We are pleased to announce the new edition of our WMT 2026 Shared Task on Multitask LLMs with Limited Resources, co-located with EMNLP 2026!
Key links:
- Shared Task Website: https://www2.statmt.org/wmt26/limited-resources-llm.html
- Google Group of LLMs with Limited Resources: https://groups.google.com/g/llms-with-limited-resources-2026
- GitHub repository: https://github.com/TUM-NLP/llms-limited-resources2026
This shared task extends the WMT 2025 edition on low-resource LLMs by introducing three additional multitask settings, on top of Machine Translation and Question Answering:
- Spell Checking
- Grammar Checking
- modern evaluation of Maths Reasoning.
The benchmark focuses on low-resource evaluation for three languages:
- Ukrainian (uk)
- Upper Sorbian (hsb)
- Lower Sorbian (dsb)
Participants are required to train a single multitask model based on the Qwen3.5-2B model and jointly evaluate it on all five tasks:
- Machine Translation
- Multiple-Choice Question Answering
- Spell Checking
- Grammar Checking
- Maths Reasoning
The task aims to study:
- Cross-task transfer in compact LLMs
- Multitask learning under strict parameter constraints
- Efficient adaptation for low-resource Slavic languages
- Interactions between translation, linguistic correction, QA, and reasoning capabilities
So, if you would like to challenge yourself in developing LLMs with restricted resources -- consider our shared task!
More details on our website.
Important dates (AoE):
- Training/development data release: Mid May 2026 (already released)
- Test data release: end of June 2026
- Submission deadline: July 2026
We welcome participation from both academia and industry, especially teams interested in multilingual NLP, efficient LLM adaptation, low-resource machine translation, and multitask learning.
Please join the Google Group for further updates and registration details.
All good vibes,
On behalf of the LLMs with Limited Resources Shared Task organisers