AI Paper Atlas

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

NeurIPS 2023 · Large language models › Language, safety and interpretability · approach: Controlled text generation · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
replace stable embeddings with per-sequence random Gaussian embeddingsTransformer LM · LookUp, Pile characcuracy: 48.50 -> 91.80Improved
random embeddingsTransformer LM · Permutation, Pile characcuracy: 27.66 -> 59.35Improved
random embeddingsTransformer LM · LookUp, Github 32kaccuracy: 4.25 -> 80.20Improved
random embeddingsTransformer LM · LookUp, Wiki-40B 32kaccuracy: 8.75 -> 59.35Improved
random embeddings, symbols sampled by token frequencyTransformer LM · LookUp, Wiki-40B characcuracy: 75.55 -> 63.45Hurt
random embeddings, symbols sampled by token frequencyTransformer LM · LookUp, Pile characcuracy: 72.80 -> 90.95Improved
random embeddings, symbols sampled by token frequencyTransformer LM · Permutation, Github characcuracy: 36.62 -> 70.77Improved
sample symbols by token frequency instead of uniformly (standard LM)Standard LM · LookUp, Pile characcuracy: 48.50 -> 72.80Improved
lexinvariant vs standard LM at context length 511Lexinvariant LM · Pile charperplexity: 3.38 vs 2.00Hurt
increase context lengthLexinvariant LM · Pile charperplexity gap: 9X -> less than 1XImproved
larger 32K vocabulary instead of char vocabulary (ablation)Lexinvariant LM · Pile perplexity gapperplexity gapHurt
code (Github) instead of Wiki-40B textLexinvariant LM · convergence speedperplexity gapImproved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy