AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

How to Fine-tune the Model: Unified Model Shift and Model Bias Policy Optimization

Hai Zhang, Hang Yu, Junqiao Zhao, Di Zhang, Chang Huang, Hongtu Zhou, Xiao Zhang, Chen Ye

NeurIPS 2023 · Reinforcement learning › RL from feedback · approach: RL fine-tuning of generative models · 3 citations (OpenAlex)

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 changedWhereMeasuredEffect
Add phase-2 fine-tuning that jointly minimizes model shift and model biasMBPO · MuJoCo continuous control tasksmaximum average returnImproved
Remove phase-2 fine-tuning (ablation)USB-PO (MBPO variant) · MuJoCo continuous control taskssample efficiencyHurt
Fine-tune using only the model-shift term (ablation)USB-PO · MuJoCo continuous control taskssample efficiencyHurt
Fine-tune using only the model-bias term (ablation)USB-PO · MuJoCo continuous control tasksaverage returnHurt
Use adaptive unified fine-tuning rather than limiting update magnitude through MBPO learning-rate changes (ablation)MBPO versus USB-PO · Walker2d and HalfCheetahaverage returnImproved
Vary the phase-2 learning rate (ablation)USB-PO · Walker2d and HalfCheetahaverage returnNo clear effect
Vary the CMLO model-shift threshold from the recommended setting (ablation)CMLO · Walker2daverage returnHurt
Increase the ensemble size (ablation)USB-PO · MuJoCo continuous control tasksperformanceImproved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy