Research map / Large language models / Language, safety and interpretability
Post-Training Language Models for Crosslingual Consistency
Tianyu Liu, Jirui Qi, Mrinmaya Sachan, Ryan Cotterell, Raquel Fernández, Arianna Bisazza
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 DCO (crosslingual logit matching) on parallel prompts instead of SFT on reference completions | Qwen2.5-14B · MMMLU joint 12 languages | CLC_all change: +0.6 (SFT) vs +10.6 (DCO) | Improved |
| Use DCO as a post-step after DPO | Qwen2.5-14B · MMMLU joint 12 languages | CLC_all change: +12.3 (DPO) -> +13.1 (DPO+DCO) | Improved |
| Use CALM majority-vote pseudo-labels instead of DCO | Llama3.1-8B · MMMLU joint 12 languages | non-English accuracy change: -5.0 (CALM) vs +7.6 (DCO) | Hurt |
| Use DCO without reference completions rather than DPO with them | Aya-Expanse-8B · MMMLU joint 12 languages | CLC_all change: +1.3 (DPO) vs +5.3 (DCO) | Improved |
| Apply bilingual DCO | five LMs · BMLAMA | CLC_all: +12.29 to +16.65 | Improved |
| Apply bilingual DCO | Qwen2.5-14B · XCSQA English accuracy | accuracy: -2.5 | Hurt |
| Train DCO on a single subject (HS microeconomics) and test on other subjects | Qwen2.5-14B · MMMLU out-of-domain subjects | CLC_all: +10.94 (anatomy), +10.33 (medical genetics) | Improved |
| Anchor on English (beta1=0.1, beta2=10) instead of default strengths (ablation) | DCO EN-SW · MMMLU English-Swahili | EN and SW accuracy: SW +4.24, EN +3.25 vs default EN -4.80, SW +2.29 | Improved |
| Anchor on Swahili (beta1=10, beta2=0.1) (ablation) | DCO EN-SW · MMMLU English-Swahili | EN accuracy: -11.74 EN, +0.18 SW | Hurt |
| Vary strength parameters (1:1, 10:0.1, 0.1:10) (ablation) | DCO EN-SW · MMMLU English-Swahili | CLC_all: 49.37 -> 59.21, 59.50, 56.59 | Improved |
| Use L1 norm instead of L2 norm in the DCO loss (ablation) | DCO · preliminary experiments | performance | Improved |
| Align non-English pairs with DCO (Korean-French) | Qwen2.5-14B · BMLAMA | CLC: 37.29 -> 52.24 | Improved |
Closest papers by meaning
- Meta-Learning Preferences for Multilingual LLM Alignment (NeurIPS 2026 · similarity 0.88)
- Calibrating Large Language Models with Sample Consistency (AAAI 2025 · similarity 0.87)
- Breaking Physical and Linguistic Borders: Multilingual Federated Prompt Tuning for Low-Resource Languages (ICLR 2024 · similarity 0.87)
- Model Fusion through Bayesian Optimization in Language Model Fine-Tuning (NeurIPS 2024 · similarity 0.87)
- SimPER: A Minimalist Approach to Preference Alignment without Hyperparameters (ICLR 2025 · similarity 0.87)
- Alignment of Large Language Models with Constrained Learning (NeurIPS 2025 · similarity 0.86)
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.
