LLM Interview QA - 3

10 essential LLM interview questions on configuration parameters, temperature, context windows, decoding strategies, determinism, and generation control — with diagrams and examples.

Author

Vectoring AI

Published

20 May 2026

Keywords

LLM interview, LLM parameters, temperature, top-p nucleus sampling, top-k sampling, context window, decoding strategies, greedy decoding, beam search, deterministic generation, max tokens, frequency penalty, stop sequences

Introduction

This is Part 3 of our LLM Interview QA series, focused on LLM configuration and generation control. Understanding how to configure LLM parameters — temperature, sampling strategies, context windows, and decoding methods — is essential for building reliable AI systems.

For foundational LLM concepts (transformers, attention, RAG, RLHF), see LLM Interview QA - 1. For advanced topics (scaling, quantization, agents), see LLM Interview QA - 2. For ML fundamentals, see ML Interview QA - 1.

Q1: What are the main configurable parameters when calling an LLM API?

Answer:

When making an LLM API call, several parameters control the behavior, quality, and cost of the generated output.

graph TD
    CONFIG["LLM Configuration Parameters"]
    CONFIG --> GEN["Generation Control"]
    CONFIG --> SAMP["Sampling Parameters"]
    CONFIG --> OUT["Output Control"]
    CONFIG --> SYS["System Parameters"]

    GEN --> G1["temperature"]
    GEN --> G2["top_p (nucleus)"]
    GEN --> G3["top_k"]
    GEN --> G4["seed"]

    SAMP --> S1["frequency_penalty"]
    SAMP --> S2["presence_penalty"]
    SAMP --> S3["repetition_penalty"]
    SAMP --> S4["logit_bias"]

    OUT --> O1["max_tokens / max_new_tokens"]
    OUT --> O2["stop sequences"]
    OUT --> O3["n (num_return_sequences)"]
    OUT --> O4["stream"]

    SYS --> SY1["model"]
    SYS --> SY2["system prompt"]
    SYS --> SY3["response_format"]
    SYS --> SY4["tools / functions"]

    style CONFIG fill:#56cc9d,stroke:#333,color:#fff
    style GEN fill:#6cc3d5,stroke:#333,color:#fff
    style SAMP fill:#ffce67,stroke:#333
    style OUT fill:#ff7851,stroke:#333,color:#fff

Parameter Overview

Parameter	Range	Default	Purpose
`temperature`	0.0 – 2.0	1.0	Controls randomness of output
`top_p`	0.0 – 1.0	1.0	Nucleus sampling threshold
`top_k`	1 – vocab_size	50 (varies)	Limits token candidates
`max_tokens`	1 – context_limit	Model-specific	Maximum output length
`frequency_penalty`	-2.0 – 2.0	0.0	Penalizes repeated tokens
`presence_penalty`	-2.0 – 2.0	0.0	Encourages topic diversity
`seed`	Any integer	None	Enables deterministic output
`stop`	List of strings	None	Stops generation at specific tokens
`n`	1+	1	Number of completions to generate

Practical Configuration Examples

Use Case	temperature	top_p	max_tokens	Other
Code generation	0.0 – 0.2	1.0	2048	`stop=["\n\n"]`
Creative writing	0.8 – 1.2	0.95	4096	`frequency_penalty=0.5`
Data extraction	0.0	1.0	512	`response_format=json`
Chat conversation	0.7	0.9	1024	`presence_penalty=0.3`
Factual Q&A	0.0 – 0.3	1.0	256	—

Q2: What is temperature and how does it affect LLM output?

Answer:

Temperature controls the randomness of the probability distribution over the vocabulary at each generation step. It’s applied to the logits before the softmax function.

Mathematical Definition

Given logits z_i for each token i in the vocabulary:

P(w_i) = \frac{e^{z_i / T}}{\sum_j e^{z_j / T}}

where T is the temperature.

graph LR
    subgraph T0["Temperature = 0 (Greedy)"]
        T0_1["Token A: 99.9%"]
        T0_2["Token B: 0.1%"]
        T0_3["Token C: ~0%"]
    end

    subgraph T07["Temperature = 0.7"]
        T07_1["Token A: 75%"]
        T07_2["Token B: 20%"]
        T07_3["Token C: 5%"]
    end

    subgraph T1["Temperature = 1.0 (Default)"]
        T1_1["Token A: 60%"]
        T1_2["Token B: 25%"]
        T1_3["Token C: 15%"]
    end

    subgraph T2["Temperature = 2.0"]
        T2_1["Token A: 40%"]
        T2_2["Token B: 32%"]
        T2_3["Token C: 28%"]
    end

    style T0 fill:#56cc9d,stroke:#333,color:#fff
    style T07 fill:#6cc3d5,stroke:#333,color:#fff
    style T1 fill:#ffce67,stroke:#333
    style T2 fill:#ff7851,stroke:#333,color:#fff

