Defending your own research: how do you do it in an interview?

In this article

1. Before you start
2. Why interviewers ask this
3. Defending your own research from first principles
4. Worked example
5. Trade-offs and edge cases
6. How to answer it in an interview
7. Common mistakes
8. Follow-up questions
9. Sources

Before you start

This page belongs to the research scientist interview guide, which shows where this round sits among the others. The terms used here are also collected in the glossary, and more guides are in the interview guides index.

Why interviewers ask this

The consequence of a weak defence is simple: if you cannot explain a basic choice in your own paper, the interviewer cannot tell whether you made it. Candidates report that some research roles include a job-talk-style session in which you present about ten slides on past work and "defend" it like a doctoral candidate. We found no public lab document describing this format, so treat it as reported by candidates.

A university source shows what a defence is meant to test. Waterloo lists mastery of the subject matter, proof that you are the author of the work, and the ability to engage in scholarly discourse in your research area. Candidates report questions of that kind: how a paper's dataset was labelled, how LoRA is initialised and which rank was used, and whether you have read a named interpretability paper.

A weak answer recites the abstract. A strong answer states the claim, gives a reason for each choice, and says where the evidence stops. The research taste questions test a related skill: judging which problems and results matter.

Defending your own research from first principles

Defending your own research

Defending your own research is the act of explaining a project you authored to an expert and justifying its claim, each method choice, each number and its limits, answering follow-up questions about any of them from your own understanding of the work.

Low-rank adaptation (LoRA)

Low-rank adaptation is a fine-tuning method that freezes the pretrained weights and trains a small update made of two thin matrices, B and A, whose product is added to a weight matrix, so far fewer parameters are trained than in full fine-tuning.

In plain words

Think of a driving test. The examiner does not ask whether you passed; they ask you to explain what you did at the junction and why. Defending research is the same: the interviewer asks about decisions, not outcomes. The analogy breaks in one place: a driving examiner has a fixed checklist, while a research interviewer follows whatever looks weakest.

In one precise sentence, you defend research by linking each claim to its evidence and each design choice to a reason you can state. Waterloo describes the goal as mastery of the subject, proof of authorship and the ability to engage in scholarly discourse. It is not winning a debate. The research brainstorm and problem-framing rounds test a different skill. Format varies, and Waterloo caps its defence presentation at 30 minutes, so practise a short version first.

Step 1: State the claim and the evidence

A defence starts with the claim and the evidence for it. Say the claim in one sentence, then the single experiment or result that supports it, then what a reader should conclude. Input: your whole project. Output: two sentences you can say without notes.

Everything else hangs from those two sentences. If the interviewer interrupts at slide three, you return to the claim and carry on from there. Prepare four blocks in this order: the claim and evidence, the method choices, the numbers, the limits. Each block should survive being asked about alone. You now have the frame; the next steps fill it with defensible detail.

Step 2: Defend each method choice, using LoRA as the example

Each method choice needs a reason you can state. A candidate reported being asked to introduce LoRA, say how it is initialised and why a given rank was used. Here is an answer built from Hu et al. (2021).

LoRA freezes the pretrained weight matrix W0 and learns an update ΔW equal to the product of B and A, where B is d by r, A is r by k and r is much smaller than d or k. The forward pass is h = W0x + BAx. Initialisation: "a random Gaussian initialization for A and zero for B", so the update ΔW is zero at the beginning of training. A Gaussian is a set of random numbers drawn from a bell curve, so only A starts random. The update is scaled by α/r, and the authors "set α to the first r we try and do not tune it", which is why changing r needs little re-tuning.

Our inference: training begins exactly at the pretrained model's behaviour. The toy example below shows the first update moves B only.

Step 3: Defend each number

Each number needs the experiment that produced it. When asked why you chose a rank, the strongest answer points at a sweep, not a habit.

