Research map / Adversarial robustness and security / Attacks on text and graphs
TextGrad: Advancing Robustness Evaluation in NLP by Gradient-Driven Optimization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 TextGrad rather than the strongest ASR baseline | BERT attack generation · SST-2 | ASR: 87.79 -> 93.51 | Helped |
| Use TextGrad rather than the strongest ASR baseline | BERT attack generation · MNLI-m | ASR: 93.37 -> 94.08 | Helped |
| Use TextGrad rather than the strongest ASR baseline | BERT attack generation · RTE | ASR: 68.20 -> 71.91 | Helped |
| Use TextGrad rather than the strongest ASR baseline | BERT attack generation · QNLI | ASR: 67.95 -> 70.48 | Helped |
| Use TextGrad rather than the strongest ASR baseline | BERT attack generation · AG News | ASR: 74.70 -> 74.51 | Mixed |
| Add random restarts (ablation) | TextGrad · SST-2, MNLI-m, RTE, QNLI, AG News | ASR: 90.27 -> 93.51; 91.90 -> 94.08; 67.42 -> 71.91; 65.26 -> 70.48; 68.87 -> 74.51 | Helped |
| Add randomized sampling alongside random restarts (ablation) | TextGrad · SST-2, MNLI-m, RTE, QNLI, AG News | ASR: 78.30 -> 93.51; 85.44 -> 94.08; 58.99 -> 71.91; 55.05 -> 70.48; 54.23 -> 74.51 | Helped |
| Train with TextGrad adversarial training instead of TextFooler-AT | BERT on SST-2 · SST-2 | Avg RA: 32.24 -> 41.34 | Helped |
Closest papers by meaning
- Word Level Robustness Enhancement: Fight Perturbation with Perturbation (AAAI 2022 · similarity 0.90)
- Unifying Adversarial Robustness and Training Across Text Scoring Models (ICML 2026 · similarity 0.89)
- Towards Robustness Against Natural Language Word Substitutions (ICLR 2021 · similarity 0.89)
- Is BERT Really Robust? A Strong Baseline for Natural Language Attack on Text Classification and Entailment (AAAI 2020 · similarity 0.89)
- Scaling Trends in Language Model Robustness (ICML 2025 · similarity 0.87)
- Revisiting Character-level Adversarial Attacks for Language Models (ICML 2024 · 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.