Effect of Temperature

Temperature	Distribution	Behavior	Output Character
T → 0	Extremely peaked	Always picks highest-probability token	Deterministic, repetitive, safe
T = 0.3	Slightly softened	Mostly picks top tokens, rare surprises	Conservative, coherent
T = 0.7	Moderately spread	Balanced between likely and creative	Good default for most tasks
T = 1.0	Original distribution	Model’s “natural” uncertainty	Raw model behavior
T > 1.0	Flattened	Low-probability tokens become likely	Creative but potentially incoherent
T = 2.0	Nearly uniform	Almost random selection	Chaotic, nonsensical

Intuition

Think of temperature as a “creativity knob”:

Low temperature (0–0.3): The model is confident and focused — it picks the most obvious next word. Great for factual tasks, code, structured extraction.
Medium temperature (0.5–0.8): The model is balanced — it explores alternatives while staying coherent. Best for general chat and writing.
High temperature (1.0+): The model is adventurous — it considers unlikely words, producing surprising or creative outputs.

Common Interview Follow-Up: “What does temperature=0 actually mean?”

Setting temperature=0 is a shortcut for greedy decoding — the model always selects the single highest-probability token. However:

It’s still based on floating-point arithmetic, so minor non-determinism can occur across hardware
Most APIs interpret temperature=0 as “return the argmax token” deterministically
Some providers require setting a seed parameter for guaranteed reproducibility

Q3: What is the difference between Top-p (nucleus) sampling and Top-k sampling?

Answer:

Both Top-p and Top-k are token filtering strategies that limit which tokens are considered during generation, but they differ in how they determine the candidate set.

Top-k Sampling

Select the k most probable tokens and redistribute probability among them:

V_{\text{top-k}} = \{w_1, w_2, \ldots, w_k\} \quad \text{(ordered by probability)}

Top-p (Nucleus) Sampling

Select the smallest set of tokens whose cumulative probability exceeds p:

V_{\text{top-p}} = \text{smallest } V' \text{ such that } \sum_{w \in V'} P(w) \geq p

graph TD
    subgraph TopK["Top-k = 3 (Fixed Size)"]
        direction LR
        K1["'the' (0.40) ✓"]
        K2["'a' (0.25) ✓"]
        K3["'my' (0.15) ✓"]
        K4["'his' (0.10) ✗"]
        K5["'our' (0.05) ✗"]
        K6["'their' (0.03) ✗"]
    end

    subgraph TopP["Top-p = 0.9 (Dynamic Size)"]
        direction LR
        P1["'the' (0.40) ✓ → cumulative: 0.40"]
        P2["'a' (0.25) ✓ → cumulative: 0.65"]
        P3["'my' (0.15) ✓ → cumulative: 0.80"]
        P4["'his' (0.10) ✓ → cumulative: 0.90 ≥ p"]
        P5["'our' (0.05) ✗"]
        P6["'their' (0.03) ✗"]
    end

    style TopK fill:#6cc3d5,stroke:#333,color:#fff
    style TopP fill:#56cc9d,stroke:#333,color:#fff

Key Differences

Aspect	Top-k	Top-p
Candidate set size	Fixed (always k tokens)	Dynamic (varies per step)
Adapts to distribution shape	No — same k regardless of certainty	Yes — fewer tokens when confident
Risk when distribution is peaked	Includes unlikely tokens unnecessarily	Naturally narrows to top few
Risk when distribution is flat	May exclude reasonable tokens	Naturally includes more candidates

Why Top-p is Generally Preferred

Consider two scenarios at different generation steps:

Step A (peaked distribution): Model is 95% sure the next word is “Paris”

Top-k=50: Considers 50 tokens (49 are noise)
Top-p=0.95: Considers only 1-2 tokens (adaptive!)

Step B (flat distribution): Model is uncertain, many tokens are equally likely

Top-k=50: Might miss some reasonable candidates if vocabulary is large
Top-p=0.95: Includes all tokens until 95% mass is covered (could be 100+ tokens)

Combining Top-k and Top-p

In practice, many systems use both simultaneously:

First apply Top-k to limit to k candidates
Then apply Top-p within those k candidates

This provides both an upper bound (Top-k) and adaptive filtering (Top-p).

