Research map / Large language models / Language, safety and interpretability
UNIREX: A Unified Learning Framework for Language Model Rationale Extraction
Aaron Chan, Maziar Sanjabi, Lambert Mathias, Liang Tan, Shaoliang Nie, Xiaochang Peng, Xiang Ren, Hamed Firooz
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 |
|---|---|---|---|
| Jointly train the task model and learned extractor for faithfulness and plausibility | UNIREX · Five English text classification datasets, Dataset Mean | CNRG relative gain: 32.9% | Improved |
| Use UNIREX (AA-F) instead of strongest baseline SGT | Rationale extraction without plausibility optimization · Dataset Mean | CNRG: 9.2% | Improved |
| Use UNIREX (DLM-FP) or UNIREX (SLM-FP) instead of A2R+P | Rationale extraction with plausibility optimization · Dataset Mean | CNRG: over 30% | Improved |
| Use UNIREX (DLM-FP) rationales rather than other non-oracle methods | SST user study · SST | alignment: 3.96~($\pm$0.22) vs 3.64~($\pm$0.28) | Improved |
| Use an MLP pooler instead of sum pooling (ablation) | UNIREX AA-FP · SST | AUPRC: 47.60~($\pm$2.44) -> 54.82~($\pm$1.97) | Improved |
| Use a learned Dual LM extractor instead of an attribution-algorithm extractor (ablation) | UNIREX FP · SST | AUPRC: 47.60~($\pm$2.44) -> 85.80~($\pm$0.74) | Improved |
| Optimize comprehensiveness without sufficiency (ablation) | UNIREX SLM-FP · SST | Comp: 0.302~($\pm$0.039) -> 0.350~($\pm$0.048) | Improved |
| Add sufficiency optimization to comprehensiveness optimization (ablation) | UNIREX SLM-FP · SST | Suff: 0.310~($\pm$0.049) -> 0.113~($\pm$0.013) | Improved |
| Use sufficiency alone instead of comprehensiveness plus sufficiency (ablation) | UNIREX SLM-FP · SST | Comp: 0.302~($\pm$0.039) -> 0.166~($\pm$0.003) | Hurt |
| Replace margin sufficiency loss with KL divergence (ablation) | UNIREX SLM-FP · SST | Suff: 0.113~($\pm$0.013) -> 0.131~($\pm$0.005) | Hurt |
| Replace margin sufficiency loss with MAE (ablation) | UNIREX SLM-FP · SST | Suff: 0.113~($\pm$0.013) -> 0.143~($\pm$0.008) | Hurt |
| Replace the linear extractor head with a two-layer 2048-dim MLP (ablation) | UNIREX SLM-FP · SST | Comp: 0.302~($\pm$0.039) -> 0.323~($\pm$0.071) | Improved |
Closest papers by meaning
- Learning Robust Rationales for Model Explainability: A Guidance-Based Approach (AAAI 2024 · similarity 0.87)
- Automated Assessment of Fidelity and Interpretability: An Evaluation Framework for Large Language Models’ Explanations (Student Abstract) (AAAI 2024 · similarity 0.87)
- WAGLE: Strategic Weight Attribution for Effective and Modular Unlearning in Large Language Models (NeurIPS 2024 · similarity 0.87)
- MUSE: Machine Unlearning Six-Way Evaluation for Language Models (ICLR 2025 · similarity 0.86)
- Enhancing Relation Extraction via Supervised Rationale Verification and Feedback (AAAI 2025 · similarity 0.86)
- EvaLearn: Quantifying the Learning Capability and Efficiency of LLMs via Sequential Problem Solving (NeurIPS 2025 · similarity 0.86)
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.
