Research map / Label-efficient and robust learning / Recognition and long-tail
GAPX: Generalized Autoregressive Paraphrase-Identification X
Yifei Zhou, Renyu Li, Hayden Housen, Ser-Nam Lim
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 |
|---|---|---|---|
| use only the positive model (drop negative model) at inference (ablation) | autoregressive paraphrase model · OOD average | macro F1: 52.9 -> 69.7 | Improved |
| use only the positive model (drop negative model) in-distribution (ablation) | autoregressive paraphrase model · in-distribution average | macro F1: 75.6 -> 60.4 | Hurt |
| replace fixed OODP with perplexity-weighted automatic ensembling (GAP) (ablation) | OODP · OOD average | macro F1: 69.7 -> 70.0 | No clear effect |
| replace fixed OODP with perplexity-weighted ensembling (GAP) (ablation) | OODP · in-distribution average | macro F1: 60.4 -> 68.8 | Improved |
| ensemble GAP with RoBERTa discriminative model (GAPX) (ablation) | GAP · in-distribution average | macro F1: 68.8 -> 78.1 | Improved |
| ensemble GAP with RoBERTa discriminative model (GAPX) (ablation) | GAP · OOD average | macro F1: 70.0 -> 67.7 | Hurt |
| use GAPX instead of best transformer baseline BART | GAPX · OOD average | macro F1: 56.2 -> 67.7 | Improved |
| add expert-product debiasing to BERT | BERT · OOD average | macro F1: 54.4 -> 48.7 | Hurt |
| replace perplexity with ENERGY OOD score for ensembling (ablation) | GAPX · OOD average | macro F1: 67.7 -> 60.6 | Hurt |
| replace perplexity with SOFTMAX OOD score for ensembling (ablation) | GAPX · OOD average | macro F1: 67.7 -> 55.3 | Hurt |
| replace perplexity with negative log-likelihood (ablation) | GAPX · OOD average | macro F1: 67.7 -> 66.8 | No clear effect |
| ensemble GAP with IDP instead of RoBERTa (ablation) | GAP · OOD average | macro F1: 67.7 -> 64.2 | Hurt |
Closest papers by meaning
- Paraphrasing evades detectors of AI-generated text, but retrieval is an effective defense (NeurIPS 2023 · similarity 0.86)
- Paraphrase Diversification Using Counterfactual Debiasing (AAAI 2019 · similarity 0.85)
- DNA-GPT: Divergent N-Gram Analysis for Training-Free Detection of GPT-Generated Text (ICLR 2024 · similarity 0.83)
- DiffScore: Text Evaluation Beyond Autoregressive Likelihood (NeurIPS 2026 · similarity 0.83)
- Generative Judge for Evaluating Alignment (ICLR 2024 · similarity 0.83)
- Weak-to-Strong Search: Align Large Language Models via Searching over Small Language Models (NeurIPS 2024 · similarity 0.82)
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.
