Research map / Large language models / Language, safety and interpretability
RAZOR: Sharpening Knowledge by Cutting Bias with Unsupervised Text Rewriting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes RAZOR, an unsupervised dataset debiasing method that iteratively rewrites shortcut-prone training sentences with an LLM (GPT-3.5-Turbo) guided by TF-IDF and positional-embedding shortcut scores, then trains BERT/RoBERTa/DistilBERT on the rewritten data.
Problem
Spurious correlations (shortcuts) between tokens and labels in manually constructed NLP datasets hurt the generalization of fine-tuned models, and existing debiasing methods require prior knowledge of specific biases.
Approach
Identify likely shortcut sentences via a surface-feature shortcut space (TF-IDF significance and positional encodings), rewrite them with an LLM while preserving the label (verified by a second LLM session), and keep the least-shortcut rewrite, so surface features are balanced across classes without bias priors.
Key results
- On FEVER-Adversarial with RoBERTa, RAZOR (w/ GPT) reaches 66.66% accuracy and 73.78 F1 versus 66.99 and 73.13 for the original; CrossAug is the strongest on standard FEVER but weaker on the adversarial test set.
- Cross-dataset NLI (train SNLI, test MNLI) improves BERT F1 from 62.75 to 67.86 with RAZOR (GPT), a +5.63 accuracy / +8.14 F1 gain.
- RAZOR reduces negation-term occurrences by about 36-40% without prior bias knowledge, and the paper reports a ~2x reduction in bias-related terms.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Rewrite FEVER training data with RAZOR using GPT | BERT · FEVER-Adversarial | accuracy | Helped |
| Rewrite FEVER training data with RAZOR using GPT | BERT · FEVER-Adversarial | F1 | Helped |
| Rewrite FEVER training data with RAZOR using GPT | RoBERTa · FEVER-Adversarial | accuracy | Helped |
| Rewrite FEVER training data with RAZOR using GPT | RoBERTa · FEVER-Adversarial | F1 | Helped |
| Rewrite MNLI training data with RAZOR using GPT | BERT · MNLI (MNLI) | accuracy: +2.08 | Helped |
| Rewrite MNLI training data with RAZOR using GPT | BERT · MNLI (MNLI) | F1: +1.44 | Helped |
| Rewrite MNLI training data with RAZOR using GPT | BERT · SNLI (MNLI) | F1: +0.78 | Helped |
| Rewrite SNLI training data with RAZOR using GPT | BERT · MNLI (SNLI) | F1: +8.14 | Helped |
| Rewrite SNLI training data with RAZOR using GPT | BERT · SNLI (SNLI) | accuracy: +1.69 | Helped |
| Rewrite SNLI training data with RAZOR using GPT | BERT · SNLI (SNLI) | F1: +1.55 | Helped |
| Use RAZOR's surface shortcut space instead of model token embeddings (ablation) | RAZOR on FEVER · FEVER | accuracy | Helped |
| Use RAZOR's surface shortcut space instead of model token embeddings (ablation) | RAZOR on FEVER · FEVER | F1 | Helped |
Surprising findings
- CrossAug achieves the best standard FEVER scores but falls below RAZOR on FEVER-Adversarial, which the authors attribute to shortcut-containing generated pairs.
- The rewritten training set's BLEU against the FEVER test set decreases (13.33 to 12.02) while it increases on FEVER-Adversarial (11.02 to 22.41).
Limitations
- Rewriting relies on closed-source GPT-3.5-Turbo API calls, with added cost and reproducibility concerns.
- Reported gains are modest on some settings and the standard deviations overlap with baselines.
- Theoretical lemma relies on assumptions about attention norms and is only argued in the attention setting rather than for generic attribution.
Experimental setup
Datasets: FEVER FEVER-Adversarial MNLI SNLI
Benchmarks: FEVER FEVER-Adversarial MNLI SNLI
Metrics: accuracy F1 BLEU negation term count
Baselines: EDA CrossAug ReW PoE CLEVER
Builds on: CLEVER CrossAug EDA ReW Product of Experts
Closest papers by meaning
- Any Large Language Model Can Be a Reliable Judge: Debiasing with a Reasoning-based Bias Detector (NeurIPS 2025 · similarity 0.86)
- Debiasing Algorithm through Model Adaptation (ICLR 2024 · similarity 0.85)
- Self-correcting for Debiasing Large Language Models (ICML 2026 · similarity 0.85)
- Discovering Bias in Latent Space: An Unsupervised Debiasing Approach (ICML 2024 · similarity 0.85)
- Biases Mitigation and Expressiveness Preservation in Language Models: A Comprehensive Pipeline (Student Abstract) (AAAI 2024 · similarity 0.85)
- SEBRA : Debiasing through Self-Guided Bias Ranking (ICLR 2025 · similarity 0.85)
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.