Research map / Reinforcement learning / RL from feedback
General Exploratory Bonus for Optimistic Exploration in RLHF
Wendi Li, Changdae Oh, Sharon Li
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 GEB-pi exploratory bonus to f-DPO | iterative DPO, Mistral-Instruct-v0.3 · UltraFeedback KL | win rate: 78.42 -> 81.00 | Improved |
| add GEB-arctanh bonus to f-DPO | iterative DPO, Mistral-Instruct-v0.3 · UltraFeedback Hellinger | win rate: 72.69 -> 75.69 | Improved |
| add GEB-arctanh bonus to f-DPO | iterative DPO, LLaMA-3-8B-SFT · UltraFeedback Hellinger | win rate: 71.11 -> 73.26 | Improved |
| use FEB (prior-style bonus) instead of f-DPO | iterative DPO, Mistral-Instruct-v0.3 · UltraFeedback forward KL | win rate: 51.11 -> 47.53 | Hurt |
| add GEB bonus (average of divergences) | f-DPO, LLaMA-3-8B-SFT · AlpacaEval2 LC | win rate: 22.75 -> 24.73 | Improved |
| add GEB bonus (average of divergences) | f-DPO, LLaMA-3-8B-SFT · MATH-500 | accuracy: 68.6 -> 70.1 | Improved |
| add GEB-1/pi bonus to DPO on hard prompts | DPO, Qwen2.5-7B · OlympiadBench | Pass@16: 57.78 -> 65.78 | Improved |
| add GEB-1/pi bonus to DPO on hard prompts | DPO, Qwen2.5-7B · AIME 2025 | Pass@16: 23.23 -> 29.48 | Improved |
| apply GEB-pi bonus to rejected responses only instead of all responses (ablation) | GEB-pi, Mistral-Instruct-v0.3 · UltraFeedback KL | win rate: 79.78 -> 81.00 | Improved |
| apply GEB-pi bonus to rejected responses only instead of all responses (ablation) | GEB-pi, Mistral-Instruct-v0.3 · UltraFeedback Hellinger | win rate: 73.40 -> 75.48 | Improved |
| add GEB bonus for distinct-n diversity (dist-4) | f-DPO · sampled corpus under KL | dist-4: 0.8418 -> 0.8448 | No clear effect |
| make bonus ratio too large (ablation) | GEB · kappa sweep | performance | Hurt |
Closest papers by meaning
- Beyond Optimism: Exploration With Partially Observable Rewards (NeurIPS 2024 · similarity 0.90)
- Value-Incentivized Preference Optimization: A Unified Approach to Online and Offline RLHF (ICLR 2025 · similarity 0.89)
- Minimax Optimal Reinforcement Learning with Quasi-Optimism (ICLR 2025 · similarity 0.88)
- Efficient Model-Based Reinforcement Learning Through Optimistic Thompson Sampling (ICLR 2025 · similarity 0.88)
- Uncertainty-Aware Reward-Free Exploration with General Function Approximation (ICML 2024 · similarity 0.88)
- Exploration from a Primal-Dual Lens: Value-Incentivized Actor-Critic Methods for Sample-Efficient Online RL (NeurIPS 2025 · similarity 0.88)
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.
