Cram sheet · 30 minutes#
Everything that fits on one screen per topic. If you only have 30 minutes, read this page and the MCQ key. The longer pages are backup.
10 multiple-choice answers#
| Question cue | Answer |
|---|---|
| Autoregressive targets | Each observed token, given the preceding context |
| \(p_i\) in \(e(x_i)+p_i\) | Position of the token |
| Self-attention | Q, K, V from the same sequence |
| Zero-shot vs few-shot | No demonstrations vs demonstrations (weights frozen in both) |
| LoRA shapes | \(B\in\mathbb{R}^{d\times r}\), \(A\in\mathbb{R}^{r\times k}\) |
| Augmented views | Related observations → consistent representations |
| Contrastive image–text | Matching pairs higher similarity than mismatched |
| Retrieval vs generation | Select an existing report vs predict tokens |
| Linear probe | Fit \(W,b\); freeze \(\theta\) |
| External validation | Independent data from outside the development source |
| ReAct observation | Returned by the tool executed by the program |
| ReAct itself | Interleave reasoning and actions |
| ReAct few-shot | Example trajectories in context, weights fixed |
| Search helps hallucination | Supplies evidence for later reasoning |
| Workflow vs agent | Predefined paths vs LLM directs process/tools |
| Extract → draft → translate | Prompt chaining |
| Classify → specialist handler | Routing (not parallelization) |
| Static MCQ benchmark misses | Deciding what to request next |
| Agent-based modelling | Simulates actions/interactions of participants |
| Simulation success → next | Real-world clinical studies + post-deployment monitoring |
6 formulas#
\[\text{PPV} = \frac{\text{sens}\cdot p}{\text{sens}\cdot p + (1-\text{spec})(1-p)} \qquad \text{NPV} = \frac{\text{spec}(1-p)}{\text{spec}(1-p)+(1-\text{sens})p}\]
\[O/E = \frac{\text{observed deaths}}{\text{mean predicted risk}\times n}\]
- \(TP=\text{sens}\times P\), \(FN=(1-\text{sens})P\), \(FP=(1-\text{spec})N\), \(TN=\text{spec}\cdot N\).
- Calibration = predicted vs observed; discrimination = AUROC. O/E needs calibration.
- Softmax over \(k\) candidates: a score is relative, not a calibrated probability.
- Mean of readings loses trend: keep slope, min, max, last value.
6 open-question templates#
- "Lower loss but worse facts" → loss averages over common tokens; teacher forcing vs own output; evaluate rare critical facts with blinded clinicians.
- "Novel perturbation/drug" → categorical embedding cannot extrapolate; need structure/feature encoder or graph prior.
- "Biased/leaky evaluation" → say the specific leakage (future data, patient overlap, balancing, treatment paradox), then the fix (causal inputs, patient-level temporal split, real prevalence, exclude treated negatives).
- "Deployment decision" → intended use → unit of output → local validation → operating point (who bears FP vs FN) → monitoring + exit plan → economics.
- "Workflow or agent?" → can the paths be enumerated in advance? Yes → workflow; no → agent/orchestrator. Not "does it use an LLM".
- "Evaluate an LLM system" → build a test set with injected errors/missing data; measure outcome, process (evidence attribution, unnecessary steps) and cost; distinguish supported from accidentally correct.
5 rules that earn points#
- Name the mechanism, then the consequence. "Attention is permutation-equivariant, therefore both reports give the same averaged vector."
- Always separate discrimination from calibration. AUROC is threshold-free; decisions need a threshold and calibration.
- Say whose error it is. False positive vs false negative, and who signs off.
- State what a result cannot establish. One split, one site, no CIs → no significance, no transportability.
- For any model claim, ask: was the reference standard independent of the model? Anchoring and "zero discrepancies" are not validation.
The one-line answer to most essay questions#
The model's score is only a ranking; the clinical decision needs an operating point, calibration, validation on the local case mix, and a named owner for the residual risk.