In the LoRA paper, with LoRA on the attention query and value matrices (Wq and Wv) of generative pre-trained transformer (GPT)-3 (Generative Pre-trained Transformer 3), rank 1 scores 73.4 on WikiSQL, rank 8 scores 73.8 and rank 64 scores 73.5. The paper puts run-to-run noise at about 0.5%, so those differences are within noise (Hu et al., Table 6). The scores stay within noise from rank 1 to rank 64 for that configuration, so a small rank was enough there. Adapting Wq alone is different: its score rises from 68.8 at rank 1 to 70.5 at rank 4 in the same table.

The same paper warns that it does "not expect a small r to work for every task or dataset". Say your numbers, your noise estimate and the condition under which they would change. The dataset follows the same rule: state its size and splits, as the follow-up below shows.

Step 4: Say where it fails and what you do not know

A defence ends with the limits and the honest unknowns. Name one limitation of your own work before the interviewer finds it. The LoRA paper states that "LoRA also has its limitations", for example that batching inputs for different tasks, each with its own A and B, is not straightforward once the weights are merged.

For work you did not originate, separate what the authors showed from what you infer. Candidates report being asked whether they had read the paper Refusal in Language Models Is Mediated by a Single Direction and how it can be improved. The authors report one direction across 13 open-source chat models; erasing it stops refusals and adding it causes refusal even on harmless prompts. The evidence is interventional, meaning they changed the model and watched the behaviour. An improvement question is open-ended; propose what you would test, such as other models or behaviours, and label it as your idea, not a finding.

1 · Claim2 · Choices3 · Numbers4 · LimitsReplay

Step-by-step description

1. State the claim and its evidence.
2. Give a reason for each method choice.
3. Tie each number to its experiment.
4. End with limits and unknowns.

• A defence starts with the claim and the evidence for it.
• Each method choice needs a reason you can state.
• Each number needs the experiment that produced it.
• A defence ends with the limits and the honest unknowns.

Worked example

This example counts what LoRA trains and checks the zero start with a toy. The model size comes from the paper: GPT-3 has a model width of 12,288 and 96 layers, and adapting two attention matrices per layer gives 192 adapted matrices (Hu et al., 2021). The parameter count of the paper is 2 times the number of adapted matrices times the width times the rank.

For one 12,288 by 12,288 matrix, full fine-tuning trains 150,994,944 weights, while LoRA at rank 8 trains 196,608, which is 0.1302% of them.¹ Across all 192 matrices, rank 1 trains 4,718,592 parameters and rank 4 trains 18,874,368, which match the roughly 4.7 million row of the paper's Table 4 and the 18.8 million figure in the paper's parameter-budget tables.¹

The table's computed rows come from the saved script.¹ The paper reports 10,000 times fewer trainable parameters and three times less graphics processing unit (GPU) memory than full fine-tuning of GPT-3 with Adam (Hu et al., 2021). Adam is a common training optimiser, the rule that updates the weights.

The rank sweep from Step 3 is charted below for LoRA on Wq and Wv; the scores stay within the paper's noise from rank 1 to rank 64 (Hu et al., Table 6).

Text description of the diagram

• With LoRA on Wq and Wv, the GPT-3 WikiSQL score is 73.4 at rank 1 and 73.8 at rank 8.
• The scores stay within noise from rank 1 to rank 64.

The zero start can be checked on a toy with illustrative numbers: W0 is a 3 by 2 matrix, x is (1, 2), A is (0.3, −0.2) and B is all zeros. The output W0x is (2, 2, 0) before and after adding BAx, so nothing changes at step zero.¹ The gradient of the loss with respect to A is zero while B is zero, because A's effect on the output is multiplied by B. The gradient with respect to B is nonzero, so the first update moves B only. This follows from the arithmetic of our toy; it is not a claim from the paper.¹ After B becomes (0.1, 0, −0.1), the output is (1.99, 2.0, 0.01).¹

Trade-offs and edge cases

Every defensible method has a cost, and naming it is part of the defence. LoRA adds no inference latency because the update can be merged into the weights, while adapters add latency, up to about 30% in small-batch online settings (Hu et al., 2021). Training memory also falls, because the frozen weights need no gradients or optimiser state, and the checkpoint shrinks because only A and B are saved.

