LXT ran a red-teaming exercise across languages to expose cases where multilingual models produced fluent output that lost a puzzle.
Public source
Publisher name
Public post
Riddles make a useful adversarial test because they break the thing translation systems lean on hardest: literal meaning. A riddle works through wordplay, ambiguity, and…
Company
LXT
High-quality AI data to accelerate the success of your AI projects.
- Industry
- Information Services
- Location
- Mississauga, CA
- Company size
- 201–500 employees
About LXT
LXT is an emerging leader in AI training data to power intelligent technology for global organizations. In partnership with an international network of contributors, LXT collects and annotates data across multiple modalities with the speed, scale and agility required by the enterprise. Our global expertise spans more than 145 countries and over 1000 language locales. Founded in 2010, LXT is headquartered in Canada with presence in the United States, UK, Egypt, India, Turkey and Australia. The company serves customers in North America, Europe, Asia Pacific and the Middle East. Learn more at lxt.ai.
See moreLatest activity
Latest activity from LXT
3 signals
Research & Knowledge
LXT published research on four African languages (arXiv 2608.11146) finding harmful prompts keeping under 10% of the English refusal signal.
Research & Knowledge
LXT researchers built LoDNA, a safety dataset for Twi, Hausa, Amharic, and Swahili that pairs literal translations of harmful prompts with culturally localized versions.
Discover more
Similar signals
Similar public activity from other companies.
Research & Knowledge
alphaXiv
alphaXiv published a paper identifying that LLMs fail during pretraining due to disjoint token spaces, even between two identical copies of English, suggesting that mapping languages into a shared semantic token space improves cross-lingual learning by up to 14x.
Research & Knowledge
Easyling.com
Easyling.com explained how cross-model agreement is used as a confidence signal in the extraction process, with three engines running over a corpus including TF-IDF, YAKE, and an LLM doing semantic extraction.
Research & Knowledge
NEC Laboratories America, Inc.
NEC Laboratories America, Inc. published a paper titled Mind the Gap in Cultural Alignment: Task-Aware Culture Management for Large Language Models on July 7, 2026.
Research & Knowledge
Liquid AI
Liquid AI published a paper showing that checkpoints with worse loss and worse pass@1 can turn into better models after RL, with TailSFT proposing a fix to record training example losses under the initial model and skip sequences whose loss improved the most.
Research & Knowledge
RWS