Xian Li
As of October 2026, Xian Li is a Research Scientist at Meta whose work spans machine translation, language-model training and reasoning agents. Her papers from 2019 to 2021 studied machine-translation robustness and zero-shot translation. From 2023 to 2024, she coauthored work on instruction data, self-rewarding language models and evaluators trained with synthetic examples. In 2025 and 2026, she coauthored research on agents that create their own tasks and models that help select AI research experiments.
Timeline
-
2019–2020
Machine translation robustness
In 2019, she coauthors the first shared-task report on machine-translation robustness, which evaluates English-French and English-Japanese systems on noisy Reddit comments. A second report in 2020 provides English-German and English-Japanese test sets for zero-shot and few-shot translation.
-
2021
Zero-shot translation by position
Her 2021 paper, Improving Zero-Shot Translation by Disentangling Positional Information, removes residual connections in an encoder layer and gains up to 18.5 BLEU points on zero-shot translation.
-
2023
Instruction backtranslation
In 2023, she coauthors Self-Alignment with Instruction Backtranslation, which uses web documents to generate candidate instruction examples and selects high-quality examples for model training.
-
2024
Self-rewarding language models
Self-Rewarding Language Models appears in 2024 with her as a coauthor; the method lets a language model provide its own rewards during training. Iterative DPO improves instruction following and the model's ability to judge those rewards.
-
2024
Training language model evaluators
In 2024, she coauthors Self-Taught Evaluators, which trains an LLM-as-a-Judge from synthetic comparisons and reasoning traces.
-
2025
Self-challenging language agents
She coauthors Self-Challenging Language Model Agents in 2025; the framework has agents create tool-based Code-as-Task challenges and train with reinforcement learning on evaluation feedback.
-
2026
AI research preference models
In 2026, she coauthors AI Research Preference Models, which helps AI research agents choose promising candidate solutions before full runs. One variant reasons over plans and code; another runs small pilot experiments.
-
As of October 2026
Research Scientist at Meta
As of October 2026, she is a Research Scientist at Meta and works on language models.
Sources and credits
Photo: We have no published photo of Xian Li yet. Send us a free-licensed photo.
- Findings of the First Shared Task on Machine Translation Robustness
- Findings of the WMT 2020 Shared Task on Machine Translation Robustness
- Improving Zero-Shot Translation by Disentangling Positional Information
- Self-Alignment with Instruction Backtranslation
- Self-Rewarding Language Models
- Self-Taught Evaluators
- Self-Challenging Language Model Agents
- AI Research Preference Models
- Xian Li | 500AI