Hiring

Four postdoc positions in LLM alignment at Imperial College London.

I am hiring four postdocs in LLM alignment at Imperial College London. Each postdoc leads one of the five areas below. All five study how large language models behave in African and other low-resource languages, where most safety work has never been tested.

Areas

1. Multilingual red teaming

Safety filters trained on English often fail when the same request arrives in Hausa, Yoruba or Swahili, or in a mix of two languages. You will build attack sets in African languages, measure how often current models break, and train automated red-teamers that find new failures without a human writing each prompt.

2. Pluralistic alignment

A model that gives one answer on family, religion or authority gives a Western answer by default. You will collect preference data from communities in Nigeria and across Africa, and train models that can represent more than one set of values and say which one they are using.

3. Safety evaluation in low-resource languages

We need to know whether a safety score in a language with little data means anything. You will build benchmarks for hate speech, polarisation and misinformation, extending AfriHate, POLAR and BRIGHTER, and test whether LLM judges can be trusted to grade answers in languages they barely know.

4. Cross-lingual alignment methods

Refusal and honesty are trained in English and assumed to carry over to other languages. You will test where that transfer breaks, find the cause inside the model with interpretability tools, and develop fine-tuning and steering methods that fix it.

5. Speech safety

Most Hausa speakers will talk to these models, not type to them. You will test whether voice assistants and speech-to-speech models keep their safety behaviour when the request is spoken in an African language, with real accents, background noise and code-switching, and build the spoken test sets needed to measure it.

Who should apply

You have a PhD in NLP or machine learning, or will submit within six months. You have published at ACL, EMNLP, NAACL, NeurIPS, ICLR or a similar venue. Speaking an African language helps and is not required.

How to apply

Email s.muhammad@imperial.ac.uk with the subject line “Postdoc: “ followed by the area you want. Attach:

  • your CV;
  • your two best papers;
  • one paragraph on what you would do in your first six months.

I read every application and reply to those that include all three.

PhD students and visitors

PhD applicants apply through Imperial’s Department of Computing. Name me as a possible supervisor and email me once you have applied. If you are in Nigeria and want to start in NLP research, join HausaNLP or the Arewa Data Science Academy.