Research map / Reinforcement learning / RL from feedback
RLAnything: Forge Environment, Policy, and Reward Model in Completely Dynamic RL System
Yinjie Wang, Tianbao Xie, Ke Shen, Mengdi Wang, Ling Yang
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 dynamic reward model optimization (consistency feedback) on top of policy training (ablation) | RLAnything GUI agent · OSWorld in-domain | accuracy: 35.3 -> 36.9 | Improved |
| add environment adaptation on top of policy + reward (ablation) | RLAnything GUI agent · OSWorld in-domain | accuracy: 36.9 -> 40.6 | Improved |
| add environment adaptation on top of policy + reward (ablation) | RLAnything LLM agent · AlfWorld in-domain | accuracy: 55.6 -> 60.2 | Improved |
| add environment adaptation to policy + reward (ablation) | reward model · AlfWorld process accuracy | process accuracy: 55.2 -> 59.4 | Improved |
| add environment adaptation to policy + reward (ablation) | coding unit-test generator · LiveBench Detect | detect accuracy: 37.9 -> 48.5 | Improved |
| add joint reward model training to policy (ablation) | coding LLM · LiveBench UT accuracy | UT accuracy: 27.8 -> 73.3 | Improved |
| add environment adaptation (ablation) | reward model outcome accuracy in GUI/LLM agents · AlfWorld outcome accuracy | outcome accuracy: 61.1 -> 61.7 | Improved |
| use integrated step-wise plus outcome reward instead of outcome-only (GRPO) (ablation) | policy training · AlfWorld and OSWorld training curves | training reward curve | Improved |
| train policy with only optimized reward model signal, no evaluator-script outcome (ablation) | GUI policy · OSWorld training curve | training curve | Improved |
| vary lambda from 1 to 1/4 in policy reward (ablation) | RLAnything policy · AlfWorld | policy accuracy: 54.1 -> 48.0 | Hurt |
| increase lambda from 1 to 4 (ablation) | RLAnything policy · AlfWorld | policy accuracy: 54.1 -> 53.3 | Hurt |
| add self-consistency signal (lambda=1 vs 0) to reward model training (ablation) | reward model · AlfWorld | step accuracy: 48.2 -> 54.7 | Improved |
Closest papers by meaning
- A Clean Slate for Offline Reinforcement Learning (NeurIPS 2025 · similarity 0.87)
- Uni-RLHF: Universal Platform and Benchmark Suite for Reinforcement Learning with Diverse Human Feedback (ICLR 2024 · similarity 0.87)
- Prompted Policy Search: Reinforcement Learning through Linguistic and Numerical Reasoning in LLMs (NeurIPS 2025 · similarity 0.87)
- PoEM: Predicting New RL Outcomes from Existing Policies (NeurIPS 2026 · similarity 0.87)
- Asymmetric REINFORCE for off-Policy Reinforcement Learning: Balancing positive and negative rewards (NeurIPS 2025 · similarity 0.87)
- Policy Filtration for RLHF to Mitigate Noise in Reward Models (ICML 2025 · similarity 0.87)
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.
