Research map / Large language models / Alignment and preferences
mR3: Multilingual Rubric-Agnostic Reward Reasoning Models
David Anugraha, Shou-Yi Hung, Zilu Tang, En-Shiun Annie Lee, Derry Tanti Wijaya, Genta Indra Winata
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 |
|---|---|---|---|
| multilingual SFT on 100k rubric-based reasoning data | Qwen3-14B · m-RewardBench | accuracy: 88.46 -> 89.18 | Improved |
| multilingual SFT | Qwen3-4B · MM-Eval | accuracy: 80.07 -> 82.62 | Improved |
| multilingual SFT | Qwen3-14B · RTP-LX | F1: 78.36 -> 90.19 | Improved |
| multilingual SFT versus English-only R3 training | Qwen3-14B · MM-Eval | accuracy: 84.04 -> 86.05 | Improved |
| multilingual SFT versus English-only R3 training | Qwen3-14B · RewardBench (English) | accuracy: 91.00 -> 90.79 | Hurt |
| scale model from 4B to 14B | mR3 · m-RewardBench | accuracy: 87.61 -> 89.18 | Improved |
| use English reasoning instead of target-language reasoning (ablation) | mR3 SFT · pairwise benchmarks | average accuracy | Improved |
| train with target-language reasoning (ablation) | mR3 SFT · pairwise benchmarks | accuracy | Improved |
| easy-to-hard curriculum vs random and other orderings (ablation) | mR3 SFT · HelpSteer3 validation | accuracy | Improved |
| DPO with mR3-Qwen3-14B reward | Qwen3-30B-A3B-Instruct-2507 · m-ArenaHard-v2.0 (23 langs) | win rate: 39.1 -> 45.2 | Improved |
| DPO with mR3-Qwen3-14B reward | Qwen3-30B-A3B-Instruct-2507 · MCLM MT-MATH100 | accuracy: 90.47 -> 92.08 | Improved |
| DPO with mR3 instead of Nemotron-Multilingual-49B reward | Qwen3-30B-A3B-Instruct-2507 · m-ArenaHard-v2.0 (23 langs) | win rate: 47.0 -> 45.2 | Hurt |
Closest papers by meaning
- RM-R1: Reward Modeling as Reasoning (ICLR 2026 · similarity 0.89)
- RM-Bench: Benchmarking Reward Models of Language Models with Subtlety and Style (ICLR 2025 · similarity 0.88)
- Language Imbalance Driven Rewarding for Multilingual Self-improving (ICLR 2025 · similarity 0.88)
- Beyond English-Centric Training: How Reinforcement Learning Improves Cross-Lingual Reasoning in LLMs (ICLR 2026 · similarity 0.88)
- BaseReward: A Strong Baseline for Multimodal Reward Model (ICLR 2026 · similarity 0.88)
- Step-wise Rubric Rewards for LLM Reasoning (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.
