Research map / Reinforcement learning / RL from feedback
Preference Goal Tuning: Post-Training as Latent Control for Frozen Policies
Guangyu Zhao, Kewei Lian, Haoxuan Ru, Borong Zhang, Haowei Lin, Zhancun Mu, Haobo Fu, QIANG FU, Shaofei Cai, Zihao Wang, Yitao Liang
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 |
|---|---|---|---|
| use DPO instead of behavior cloning on positive trajectories to tune the latent goal (ablation) | PGT latent-goal-only · Minecraft task (pretrained 42.0) | success rate (%): 18.2 -> 57.2 | Improved |
| tune only the latent goal instead of full fine-tuning (in distribution) (ablation) | GROOT with DPO · Minecraft SkillForge tasks, ID | success rate / reward: 57.2 vs 62.2; 6.58 vs 6.00; 57.8 vs 58.4 | No clear effect |
| tune only the latent goal instead of full fine-tuning (out of distribution) (ablation) | GROOT with DPO · Minecraft SkillForge tasks, OOD | performance: surpasses full fine-tuning by 13.4% | Improved |
| apply PGT to frozen policy versus pretrained policy with expert prompt | GROOT · Minecraft SkillForge, 17 tasks, ID | relative improvement: +72.0% | Improved |
| apply PGT to frozen policy versus pretrained STEVE-1 | STEVE-1 · Minecraft SkillForge, 17 tasks, OOD | relative improvement: +36.9% | Improved |
| run iterative preference rounds up to three (ablation) | PGT · Minecraft tasks ID and OOD | performance | Improved |
| start PGT from any of five initial video prompts including suboptimal ones (ablation) | PGT on collect wood · Minecraft collect wood | cumulative reward | Improved |
| use PGT latent goals rather than multi-task learning (out of distribution) | GROOT · Minecraft 6-task multi-task | aggregate performance | Mixed |
| add PGT latent goals to GROOT with JARVIS-1 planner | GROOT · long-horizon crafting | success rate (%): 37.8 -> 64.9 | Improved |
| tune latent goal instead of LoRA, BitFit or VeRA under same preference data (ablation) | GROOT · Minecraft ID and OOD | performance | Improved |
| apply PGT to Minecraft tasks where success rate is saturated or zero baseline | GROOT/STEVE-1 · Minecraft SkillForge, per-task | success rate: e.g. -65.5%, -80.0%, -0.5% on some tasks | Mixed |
Closest papers by meaning
- Pre-Trained Multi-Goal Transformers with Prompt Optimization for Efficient Online Adaptation (NeurIPS 2024 · similarity 0.88)
- BOTS: A Unified Framework for Bayesian Online Task Selection in LLM Reinforcement Finetuning (ICLR 2026 · similarity 0.88)
- Reflective Prompted Policy Optimization: Trajectory-Grounded Revision and Salience Bias (NeurIPS 2026 · similarity 0.87)
- Hindsight PRIORs for Reward Learning from Human Preferences (ICLR 2024 · similarity 0.87)
- Implicit Reward as the Bridge: A Unified View of SFT and DPO Connections (NeurIPS 2025 · similarity 0.87)
- AlphaDPO: Adaptive Reward Margin for Direct Preference Optimization (ICML 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.
