Research map / Large language models / Alignment and preferences
MENLO: From Preferences to Proficiency – Evaluating and Modeling Native-like Quality Across 47 Languages
Chenxi Whitehouse, Sebastian Ruder, Tony Zhiyang Lin, Oksana Kurylo, Haruka Takagi, Janice Lam, Nicolò Busetto, Denise Diaz
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 |
|---|---|---|---|
| grade two responses side by side (pairwise) instead of one at a time (ablation) | zero-shot LLM judges · MENLO test | Preference accuracy: Qwen3-4B 40.54 -> 57.13 | Improved |
| add three graded few-shot examples to pointwise judging (ablation) | LLM judges · MENLO test | Preference accuracy: Qwen3-4B 40.54 -> 39.35 | No clear effect |
| add detailed 5-point grading rubrics (ablation) | pointwise LLM judges · MENLO test | Macro-F1: Qwen3-4B 16.00 -> 23.06 | Improved |
| add rubrics to pairwise judging (ablation) | pairwise LLM judges · MENLO test | Macro-F1: Qwen3-32B 38.10 -> 37.48 | No clear effect |
| train judge with SFT (no CoT) instead of zero-shot (ablation) | Qwen3-4B · MENLO test | Macro-F1: 35.46 -> 33.44 | Hurt |
| train judge with SFT instead of zero-shot (ablation) | Llama4-Scout · MENLO test | Macro-F1: 36.11 -> 44.17 | Improved |
| train judge with GRPO RL instead of zero-shot (ablation) | Qwen3-4B · MENLO test | Macro-F1: 35.46 -> 39.44 | Improved |
| initialize RL from the SFT checkpoint (ablation) | Qwen3-4B / Llama4-Scout · MENLO test | Macro-F1: 39.44 -> 39.33 (Qwen3-4B) | No clear effect |
| add reward smoothing alone to binary reward (ablation) | Qwen3-4B RL judge · MENLO test | Preference accuracy: 58.27 -> 51.47 | Hurt |
| add smoothing and preference bonus to binary reward (ablation) | Qwen3-4B RL judge · MENLO test | Macro-F1: 37.11 -> 39.44 | Improved |
| train RL on English data only (ablation) | Qwen3-4B RL judge · 47 language varieties | performance | Hurt |
| use RL on Localized Factuality dimension (ablation) | Qwen3-4B · Localized Factuality | Macro-F1: +0.6 single-task RL; regressions multitask | No clear effect |
Closest papers by meaning
- Peering Through Preferences: Unraveling Feedback Acquisition for Aligning Large Language Models (ICLR 2024 · similarity 0.90)
- Meta-Learning Preferences for Multilingual LLM Alignment (NeurIPS 2026 · similarity 0.89)
- Reward-Augmented Data Enhances Direct Preference Alignment of LLMs (ICML 2025 · similarity 0.88)
- On Evaluating LLM Alignment by Evaluating LLMs as Judges (NeurIPS 2025 · similarity 0.88)
- Self-Boosting Large Language Models with Synthetic Preference Data (ICLR 2025 · similarity 0.88)
- CLAVE: An Adaptive Framework for Evaluating Values of LLM Generated Responses (NeurIPS 2024 · 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.
