Research map / Large language models / Alignment and preferences
GEM: Generative Entropy-Guided Preference Modeling for Few-Shot Alignment of LLMs
Yiyang Zhao, Huiyu Bai, Xuejiao Zhao
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 Cognitive Filtering and SEGA (ablation) | GEM on Llama-3-8B-Instruct · UltraFeedback | preference-prediction accuracy (%): 69.0 -> 77.1 | Improved |
| add Cognitive Filtering and SEGA (ablation) | GEM on Llama-3-8B-Instruct · GSM8K | accuracy (%): 48.3 -> 55.6 | Improved |
| add final-answer entropy term to the token score (ablation) | GEM · GSM8K | accuracy (%): 50.1 -> 55.6 | Improved |
| add fork-token entropy term to the token score (ablation) | GEM · GSM8K | accuracy (%): 52.7 -> 55.6 | Improved |
| use SEGA instead of pairwise DPO on filtered CoTs (ablation) | GEM · Medical QA expert agreement | agreement (%): 73.0 -> 78.2 | Improved |
| use SEGA instead of pairwise DPO on filtered CoTs (ablation) | GEM · UltraFeedback | preference-prediction accuracy (%): 74.5 -> 77.1 | Improved |
| use GEM instead of DPO on few-shot preference pairs | Llama-3-8B-Instruct · GSM8K | accuracy (%): 50.2 -> 55.6 | Improved |
| use GEM instead of DPO on few-shot preference pairs | Llama-3-8B-Instruct · RewardBench | preference-prediction accuracy (%): 63.2 -> 75.4 | Improved |
| use GEM instead of Reward Model + PPO | Llama-3-8B-Instruct · MT-Bench (win-rate vs SFT baseline) | win-rate (%): 47 -> 68 | Improved |
| use GEM instead of Reward Model + PPO on medical QA | Llama-3-8B-Instruct · Medical QA expert agreement | agreement (%): 72.5 -> 78.2 | Improved |
| add Cognitive Filtering and SEGA (ablation) | GEM on Llama-3-8B-Instruct · Medical QA expert agreement | agreement (%): 70.5 -> 78.2 | Improved |
Closest papers by meaning
- Reward-Augmented Data Enhances Direct Preference Alignment of LLMs (ICML 2025 · similarity 0.91)
- Leveraging robust optimization for llm alignment under distribution shifts (NeurIPS 2025 · similarity 0.91)
- Sequential Preference Optimization: Multi-Dimensional Preference Alignment with Implicit Reward Modeling (AAAI 2025 · similarity 0.91)
- Panacea: Pareto Alignment via Preference Adaptation for LLMs (NeurIPS 2024 · similarity 0.91)
- Group Preference Optimization: Few-Shot Alignment of Large Language Models (ICLR 2024 · similarity 0.91)
- Stackelberg Self-Annotation: A Robust Approach to Data-Efficient LLM Alignment (NeurIPS 2025 · similarity 0.91)
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.
