LookUp, Wiki-40B char | accuracy: 75.55 -> 63.45 | Hurt |
| random embeddings, symbols sampled by token frequency | Transformer LM · LookUp, Pile char | accuracy: 72.80 -> 90.95 | Improved |
| random embeddings, symbols sampled by token frequency | Transformer LM · Permutation, Github char | accuracy: 36.62 -> 70.77 | Improved |
| sample symbols by token frequency instead of uniformly (standard LM) | Standard LM · LookUp, Pile char | accuracy: 48.50 -> 72.80 | Improved |
| lexinvariant vs standard LM at context length 511 | Lexinvariant LM · Pile char | perplexity: 3.38 vs 2.00 | Hurt |
| increase context length | Lexinvariant LM · Pile char | perplexity gap: 9X -> less than 1X | Improved |
| larger 32K vocabulary instead of char vocabulary (ablation) | Lexinvariant LM · Pile perplexity gap | perplexity gap | Hurt |
| code (Github) instead of Wiki-40B text | Lexinvariant LM · convergence speed | perplexity gap | Improved |
Closest papers by meaning
- The Foundations of Tokenization: Statistical and Computational Concerns (ICLR 2025 · similarity 0.89)
- Token Embeddings Violate the Manifold Hypothesis (NeurIPS 2025 · similarity 0.88)
- Interchangeable Token Embeddings for Extendable Vocabulary and Alpha-Equivalence (ICML 2025 · similarity 0.88)
- On the Origins of Linear Representations in Large Language Models (ICML 2024 · similarity 0.87)
- Rep2Text: Decoding Full Text from a Single LLM Token Representation (NeurIPS 2026 · similarity 0.87)
- Language Models over Canonical Byte-Pair Encodings (ICML 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.
