LLM Research Scientist (Pre-training & Computer Vision & Adversarial Robustness)

We're looking for experienced machine learning researchers with hands-on experience training and improving deep learning models end-to-end, across vision and language. You'll work on well-scoped empirical open-ended ML research problems.

Responsibilities

• Train image classifiers and generative image models from scratch, and fine-tune open-weight language models.

• Get the most out of limited data, compute, and model-size budgets.

• Make models robust — to adversarial inputs and to adversarial conversations.

• Compress models to meet hard size and latency constraints without sacrificing accuracy.

• Diagnose and resolve training issues.

Requirements

We are looking for candidates with strong expertise in one or more of the following areas:

Adversarial Robustness

Experience with:

• Adversarial training of image classifiers (e.g. PGD-based training, TRADES).

• Evaluating robust accuracy under standard threat models (e.g. L∞ attacks, AutoAttack) and avoiding gradient-masking pitfalls.

• Managing the robustness–accuracy trade-off and robust overfitting.

Efficient Computer Vision

• Training image classifiers end-to-end, especially for fine-grained recognition (many visually similar classes, few examples per class).

• Model compression: quantization, pruning, and knowledge distillation from large teachers into small students.

• Deploying models under hard size or latency budgets (on-device, edge, or embedded settings).

Generative Image Modeling

• Training image generative models from scratch: diffusion models, GANs, VAEs, or flow-based models.

• Iterating against sample-quality metrics such as FID.

• Training-efficiency tricks that produce good generators quickly and at small parameter counts.

LLM Post-Training & Behavioral Robustness

Hands-on experience with one or more of:

• Supervised fine-tuning and preference optimisation (DPO, RLHF, RLAIF) of open-weight language models, including building your own datasets via synthetic generation, noisy or weak supervision, and rejection sampling.

• Shaping conversational behaviour over multiple turns: resistance to persuasion and sycophancy, calibrated confidence, and knowing when to accept corrections.

• Alignment-style fine-tuning that changes a specific behaviour while preserving general capability.

Multilingual Pre-training

• Training multilingual or low-resource-language models from scratch.

• Tokenizer design across scripts and typologically diverse languages.