Research map / Label-efficient and robust learning / Recognition and long-tail
Refining Language Models with Compositional Explanations
Huihan Yao, Ying Chen, Qinyuan Ye, Xisen Jin, Xiang Ren
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 |
|---|---|---|---|
| add soft-matched regularization advice R_soft on top of strict noisy labels, versus labels alone (ablation) | ReMoTE BERT-Large · HatEval->GHC target | Target F1: 37.5 -> 46.1 | Improved |
| ReMoTE explanations versus fine-tune with same-time labeled sample | BERT-Large · HatEval->GHC target | Target F1: 40.7 -> 46.1 | Improved |
| regularize both attribution and interaction rather than attribution only (ablation) | ReMoTE R_soft BERT-Base · HatEval->GHC target | Target F1: 36.8 -> 39.9 | Improved |
| add interaction regularization alone (no attribution) (ablation) | ReMoTE R_soft BERT-Base · HatEval->GHC target | Target F1: 31.8 vs 36.8 (attribution only) | Hurt |
| use only hate-labeled rules instead of all rules (ablation) | ReMoTE R_soft BERT-Base · HatEval->GHC target | Target F1: 35.7 -> 39.9 | Improved |
| add hate-labeled rules vs nonhate-labeled rules only (ablation) | ReMoTE R_soft BERT-Base · Stormfront->GHC target | Target F1: 43.2 -> 45.7 | Improved |
| fine-tune on strict-matched noisy labels instead of random labeled sample | BERT-Base · HatEval->GHC target | Target F1: 45.6 -> 49.0 | Improved |
| add R_soft regularization to C_strict fine-tuning (IG variant) (ablation) | BERT-Base · HatEval->GHC | FPRD: 80.2 -> 129.5 | Hurt |
| add R_soft regularization to C_strict fine-tuning (SOC variant) (ablation) | BERT-Base · HatEval->GHC target | Target F1: 45.1 -> 46.6 | Improved |
| increase number of explanations (ablation) | ReMoTE BERT-Base · HatEval->GHC target | Target F1 | Improved |
| ReMoTE R_soft+C_strict versus source model | RoBERTa-Base · Stormfront->GHC | FPRD: 43.8 -> 77.5 | Hurt |
Closest papers by meaning
- Automated Assessment of Fidelity and Interpretability: An Evaluation Framework for Large Language Models’ Explanations (Student Abstract) (AAAI 2024 · similarity 0.86)
- Compositional Generalization for Multi-Label Text Classification: A Data-Augmentation Approach (AAAI 2024 · similarity 0.85)
- Patchscopes: A Unifying Framework for Inspecting Hidden Representations of Language Models (ICML 2024 · similarity 0.85)
- Biases Mitigation and Expressiveness Preservation in Language Models: A Comprehensive Pipeline (Student Abstract) (AAAI 2024 · similarity 0.85)
- Training Language Models to Explain Their Own Computations (NeurIPS 2026 · similarity 0.84)
- metabench - A Sparse Benchmark of Reasoning and Knowledge in Large Language Models (ICLR 2025 · similarity 0.84)
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.