Recommended Settings

Task	top_k	top_p	Rationale
Deterministic (code, facts)	1	1.0	Equivalent to greedy
Balanced (chat)	40-50	0.9	Diverse but coherent
Creative (stories)	100+	0.95	Wide exploration
Structured output (JSON)	5-10	0.8	Limited, safe choices

Q4: What is the context window and how does it constrain LLM behavior?

Answer:

The context window (also called context length or maximum sequence length) is the total number of tokens an LLM can process in a single inference call — this includes both input tokens and output tokens.

\text{Context Window} = \text{Input Tokens (prompt)} + \text{Output Tokens (completion)}

graph LR
    subgraph CW["Context Window (e.g., 128K tokens)"]
        direction LR
        SYS["System Prompt<br/>(500 tokens)"]
        CTX["Retrieved Context / RAG<br/>(10,000 tokens)"]
        HIST["Conversation History<br/>(5,000 tokens)"]
        USER["User Message<br/>(200 tokens)"]
        RESP["Model Response<br/>(max_tokens: 4,096)"]
    end

    style CW fill:#56cc9d,stroke:#333,color:#fff
    style RESP fill:#ffce67,stroke:#333

Context Window Sizes (2024–2026)

Model	Context Window	Notes
GPT-3.5 Turbo	16K tokens	~12K words
GPT-4	128K tokens	~96K words
GPT-4o	128K tokens	~96K words
Claude 3.5 Sonnet	200K tokens	~150K words
Gemini 1.5 Pro	1M–2M tokens	Longest available
LLaMA 3.1	128K tokens	Open-source
Mistral Large	128K tokens
DeepSeek-V3	128K tokens

What Happens When You Exceed the Context Window?

Behavior	Description
Truncation	Oldest tokens are dropped (APIs return error or truncate)
Error	API rejects the request if input exceeds limit
Degraded performance	Even within limits, performance drops in the “middle”

Context Window vs. Effective Context

Key insight for interviews: The advertised context window is not the same as effective context:

Concept	Meaning
Maximum context	Technical limit the model supports
Effective context	Length at which performance remains high
“Lost in the middle”	Information in the center of long contexts is often missed
Needle-in-a-haystack	Benchmark: can the model find a fact placed at position X?

Strategies for Context Window Management

Strategy	How it works
Chunking + RAG	Only retrieve relevant chunks, don’t stuff everything
Summarization	Compress conversation history into summaries
Sliding window	Keep recent messages + system prompt, drop old middle
Hierarchical context	Summary of old messages + full recent messages
Prompt compression	Use tools like LLMLingua to compress prompts

Cost Implications

Context window directly affects cost:

\text{Cost} = (\text{Input tokens} \times \text{price/input token}) + (\text{Output tokens} \times \text{price/output token})

Longer contexts mean higher costs, higher latency, and more KV-cache memory usage.

Q5: Is LLM generation deterministic? How do you achieve reproducible outputs?

Answer:

By default, LLM generation is non-deterministic — the same prompt can produce different outputs across calls. This is intentional but can be controlled.

graph TD
    subgraph NonDet["Non-Deterministic (Default)"]
        ND1["Same prompt"]
        ND2["Run 1: 'The capital is Paris.'"]
        ND3["Run 2: 'Paris is the capital of France.'"]
        ND4["Run 3: 'France's capital city is Paris.'"]
        ND1 --> ND2
        ND1 --> ND3
        ND1 --> ND4
    end

    subgraph Det["Deterministic (Configured)"]
        D1["Same prompt + seed + temp=0"]
        D2["Run 1: 'The capital is Paris.'"]
        D3["Run 2: 'The capital is Paris.'"]
        D4["Run 3: 'The capital is Paris.'"]
        D1 --> D2
        D1 --> D3
        D1 --> D4
    end

    style NonDet fill:#ffce67,stroke:#333
    style Det fill:#56cc9d,stroke:#333,color:#fff

Sources of Non-Determinism

Source	Explanation	Controllable?
Sampling (temperature > 0)	Random token selection from distribution	Yes — set temperature=0
Top-p / Top-k filtering	Random selection within candidate set	Yes — set top_p=1, top_k=1
Floating-point non-determinism	GPU parallel operations not strictly ordered	Partially — depends on hardware
Batching effects	Different batch compositions may affect computation	No (server-side)
Model updates	Provider may update model without notice	No (use versioned models)
System prompt caching	Some providers cache and may route differently	No

How to Achieve Deterministic Output

Method	What it does	Guarantee Level
`temperature=0`	Greedy decoding (argmax)	High — nearly deterministic
`seed` parameter	Fixes random state for sampling	High (API-dependent)
`temperature=0` + `seed`	Both greedy and fixed state	Highest available
Self-hosted + fixed seed + deterministic CUDA	Full control over hardware	True determinism

