AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

QUARK: Controllable Text Generation with Reinforced Unlearning

Ximing Lu, Sean Welleck, Jack Hessel, Liwei Jiang, Lianhui Qin, Peter West, Prithviraj Ammanabrolu, Yejin Choi

NeurIPS 2022 · Large language models › Language, safety and interpretability · approach: Controlled text generation

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
Use Quark instead of PPOGPT2-large · RealToxicityPromptsavg. max. toxicity: 0.218 -> 0.196Improved
Use Quark instead of PPOGPT2-large · RealToxicityPromptstoxic probability: 0.044 -> 0.035Improved
Use Quark instead of PPOGPT2-large · Sentiment unlearn negative% positive: 43.13 -> 46.55Improved
Use Quark instead of PPOGPT2-large · Sentiment unlearn positive% positive: 32.22 -> 27.50Improved
Add unlikelihood loss to QuarkGPT2-base · Wikitext-103rep-2: 39.89 -> 18.76Improved
Add unlikelihood loss to QuarkGPT2-base · Wikitext-103MAUVE: 0.74 -> 0.82Improved
Add unlikelihood loss to QuarkGPT2-base · Wikitext-103perplexity: 26.22 -> 27.97Hurt
Exact token-level KL vs approximate point-wise KL (ablation)Quark · RealToxicityPrompts valoutput perplexity: 13.86 -> 12.72Improved
Add exact KL term (vs none) (ablation)Quark · RealToxicityPrompts valoutput perplexity: 13.29 -> 12.72Improved
Increase KL coefficient beta (ablation)Quark · RealToxicityPrompts valoutput perplexityMixed
Increase number of quantiles (ablation)Quark · RealToxicityPrompts valtoxicityMixed
Train only on best quantile instead of all (ablation)Quark · RealToxicityPrompts valoutput perplexity: 12.72 -> 21.92Hurt

Later papers that compared with this method

1 later accepted paper reports Quark as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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