Research map / Reinforcement learning / RL from feedback
Learning to Route Languages for Multilingual Policy Optimization
Geyang Guo, Hiromi Wakaki, Yuki Mitsufuji, Alan Ritter, Wei Xu
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 |
|---|---|---|---|
| replace GRPO with LRPO (routed multilingual rollouts) | Qwen2.5-1.5B-it · mGSM-v2 average | score: 32.33 -> 38.25 | Improved |
| replace GRPO with LRPO | Qwen2.5-1.5B-it · Overall average (5 benchmarks) | average score: 30.42 -> 32.15 | Improved |
| replace GRPO with LRPO | Llama3.2-1B-it · mGSM-v2 average | score: 21.24 -> 20.55 | Hurt |
| replace GRPO with LRPO | Gemma3-4B-it · CARE-pro average | score: 18.92 -> 18.45 | Hurt |
| use multilingual rollouts instead of monolingual (ablation) | LRPO fixed router · Overall average (Qwen2.5-1.5B-it) | average score: 30.42 -> 31.89 | Improved |
| sample languages uniformly instead of EN-dominant mix (ablation) | LRPO fixed router · Overall average (Qwen2.5-1.5B-it) | average score: 31.89 -> 31.36 | Hurt |
| use dynamic learned router instead of fixed router (ablation) | LRPO · Overall average (Qwen2.5-1.5B-it) | average score: 31.89 -> 32.15 | Improved |
| add uniform initialization of the router (ablation) | LRPO router · Overall average (Qwen2.5-1.5B-it) | average score: 31.48 -> 32.15 | Improved |
| add temperature annealing to the router (ablation) | LRPO router · Overall average (Qwen2.5-1.5B-it) | average score: 31.67 -> 32.15 | Mixed |
| add warm-start SFT for language control (ablation) | LRPO · Overall average (Qwen2.5-1.5B-it) | average score: 30.44 -> 32.15 | Improved |
| LRPO without warm-start vs GRPO (ablation) | LRPO-Zero · Overall average (Qwen2.5-1.5B-it) | average score: 30.42 -> 30.44 | No clear effect |
| add LRPO on multiple-choice tasks | LRPO · Global-MMLU-Lite / Include-Lite | accuracy | No clear effect |
Closest papers by meaning
- Q-SFT: Q-Learning for Language Models via Supervised Fine-Tuning (ICLR 2025 · similarity 0.88)
- Meta-Learning Preferences for Multilingual LLM Alignment (NeurIPS 2026 · similarity 0.88)
- DAPO : Improving Multi-Step Reasoning Abilities of Large Language Models with Direct Advantage-Based Policy Optimization (NeurIPS 2025 · similarity 0.88)
- Enhancing Decision-Making of Large Language Models via Actor-Critic (ICML 2025 · similarity 0.88)
- Prompted Policy Search: Reinforcement Learning through Linguistic and Numerical Reasoning in LLMs (NeurIPS 2025 · similarity 0.88)
- Reinforcement Mid-Training (ICLR 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.
