Multiple choice · practice

20% of the grade20 questionsClick to answer

Pick an option for each question. The correct one turns green, a wrong pick turns red and shows the right answer, and the explanation appears underneath. The percentages are how the class voted — a low percentage marks a trap.

0 correct · 0 answered

Reset

Lectures 1–2

1. In autoregressive pretraining, where do the prediction targets come from?

A separate expert annotator assigns a class to each sentence Each observed token is the target for the preceding context The model's own generated continuation is used as the target Randomly hidden tokens are predicted from both sides

2. In $h_i^{(0)} = e(x_i) + p_i$, what information does $p_i$ supply?

The probability of the token as the next output How often the token appeared in the training set The position of the token in the sequence The model's confidence that the token is factually correct

3. What defines self-attention, as opposed to cross-attention?

Every token can attend only to itself Attention is computed without trainable parameters Queries, keys and values use an identical learned projection matrix Queries, keys and values come from the same input sequence

4. What distinguishes zero-shot from few-shot prompting?

Zero-shot freezes the weights; few-shot trains them Zero-shot uses no pretrained knowledge; few-shot does Zero-shot supplies no task demonstrations; few-shot supplies examples Zero-shot supplies no instruction; few-shot supplies one

5. For $W \in \mathbb{R}^{d \times k}$ and $\Delta W = BA$ of rank at most $r$, which shapes are compatible?

$B \in \mathbb{R}^{d \times r}$ and $A \in \mathbb{R}^{r \times k}$ $B \in \mathbb{R}^{d \times k}$ and $A \in \mathbb{R}^{r \times r}$ $B \in \mathbb{R}^{k \times r}$ and $A \in \mathbb{R}^{r \times d}$ $B \in \mathbb{R}^{r \times d}$ and $A \in \mathbb{R}^{k \times r}$

6. Why are different augmented views of an image used in visual self-supervision?

They guarantee preservation of every clinically important detail They provide related observations from which to learn consistent representations They ensure all images get identical feature vectors They replace the need for an optimisation objective

7. What is the purpose of a contrastive image–text training objective?

To give matching image–text pairs higher similarity than mismatched pairs To predict successive words of a caption from its image To reconstruct image pixels from text descriptions To assign images and text independently to fixed diagnostic labels

8. Which statement distinguishes image-to-report retrieval from report generation?

Retrieval predicts report tokens; generation ranks stored reports Retrieval selects an existing report; generation predicts a sequence of report tokens Retrieval uses a text query only; generation uses an image query only Both select stored reports, using different similarity metrics

9. For fixed features $z = f_\theta(x)$, a linear probe uses $p = \mathrm{softmax}(Wz + b)$. Which parameters are fitted?

$\theta$, $W$ and $b$ jointly No parameters; the class probabilities are known before training $W$ and $b$, while $\theta$ is fixed $\theta$ only, while $W$ and $b$ are fixed

10. What is external validation of a biomedical AI model, as opposed to an internal held-out test?

Evaluation on independently collected data from outside the development source Evaluation on a random subset withheld from the same dataset Evaluation after selecting the best model on the original validation subset Evaluation on augmented copies of the training images

Lecture 6

11. When a model requests a search in a ReAct-style system, where does the resulting observation come from?

The search tool executed by the surrounding program The model's prediction of what the search should return The description of the search tool supplied in the prompt The reasoning text that introduced the search request

12. Which description best captures the ReAct approach?

Complete a reasoning plan before acting, without further model decisions Interleave reasoning and actions, using observations to inform subsequent steps Generate the final answer first, then a reasoning trace to explain it Assign reasoning and action selection to two separately trained models

13. In ReAct's few-shot prompting setup, how do demonstrations influence the model?

They supply labelled examples for updating every model parameter They train a separate controller that chooses actions They provide example trajectories in the input context while the weights stay fixed They replace the external environment with stored answers

14. Why can external search help address the factual hallucinations discussed in ReAct?

It increases the amount of reasoning generated before any evidence is retrieved It supplies evidence that the model can use in subsequent reasoning It transfers responsibility for the final answer to the search engine It updates the model's weights after each search

15. What is the main distinction between workflows and agents?

Workflows use one model; agents require several cooperating models Workflows process text; agents must also process images Workflows use predefined code paths; agents let the LLM direct the process and tool use Workflows keep the weights fixed; agents update them after each action

16. One LLM call extracts findings, a second drafts a summary, a third translates it. Which pattern is this?

Routing Prompt chaining Parallelization Evaluator–optimizer

17. A system classifies a request and sends it to one of several specialist handlers. Which pattern is this?

Prompt chaining Evaluator–optimizer Parallelization Routing

18. Which capability is least directly tested by a static medical multiple-choice benchmark?

Selecting a diagnosis from a supplied list of alternatives Recalling a medical fact needed to answer a question Deciding what information to request as a consultation develops Recognising an association between symptoms and a disease

19. What is the role of agent-based modelling in the proposed clinical evaluations?

To simulate the actions and interactions of participants in a clinical environment To replace a clinical encounter with a single fixed question and answer To train the language model's next-token predictor on clinical textbooks To rank language models by parameter count and inference speed

20. An agent performs well in a clinical simulation. Which further step best addresses whether that result transfers to actual care?

Increase the number of runs in the same unchanged simulation Expand the set of static medical knowledge questions Test real-world impact in clinical studies and monitor after deployment Increase the similarity between generated explanations and reference explanations

Traps to remember

Tempting option	Why it is wrong
Workflows train or update weights	Both workflows and agents keep the weights fixed
Parallelization for a router	Routing selects one path; parallelization runs all
Tool description equals observation	The program executes the tool; the description is only instructions
Few-shot equals fine-tuning	Demonstrations are in context; weights unchanged
Zero-shot equals no pretrained knowledge	It uses pretraining, just no demonstrations
Retrieval generates reports	Retrieval selects an existing report
Balanced test set gives real precision	Precision depends on prevalence
0.90 softmax equals 90% disease probability	Softmax is relative to the candidate set
AUROC 0.97 settles the threshold	AUROC is threshold-free; the point is a clinical decision