🌐 Multilingual & Translation

🧪 ICML2026 · 3 paper notes

📌 Same area in other venues: 🔬 ICLR2026 (8) · 💬 ACL2026 (64) · 🤖 AAAI2026 (9) · 🧠 NeurIPS2025 (11) · 📹 ICCV2025 (1) · 🧪 ICML2025 (1)

Edit-Based Refinement for Parallel Masked Diffusion Language Models

ME-DLM introduces a lightweight "decode-then-edit" refinement stage to masked diffusion language models (e.g., LLaDA). The first stage generates a draft via standard parallel unmasking, while the second stage performs parallel corrections using replace/delete/insert actions supervised by the shortest edit distance scripts. Using only 1/8 of the diffusion step budget, it outperforms LLaDA-Instruct by +11.6 on HumanEval and +33.6 on GSM8K.

Optimizing Language Models for Crosslingual Knowledge Consistency

This paper addresses the issue of multilingual LLMs providing conflicting answers to the same question across different languages. It designs an RL objective using the "log-likelihood of the answer in another language" as a reward, proving that the optimal policy follows a product-of-experts form and guarantees crosslingual preference consistency when \(\gamma_1\gamma_2=\beta^2\). Based on this, the authors derive DCO (Direct Consistency Optimization), a reward-model-free and online-sampling-free algorithm. Experiments across 9 LLMs, 3 multilingual QA benchmarks, and 26 languages demonstrate simultaneous improvements in crosslingual consistency (RankC) and response accuracy.

Toward Robust Multilingual Adaptation of LLMs for Low-Resource Languages

LiRA inserts a lightweight fine-tuning module featuring "anchoring + consistency regularization" between a frozen multilingual encoder and an English LLM. It constrains the sentence vectors of low-resource languages into a shared English semantic space through two theoretically controllable quantities: \(\epsilon_1\) (anchoring error) and \(\epsilon_2\) (translation KL distance), achieving stable improvements across retrieval, ranking, and reasoning tasks.