Research map / Reinforcement learning / RL from feedback
MRACL: Multi-Reward Space Guided Adaptive Curriculum Reinforcement Learning for LLMs
Wenxuan Liu, Liangyu Huo, Yi Jing, Xiyuan Zhang, Jian Xie
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 MRACL multi-reward curriculum instead of DUMP | Reinforce++ on Qwen2.5-7B-Base · UltraFeedback in-distribution | convergence speed: 1.62x faster than DUMP | Improved |
| use MRACL curriculum instead of no curriculum | Reinforce++ on Qwen2.5-7B-Base · UltraFeedback in-distribution | convergence speed: 2.55x faster (abstract) vs 2.54x (results text) | Improved |
| use MRACL instead of DUMP | Qwen2.5-7B-Base · win rate (step 100 avg) | win rate (%): 60.0 -> 63.1 | Improved |
| use MRACL instead of the best baseline on rule-based benchmarks | Qwen2.5-7B-Base · TriviaQA | score: 51.71 -> 50.31 (SpeedRL vs MRACL) | Hurt |
| add capability distance term (ablation) | MRACL · UltraFeedback hard subset | sum reward | Improved |
| add optimization direction term (ablation) | MRACL · UltraFeedback hard subset | sum reward: 25.0 -> 26.6 | Improved |
| add historical evolution term (ablation) | MRACL · UltraFeedback easy/median/hard | sum reward | Improved |
| add exploration (UCB bias) (ablation) | MRACL · UltraFeedback | sum reward (later training) | Improved |
| use fixed alpha=beta=0.5 instead of time-varying schedule (ablation) | MRACL · UltraFeedback | final sum reward | Hurt |
| reverse the alpha/beta schedule (ablation) | MRACL · UltraFeedback | sum reward | Hurt |
| reduce rollouts from 4 to 1 (ablation) | DUMP · UltraFeedback sum reward | final sum reward: 33.8 -> 32.1 | Hurt |
| apply MRACL to a stronger instruct model | Qwen2.5-7B-Instruct · UltraFeedback easy/median vs hard | sum reward: hard subset 25.1 -> 30 | Mixed |
Closest papers by meaning
- AdaCuRL: Adaptive Curriculum Reinforcement Learning with Invalid Sample Mitigation and Historical Revisiting (AAAI 2026 · similarity 0.91)
- Reshaping Reasoning in LLMs: A Theoretical Analysis of RL Training Dynamics through Pattern Selection (ICLR 2026 · similarity 0.90)
- CurES: From Gradient Analysis to Efficient Curriculum Learning for Reasoning LLMs (ICLR 2026 · similarity 0.89)
- Reward Is Enough: LLMs Are In-Context Reinforcement Learners (ICLR 2026 · similarity 0.89)
- On Predictability of Reinforcement Learning Dynamics for Large Language Models (ICLR 2026 · similarity 0.88)
- Revisiting Reinforcement Learning for LLM Reasoning from A Cross-Domain Perspective (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.