Text description of the diagram

• Adapters add inference latency, up to about 30% in small-batch online settings.
• LoRA adds no inference latency because the update can be merged into the weights.

Edge cases to prepare:

• Small rank on a distant task. The paper does not expect a small rank to work everywhere, so a task far from the pretraining data may need more.
• Which weights you adapt. In the paper most experiments adapt only the attention weights and freeze the rest. Under a fixed budget of 18 million parameters, it found that adapting more matrices at a lower rank beat one matrix at a higher rank, and adapting two attention matrices at rank 4 gave the best overall result (Hu et al., Section 7.1).
• Mechanism claims. A single direction is a finding about how refusal is represented, not proof that safety is solved or fully understood. The same paper lists jailbreaking as an application. Jailbreaking is getting a model to answer requests it is meant to refuse, so the finding cuts both ways.
• Team work. Say which parts you built. Interviewers probe the part you did, and overclaiming is the fastest way to lose them. Do not share confidential employer details.

How to answer it in an interview

Replace the sweep claim with your own results. The spoken structure: say what the method is, say the initialisation and its reason, give the evidence for the choice, then name the caveat.

Common mistakes

• "LoRA always matches full fine-tuning." (wrong) The paper reports on-par or better results on the models and tasks it tested and warns a small rank may not work everywhere. (right) "It matched on the tasks tested; here is where I would not expect that."
• "I used a large rank because higher is better." (wrong) When adapting Wq and Wv on GPT-3, the scores are flat from rank 1 to rank 64 within noise. (right) "I swept ranks and picked the smallest one inside the noise."
• "Both matrices start random." (wrong) Only A is Gaussian; B is zero. (right) "A is random, B is zero, so the update starts at zero."
• "The α over r scaling is a tuned hyperparameter." (wrong) The paper sets α to the first r it tries and does not tune it. (right) "I kept the default scaling and varied only r."
• "Defending means winning the argument." (wrong) Waterloo describes it as showing mastery of the subject, proof of authorship and the ability to engage in scholarly discourse. (right) "I explain, concede real limits and say what I would test next."

Follow-up questions

How do you describe the dataset of your research paper?

Give its source, size, splits, how labels were made and its known biases, in that order. For example, the LoRA paper describes WikiSQL as having 56,355 training and 8,421 validation examples (Hu et al., Appendix C). Ground truth is the label you treat as correct when scoring, so say who or what produced it and how disagreements were settled.

How do you remove outlier labels from annotators in regression?

Remove outlier annotator labels with a rule fixed in advance. This is general practice, not a sourced result. Check annotator agreement first, then drop labels far from the median of several annotators or train with a loss that resists outliers, such as the Huber loss, and report what you removed.

Have you read the refusal-direction paper, and how can it be improved?

Say what it shows: refusal is mediated by a one-dimensional direction in the residual stream. The residual stream is the running vector each transformer layer reads from and adds to. Then offer one testable extension, labelled as your idea, such as other models or behaviours.

What if the interviewer asks about a part of the project you did not build?

Say so directly, describe what you understand of it and what you did own. Candidates are better served by a bounded, honest answer than by a guess that the next follow-up exposes.

What was the hardest part of tuning, and what was the budget?

Name the real constraint (compute, time or data), what you tried within it and what you dropped, using figures from your own logs. The LoRA paper models this: it fixed a budget of 18 million trainable parameters on GPT-3 and compared ways of spending it (Hu et al., Section 7.1).

Sources

1. LoRA: Low-Rank Adaptation of Large Language Models (arXiv 2106.09685 v2) — arXiv (Hu et al.), ICLR 2022, 16 Oct 2021
2. Refusal in Language Models Is Mediated by a Single Direction (arXiv 2406.11717 v3) — arXiv (Arditi et al.), 30 Oct 2024
3. Preparing for your thesis defence — University of Waterloo (undated page, accessed 2026-10-26), 26 Oct 2026