What Really Matters? Multilinguality in the Era of Agentic AI

AgentsMultilingual NLPEvaluation & benchmarks

Is multilinguality best built into the agent, or added through a harness? Comparing both paradigms, and hybrids, in noisy agentic environments.

About the project

In the recent era of agentic AI, the relevance of natural language processing (NLP) research is largely questioned, one of which is also reflected in the special theme of Empirical Methods in Natural Language Processing (EMNLP) 2026 on “New Missions for NLP Research”. In the spirit of answering this very question, we conduct an in-depth study on one of the specific field in NLP, i.e., multilingual NLP. In this work, we would like understand the impact of two different paradigms of enabling multilingual capability under complex and noisy agentic AI environments. The first paradigm incorporates multilinguality as one of the agentic AI capabilities itself, allowing the agent to natively understand and produce sentences in any languages. The other paradigm brings multilinguality as a part of an agentic harness/scaffold allowing multilingual support through external tool use or a specialized workflow making multilinguality as an external independent capability from agentic AI.

However it is also important to note that most existing agentic AI models have already supported multilinguality following the prior paradigm, however the capability is still considerably weaker in comparison to major languages like English or Chinese. In accordance to these two paradigms, we would further explore hybrid approaches to enable multilingual capability in agentic AI, meaning allowing a part of multilingual capability through the agent internal capabilities itself while leaving the rest to customized harnesses and scaffolds to find a perfect balance between efficiency and quality trade-off to deliver best multilingual capability in agentic AI. Through this study, we aim to provide a clear picture on how we define new mission for multilingual NLP in the era of agentic AI.

What mentees potentially work on

  • Explore multilingual agentic datasets
  • Design experiments comparing a multilingual LLM against a multilingual harness around an LLM
  • Conduct extensive experiments and behavioral analysis comparing the two approaches

Requirements and recommended background

  • Basic programming
  • Experience building high-quality data for LLMs
  • Experience using coding agents in day-to-day work
  • Prior experience with multi-turn and tool-use evaluation is preferred
  • Prior experience conducting mass surveys and interviews
  • Good communication skills

There are no formal eligibility limits. See who can apply; you can still apply if you don't match every item.

Relevant publications

Mentors

Samuel Cahyawijaya
Samuel Cahyawijaya
Cohere
Member of Technical Staff
Amit Agarwal
Amit Agarwal
Oracle
Senior Principal Applied Scientist