Research map / Reinforcement learning / RL from feedback
Noise-corrected GRPO: From Noisy Rewards to Unbiased Gradients
Omar El mansouri, Fathinah Asma Izzati, Mohamed El Amine Seddik, Salem Lahlou
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 noise correction to Dr.GRPO (synthetic noise 0.2,0.3) | Dr.GRPO, Llama-3.2-1B · GSM8K | clean accuracy (%): 52.99 -> 68.51 | Improved |
| add noise correction to Dr.GRPO (synthetic noise 0.4,0.5) | Dr.GRPO, Qwen3-0.6B · GSM8K | clean accuracy (%): 48.38 -> 56.34 | Improved |
| add noise correction under light noise (0.1,0.2) | Dr.GRPO, Llama-3.2-1B · GSM8K | clean accuracy (%): 65.69 noiseless -> 71.04 | Improved |
| add noise correction across model scales | Dr.GRPO, Qwen3-4B · MATH (synthetic noise 0.3,0.4) | clean accuracy (%): 24.87 -> 31.58 | Improved |
| add noise correction | Dr.GRPO, Llama-3.2-3B · MATH (synthetic noise 0.1,0.2) | clean accuracy (%): 51.01 -> 51.70 | Improved |
| apply mean (Natarajan) correction only, without variance adjustment (ablation) | GRPO, Qwen3-0.6B · GSM8K synthetic noise (0.2,0.1) | clean accuracy (%): 75.73 -> 75.44 | No clear effect |
| apply mean correction only without variance adjustment (real reward model) (ablation) | GRPO, Qwen3-0.6B · GSM8K (Llama3.1-8B-ORM-Mistral-Data) | accuracy (%): 76.47 -> 72.82 | Hurt |
| add variance adjustment M=sqrt(Z) on top of mean correction (ablation) | GRPO, Qwen3-0.6B · GSM8K (Llama3.1-8B-ORM-Mistral-Data) | accuracy (%): 72.82 -> 79.12 | Improved |
| add noise correction with AceMath-7B-RM reward model | Dr.GRPO, Llama-3.2-1B · GSM8K | accuracy (%): 52.52 -> 59.21 | Improved |
| add noise correction with RM-Mistral-7B reward model | Dr.GRPO, Qwen3-0.6B · APPS | accuracy (%): 9.01 -> 10.50 | Improved |
| add noise correction with RM-Mistral-7B reward model | Dr.GRPO, Llama-3.2-1B · APPS | accuracy (%): 8.55 -> 8.68 | No clear effect |
| add gradient clipping instead of noise correction (ablation) | GRPO-style training · noisy rewards | accuracy | No clear effect |
Closest papers by meaning
- Learning to Explore with Parameter-Space Noise: A Deep Dive into Parameter-Space Noise for Reinforcement Learning with Verifiable Rewards (NeurIPS 2026 · similarity 0.90)
- TIC-GRPO: Provable and Efficient Optimization for Reinforcement Learning from Human Feedback (NeurIPS 2026 · similarity 0.90)
- On the Effect of Negative Gradient in Group Relative Deep Reinforcement Optimization (NeurIPS 2025 · similarity 0.88)
- Policy Filtration for RLHF to Mitigate Noise in Reward Models (ICML 2025 · similarity 0.88)
- DACA-GRPO: Denoising-Aware Credit Assignment for Reinforcement Learning in Diffusion Language Models (NeurIPS 2026 · similarity 0.88)
- $f$-GRPO & Beyond: Divergence-Based Reinforcement Learning Algorithms for General LLM Alignment (NeurIPS 2026 · 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.
