Career roadmap
ML Research Scientist
Push the method forward: read the literature, form hypotheses, run rigorous experiments, and publish what holds.
Before you start Research Scientist
- Linear algebra, calculus and probability at degree level
- Strong Python and PyTorch
- Comfort reading academic papers
Mathematical foundations
Research requires derivation, not just application. This stage cannot be shortened.
The language of every architecture and every proof you will read.
- Matrix decompositions and eigenvalues
- Vector spaces and projections
- Matrix calculus and gradients
- Numerical stability
Generative models and uncertainty quantification are built on it.
- Distributions and conjugacy
- Maximum likelihood and Bayesian inference
- Information theory: entropy, KL divergence
- Concentration inequalities
Why training works, when it fails, and what the optimiser is actually doing.
- Convexity and gradient descent
- Stochastic optimisation and variance
- Adaptive methods and their behaviour
- Loss landscapes and initialisation
Generalisation, capacity and the questions the field still cannot answer.
- Bias-variance and generalisation bounds
- Overparameterisation and double descent
- Regularisation as inductive bias
- Scaling laws
You do not understand a method until you have implemented it without a library.
- Autograd from first principles
- Attention and transformer blocks
- Optimisers implemented manually
- Verifying gradients numerically
BuildDerive and implement backpropagation, an optimiser and attention from scratch, with tests.
Reading and reproducing
The research skill that compounds: reading critically and reproducing reliably.
Volume plus scepticism. Most claims do not survive careful reading.
- Efficient reading strategies
- Identifying the actual contribution
- Spotting weak baselines and cherry-picking
- Tracking a research area over time
The most respected portfolio artefact for someone without publications.
- Working from paper to implementation
- Handling missing hyperparameters
- Compute-constrained reproduction
- Reporting failures honestly
Rigour is what distinguishes research from tinkering.
- Ablations that isolate the mechanism
- Seeds, variance and error bars
- Fair baseline construction
- Avoiding test set contamination
Bad tooling wastes more research time than bad ideas.
- Experiment tracking and configuration
- Reproducible environments and seeds
- Efficient data loading
- Debugging training instability
Almost everyone outside a frontier lab is compute-constrained.
- Scaled-down proxy experiments
- Mixed precision and gradient accumulation
- Distributed training basics
- Choosing experiments by information gained
BuildReproduce three papers from scratch and publish the code with an honest report on what did not replicate.
Specialisation
Research careers are built on depth in one area, not breadth across many.
How current models are actually built and trained at scale.
- Transformer variants and efficiency
- Pretraining objectives
- Scaling laws and compute allocation
- Long context approaches
One of the most active and best-funded research areas.
- Instruction tuning and preference learning
- RLHF and its alternatives
- Reward modelling and reward hacking
- Evaluation of alignment
Understanding what models compute internally, and a growing hiring area.
- Probing and representation analysis
- Circuit and feature analysis
- Sparse autoencoders
- Causal intervention methods
Quantisation, sparsity and inference optimisation have direct commercial value.
- Quantisation methods and quality
- Sparsity and mixture of experts
- Distillation approaches
- Inference-time compute scaling
Benchmarks are saturating and contaminated. Better evaluation is genuinely open work.
- Benchmark design and contamination
- Capability elicitation
- Human evaluation methodology
- Measuring reasoning claims
BuildReach the frontier of one subfield: read everything current and identify an unanswered question.
Producing research
Original contribution, written up to a standard reviewers accept.
The hardest and least taught skill in research.
- Gaps in the literature
- Questions that are tractable with your compute
- Negative results worth reporting
- Scoping to a publishable unit
A good result written badly does not get accepted or cited.
- Paper structure and narrative
- Figures that carry the argument
- Precise claims and stated limitations
- Related work positioning
Reviewing teaches you what reviewers look for in your own work.
- Reviewing papers constructively
- Responding to reviews
- Rebuttal strategy
- Venue selection
Modern research is collaborative; solo work is rare and slow.
- Working with co-authors
- Open-source research contributions
- Building a public research profile
- Finding mentors and advisors
Dual use, data provenance and honest reporting.
- Data licensing and provenance
- Dual use considerations
- Reporting standards and reproducibility
- Responsible disclosure of capabilities
BuildComplete one original project from question to preprint, with code and reproducible results.
Interview preparation
Research interviews test depth, taste and the ability to defend your own work.
Derivations on a whiteboard are standard for research positions.
- Derive backpropagation for a given layer
- Attention complexity and memory
- Probability and expectation problems
- Optimisation behaviour questions
Discuss a recent paper critically, including what you think is wrong with it.
- Summarising a contribution precisely
- Identifying weaknesses in method
- Proposing follow-up experiments
- Placing it in the literature
The core of a research interview: your project, questioned hard.
- Explaining your contribution clearly
- Justifying design choices
- Discussing limitations honestly
- What you would do with more compute
Implement a method from a description, in PyTorch, correctly.
- Implementing a layer or loss from a paper
- Debugging a training loop
- Efficient tensor operations
- Numerical stability
Without publications, reproductions and preprints are the accepted substitute.
- Reproductions with honest reports
- A preprint or workshop paper
- Open-source research code
- A public research blog
BuildA public research portfolio: reproductions, a preprint, and clean, documented code.
Research Scientist tools on your CV
- PyTorch
- JAX
- Weights & Biases
- Hugging Face
- LaTeX
- Slurm / distributed training
- arXiv
What Research Scientist employers ask to see
- Three paper reproductions with honest replication reports
- A preprint or workshop paper
- Open-source research code others have used
- A public research blog with technical depth
Small, highly competitive and concentrated in labs, big tech and well-funded startups. A PhD is common but demonstrated publications or reproductions can substitute.
Content last reviewed 2026-08-31. Guidance only — no institute or paid placement is endorsed anywhere in this book.