AI Paper Atlas

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

ICML 2026 · Large language models › Language, safety and interpretability · approach: Multilingual LLM data and culture · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
Add DCO (crosslingual logit matching) on parallel prompts instead of SFT on reference completionsQwen2.5-14B · MMMLU joint 12 languagesCLC_all change: +0.6 (SFT) vs +10.6 (DCO)Improved
Use DCO as a post-step after DPOQwen2.5-14B · MMMLU joint 12 languagesCLC_all change: +12.3 (DPO) -> +13.1 (DPO+DCO)Improved
Use CALM majority-vote pseudo-labels instead of DCOLlama3.1-8B · MMMLU joint 12 languagesnon-English accuracy change: -5.0 (CALM) vs +7.6 (DCO)Hurt
Use DCO without reference completions rather than DPO with themAya-Expanse-8B · MMMLU joint 12 languagesCLC_all change: +1.3 (DPO) vs +5.3 (DCO)Improved
Apply bilingual DCOfive LMs · BMLAMACLC_all: +12.29 to +16.65Improved
Apply bilingual DCOQwen2.5-14B · XCSQA English accuracyaccuracy: -2.5Hurt
Train DCO on a single subject (HS microeconomics) and test on other subjectsQwen2.5-14B · MMMLU out-of-domain subjectsCLC_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-SwahiliEN and SW accuracy: SW +4.24, EN +3.25 vs default EN -4.80, SW +2.29Improved
Anchor on Swahili (beta1=10, beta2=0.1) (ablation)DCO EN-SW · MMMLU English-SwahiliEN accuracy: -11.74 EN, +0.18 SWHurt
Vary strength parameters (1:1, 10:0.1, 0.1:10) (ablation)DCO EN-SW · MMMLU English-SwahiliCLC_all: 49.37 -> 59.21, 59.50, 56.59Improved
Use L1 norm instead of L2 norm in the DCO loss (ablation)DCO · preliminary experimentsperformanceImproved
Align non-English pairs with DCO (Korean-French)Qwen2.5-14B · BMLAMACLC: 37.29 -> 52.24Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy