Prompt Engineering, Alignment, and Summary

<p class="kicker">GRA 4164 · Lecture 11 · Fall 2026</p>

## Prompt Engineering,<br>Alignment, and Summary

### Leif Anders Thorsrud & Vegard H. Larsen — BI Norwegian Business School

---

The cheapest way to program a computer ever invented: **ask it**.

Route 3 from lecture 9 — adaptation with no training at all.

---

## Prompting

---

## Specificity is everything

- Prompt: `prompt` → *"It looks like you might have forgotten to include the details…"*
- Prompt: `prompt engineering` → a structured, detailed explanation
- The model **completes your text** — give it more to condition on

---

## Know what you're prompting

- **Cut-off & tools** — what does it know, can it search or run code?
- **Raw model or chatbot** — completion vs. aligned assistant
- **Context limit** — how much fits (lecture 9: tokens, truncation)
- Capabilities **change monthly** — re-check your assumptions

</div>

---

## Crafting the prompt

- **Clear and precise** — vague in, vague out
- **Context** — paste the document, the data, the example
- **Tone and format** — say what you want the output to look like
- **Iterate** — refine based on what comes back

---

## Show, don't tell: shot prompting

```text
Zero-shot:   Add 2+2

One-shot:    Add 2+3: 5
             Add 2+2:
```

- Examples in the prompt = **in-context learning** — no training, no weights touched

---

## Chain of thought

```text
Q: Roger has 5 tennis balls. He buys 2 cans of 3. How many now?
A: Roger started with 5. 2 cans of 3 is 6. 5+6=11. The answer is 11.
Q: The cafeteria had 23 apples. They used 20, bought 6 more. How many?
```

- Model answers: *"23 − 20 = 3, then 3 + 6 = 9. The answer is 9."*
- Zero-shot version: just append **"Let's think step by step"**

---

## Pitfalls

**Hallucinated citations · bias · confident nonsense · prompt injection.** The failure modes from lecture 7 don't disappear because you prompt well.

</div>

Prompt injection: instructions hidden in *data* the model reads.

---

## Alignment

---

## A raw model is not an assistant

- A base model **continues text** — ask it a question, it may continue with *more questions*
- Helpful? Harmless? Honest? **Not its objective** — next-token probability is
- Try it: *"Give me a reference for a paper by Leif Anders Thorsrud"*

---

## The alignment recipe

- **Instruction tuning** — fine-tune on instruction → answer pairs (lecture 9, route 2)
- **Preference alignment** — a separate model scores responses against human preferences; the base model is tuned against it
- Result: the **assistant persona** — helpfulness, refusals, hedging

---

## What alignment doesn't fix

- Hallucination is **reduced, not removed** — the objective still rewards fluency
- Alignment reflects **whose** preferences? A design choice, not a law of nature
- Lecture 7's rule survives: **starting point, never definitive answer**

---

## Grounding the model

---

## Retrieval-augmented generation

- The model can't cite what it never saw: **your** documents, **this week's** data
- So: **retrieve** the relevant text first, put it **in the prompt**, then ask
- The answer is grounded in retrievable, citable sources

---

## You built this already

![RAG pipeline: documents embedded into a vector store; a question is embedded, top-k chunks retrieved by cosine similarity, and an LLM answers from the retrieved context](../figures/rag-pipeline.svg)

---

RAG = **lecture 2** + **lecture 6** + an LLM.

The course's oldest tools fix the newest model's worst habit.

---

## The course in one pass

---

## The map, one last time

![Five-step pipeline from raw text to business decisions](../figures/text-as-data-workflow.svg)

---

## The arc

![Course timeline: counting words, learning from prediction, LLMs, wrap-up](../figures/course-timeline.svg)

- **Count** (L1–3) → **predict** (L4–6) → **pre-trained** (L7–11)
- Twice we saw counting and learning **coincide** — bigrams (L4), embeddings (L6)

---

One loss function ran the **whole course**.

From a 27×27 count table to GPT: −log P(next token).

---

## The exam, practically

- **3 hours, written, individual** — 70% of the grade; ECTS scale
- Revise the **in-sum slides** of every lecture + the **by-hand examples**: Naive Bayes (L3), bigram loss (L4), PPMI (L6), attention (L8)
- Methods: know **when to use what** and how to **interpret output** — the syllabus learning goals

---

## Thursday: presentations

**Lecture 12: group presentations.** 6 minutes per group + questions. Assignment allocation per group is on Itslearning. Submission of the collected assignments follows at semester's end.

</div>

---

## In sum

- **Prompting**: specific, contextual, iterative — examples and reasoning on demand
- **Alignment**: instruction tuning + preferences make the assistant — and don't make it a source
- **RAG**: when sources matter, retrieve them — lectures 2 + 6 + an LLM
- **The course**: count → predict → pre-trained, one loss the whole way

</div>

---

## Questions

## The prompting checklist

- **Model**: cut-off? tools? chat or raw? context limit?
- **Prompt**: specific ask · pasted context · output format · tone
- **Upgrade**: add examples (few-shot) · ask for steps (CoT) · iterate
- **Verify**: sources real? numbers right? injection risk?

## Instruction tuning ≠ task fine-tuning

<div class="cols">
<div class="callout">
<h3>Task fine-tuning (L9, route 1)</h3>
<p>One task, labelled data, new head — a specialist (e.g. sentiment classifier).</p>
</div>
<div class="callout">
<h3>Instruction tuning (L9, route 2)</h3>
<p>Many tasks phrased as instructions — a generalist that follows directions.</p>
</div>
</div>

## Beyond chat: tools and structure

- **Structured output** — ask for JSON with a fixed schema → text becomes *data* you can pipe onward
- **Function calling** — the model requests a tool (search, database, code); your system runs it and returns the result
- Chat is the demo; **pipelines** are the business value

---

## What's next

- **Lecture 12 (Tue): presentations** — 6 minutes per group, allocation on Itslearning
- **Submission**: the four assignments, collected, end of semester
- **Exam**: December — start from the in-sum slides