Research map / Label-efficient and robust learning / Recognition and long-tail
Prefer to Classify: Improving Text Classifiers via Auxiliary Preference Learning
Jaehyung Kim, Jinwoo Shin, Dongyeop Kang
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 auxiliary preference learning (P2C) with GPT-3 generative preference labels | RoBERTa-base classifier · CoLA, SMS Spam, Hate Speech, Emotion | test accuracy / Mcc / ECE: e.g., Mcc 63.7 to 65.4; ECE 3.6 to 2.8 on CoLA; ~11.55% relative test error reduction on average | Improved |
| add P2C with extractive preference from annotation records | RoBERTa-base classifier · Offensive, Polite-Wiki, Polite-SE, MNLI, DynaSent-R1, DynaSent-R2 | test accuracy: improves on all six datasets; ~7.59% relative test error reduction vs vanilla on average | Improved |
| P2C with extractive preference vs. disagreement learning baselines (soft-labeling, margin, filtering, weighting, multi-annotator, CS-KD) | RoBERTa-base classifier · six annotation-record datasets | test accuracy: 4.27% relative test error reduction vs previous best disagreement method on average | Improved |
| add subjective (crowd-sourced) preference labels vs. 5k extra task labels vs. generative/extractive preference | RoBERTa-base on DynaSent-R2 · DynaSent-R2 | test accuracy and ECE: subjective 71.74 / ECE 6.09 vs task labels 71.17 / ECE 9.19 | Improved |
| add 5k GPT-3 generative preference labels instead of 5k additional task labels | RoBERTa-base on DynaSent-R2 · DynaSent-R2 | test accuracy: generative 71.46 vs task labels 71.17 | Improved |
| use multiple (T=3) preference heads instead of a single head (ablation) | P2C without div/cons/sampling · DynaSent-R2 | test accuracy: 71.84 (T=1) to 71.92 (T=3) | Improved |
| add diversity regularization L_div between preference heads (ablation) | P2C · DynaSent-R2 | test accuracy: 71.92 to 72.05 | Improved |
| add diversity regularization L_div between preference heads (ablation) | P2C · Offensive | test accuracy: 76.43 to 76.67 | Improved |
| add consistency regularization L_cons between classification and preference heads (ablation) | P2C · DynaSent-R2 | test accuracy: 72.05 to 72.67 | Improved |
| add consistency regularization L_cons between classification and preference heads (ablation) | P2C · Offensive | test accuracy: 76.67 to 77.67 | Improved |
| add disagreement/inconsistency-based informative pair sampling (ablation) | P2C · DynaSent-R2 and Offensive | test accuracy: DynaSent-R2 72.67 to 73.06; Offensive 77.67 to 77.81 | Improved |
| add P2C with extractive preference to ResNet-18 scene attribute classification | ResNet-18 on SUN Attribute · five binary attribute tasks | test accuracy: 76.48 to 78.05 average (~6.66% relative test error reduction) | Improved |
Closest papers by meaning
- Spread Preference Annotation: Direct Preference Judgment for Efficient LLM Alignment (ICLR 2025 · similarity 0.85)
- InfiFPO: Implicit Model Fusion via Preference Optimization in Large Language Models (NeurIPS 2025 · similarity 0.84)
- RePO: Understanding Preference Learning Through ReLU-Based Optimization (NeurIPS 2025 · similarity 0.84)
- Discriminative Finetuning of Generative Large Language Models without Reward Models and Human Preference Data (ICML 2025 · similarity 0.84)
- PREFER: Prompt Ensemble Learning via Feedback-Reflect-Refine (AAAI 2024 · similarity 0.83)
- Compositional Preference Models for Aligning LMs (ICLR 2024 · similarity 0.83)
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.
