Research map / Large language models / Language, safety and interpretability
Collaborative Alignment of NLP Models
Fereshte Khani, Marco Tulio Ribeiro
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 CoAlign data instead of AdaTest data when adding a conflicting concept | concept data · QQP, Corig 'X = not antonym (X)' | examples broken by new concept: 24/50 -> 7/50 | Improved |
| use CoAlign data instead of AdaTest data for the conflicting concept pair | concept data · QQP, Corig 'X = not antonym (X)' | examples fixed by new concept: 2/50 -> 5/50 | Improved |
| use CoAlign data instead of AdaTest data for similar concept pair | concept data · QQP, Corig 'X = synonym (X)' | examples broken by new concept: 18/50 -> 9/50 | Improved |
| use CoAlign data instead of AdaTest data for similar concept pair | concept data · QQP, Corig 'X = synonym (X)' | examples fixed by new concept: 18/50 -> 20/50 | Improved |
| run CoAlign on top of AdaTest-debugged data | AdaTest-debugged model · QQP 6 concepts | bugs revealed | Improved |
| use CoAlign instead of naive data collection from biased seed | BERT-base-uncased · SB (full concept) | accuracy: 80.7 -> 94.5 | Improved |
| use naive data collection from biased seed instead of base model | BERT-base-uncased · SB (full concept) | accuracy: 82.6 -> 80.7 | Hurt |
| use naive data collection vs CoAlign on the biased distribution | BERT-base-uncased · biased-SB | accuracy: 98.6 -> 94.9 | Hurt |
| use CoAlign instead of random selection | RoBERTa-Large · MNLI two concepts | overall average accuracy: 89.6 -> 91.9 | Improved |
| use CoAlign instead of uncertainty sampling | RoBERTa-Large · MNLI two concepts | overall average accuracy: 91.6 -> 91.9 | No clear effect |
| use CoAlign instead of random selection | RoBERTa-Large · MNLI two concepts | worst-case accuracy: 87.7 -> 90.7 | Improved |
| use CoAlign instead of uncertainty sampling | RoBERTa-Large · Amazon 4 categories | worst-case accuracy: 85.9 -> 86.0 | No clear effect |
Closest papers by meaning
- Generative Judge for Evaluating Alignment (ICLR 2024 · similarity 0.87)
- Aligner: Efficient Alignment by Learning to Correct (NeurIPS 2024 · similarity 0.85)
- Principled Data Selection for Alignment: The Hidden Risks of Difficult Examples (ICML 2025 · similarity 0.85)
- Progressively Label Enhancement for Large Language Model Alignment (ICML 2025 · similarity 0.85)
- PaCE: Parsimonious Concept Engineering for Large Language Models (NeurIPS 2024 · similarity 0.85)
- SPRI: Aligning Large Language Models with Context-Situated Principles (ICML 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.
