Zhou Ej

first year PhD @CambridgeLTL

[Zhou Ej @EjZhou582]

2. Language-agnostic calibration: middle layer provides better confidence signal. See the paper here: https://t.co/TP74ff1qen

[Zhou Ej @EjZhou582]

Heading to Budapest for #EMNLP2026!

Especially proud of these papers so I'm promoting them here again

If you’re at EMNLP and working on multilinguality, interpretability, come find me at the posters. I will be at poster session C and F :)

#NLP #EMNLP #Multilingual

[Zhou Ej @EjZhou582]

1. The Strong Platonic Representation Hypothesis, or the universal convergence of representation across languages. See the paper here: https://t.co/MjD3HlL5Ft

[Zhou Ej @EjZhou582]

Will be in SF this week for COLM;

happy to connect!

[Zhou Ej @EjZhou582]

A 3.5-star restaurant is "top gourmet" in Japan but "ordinary" in the US. Same score, different scale.

LLM evaluators do the same thing across languages, and it's a safety problem.

Our paper was accepted to COLM 2026 🧵

[Zhou Ej @EjZhou582]

A 3.5-star restaurant is "top gourmet" in Japan but "ordinary" in the US. Same score, different scale.

LLM evaluators do the same thing across languages, and it's a safety problem.

Our paper was accepted to COLM 2026 🧵

[Zhou Ej @EjZhou582]

@bminixhofer see you there :)

[Zhou Ej @EjZhou582]

📄 Lower-Resource, Higher Scores: Language Bias in LLM Evaluators

w/ @lucasresck2, Zack Hui, @annalkorhonen (Language Technology Lab, Cambridge)

https://t.co/GcL8V52ZiE

[Zhou Ej @EjZhou582]

A 3.5-star restaurant is "top gourmet" in Japan but "ordinary" in the US. Same score, different scale.

LLM evaluators do the same thing across languages, and it's a safety problem.

Our paper was accepted to COLM 2026 🧵

[Zhou Ej @EjZhou582]

Problem: it's invisible to standard eval.

Evaluators hit >90% pairwise accuracy but show up to a 43% gap in acceptance rate across languages under one threshold.

So harmful content in lower-resource languages is likelier to pass safety filters.

[Zhou Ej @EjZhou582]

We scored semantically identical content in 23 languages with reward models and LLM-as-a-Judge.

Lower-resource languages get scored MORE generously (ρ = −0.81), even though models are weaker in them. The bias contradicts what was expected.

[Caiqi Zhang @caiqizh]

🔥Check our EMNLP paper with @vlachos_nlp and @ZhijiangG

🤔Do We Need Language-Specific Fact-Checking Models? The Case of Chinese https://t.co/1Hee6fY9nx

‼️ We find the domain and cultural biases in the Chinese fact-checking area that necessitate language-specific tools!

[Caiqi Zhang @caiqizh]

🔥Conformity in Large Language Models🔥
Our latest paper dives into how LLMs align their answers with incorrect majorities. We explore the fascinating interplay between LLM behavior and human psychology!🧠

Preprint:
https://t.co/JbuJUkdYmr

#AI #NLP #LLMs #Conformity