When Determinism Matters

Use Case	Need Deterministic?	Why
Unit testing	Yes	Reproducible test assertions
Evaluation/benchmarks	Yes	Fair comparison across models
Caching	Yes	Same input → cache hit
Audit/compliance	Yes	Reproducible decisions
Creative writing	No	Variety is desired
Chat conversations	No	Natural variation is expected

Important Caveat

Even with temperature=0 and a seed, exact determinism is not always guaranteed:

GPU floating-point operations may vary across hardware versions
API providers may route requests to different hardware
Model quantization can introduce slight variations
OpenAI states: “deterministic outputs are not guaranteed” even with seed (but are “mostly deterministic”)

Q6: What are the main decoding strategies and when should you use each?

Answer:

Decoding is the process of selecting which token to generate next given the probability distribution from the model. The choice of decoding strategy dramatically affects output quality.

graph TD
    DECODE["Decoding Strategies"]
    DECODE --> DETERM["Deterministic"]
    DECODE --> STOCH["Stochastic (Sampling)"]
    DECODE --> HYBRID["Hybrid / Advanced"]

    DETERM --> GREEDY["Greedy Search<br/>Pick argmax at each step"]
    DETERM --> BEAM["Beam Search<br/>Track top-n hypotheses"]

    STOCH --> PURE["Pure Sampling<br/>Sample from full distribution"]
    STOCH --> TOPK["Top-k Sampling<br/>Sample from top k tokens"]
    STOCH --> TOPP["Top-p Sampling<br/>Sample from nucleus"]
    STOCH --> TEMP_SAMP["Temperature Sampling<br/>Reshape distribution then sample"]

    HYBRID --> SPEC["Speculative Decoding<br/>Draft + verify"]
    HYBRID --> CONTRAST["Contrastive Decoding<br/>Subtract weak model's distribution"]
    HYBRID --> GUIDED["Guided/Constrained<br/>Enforce output structure"]

    style DECODE fill:#56cc9d,stroke:#333,color:#fff
    style DETERM fill:#6cc3d5,stroke:#333,color:#fff
    style STOCH fill:#ffce67,stroke:#333
    style HYBRID fill:#ff7851,stroke:#333,color:#fff

Detailed Comparison

Strategy	How it works	Pros	Cons
Greedy	Always pick highest probability token	Fast, deterministic, simple	Repetitive, misses better sequences
Beam Search	Track top-n partial sequences	Finds higher-probability sequences	Still repetitive, expensive, poor for open-ended
Top-k Sampling	Sample from top k tokens	Reduces nonsense, some diversity	Fixed k not adaptive to distribution
Top-p Sampling	Sample from smallest set covering p mass	Adaptive to uncertainty, natural	Slightly less predictable
Temperature + Sampling	Reshape distribution then sample	Fine-grained control	Need to tune parameter
Speculative Decoding	Small model drafts, large model verifies	2-3x faster, same quality	Needs draft model
Contrastive Decoding	Subtract amateur model’s preferences	Reduces repetition, more coherent	Complex setup
Constrained Decoding	Force output to follow grammar/schema	Guarantees valid structure	Limits expressiveness

Greedy Search: The Simplest Strategy

At each step, pick the token with the highest probability:

w_t = \arg\max_{w} P(w | w_{1:t-1})

Problem: Greedy search is locally optimal but not globally optimal. A low-probability token now might lead to a much better overall sequence.

Example: “The dog” (0.4) → “has” (0.9) gives sequence probability 0.36, while “The nice” (0.5) → “woman” (0.4) gives 0.20. Greedy picks “nice” first but misses the better path.

Beam Search: Exploring Multiple Paths

Maintains num_beams parallel hypotheses:

Beam 1: "The" → "dog" → "has" → "a"      (prob: 0.36 × ...)
Beam 2: "The" → "nice" → "woman" → "is"   (prob: 0.20 × ...)
Beam 3: "The" → "cat" → "sat" → "on"      (prob: 0.15 × ...)

When to use beam search:

Translation (known output length)
Summarization (structured output)
NOT for open-ended generation (causes repetition)

When to Use Which Strategy

Task	Recommended Strategy	Why
Code generation	Greedy (temp=0)	Correctness over creativity
Translation	Beam search (beams=4-5)	Quality over diversity
Creative writing	Top-p=0.95, temp=0.8	Diversity and surprise
Chat/conversation	Top-p=0.9, temp=0.7	Natural but coherent
Structured extraction	Constrained decoding	Must follow schema
JSON output	Greedy + grammar constraints	Validity guaranteed
Fast inference	Speculative decoding	Speed with no quality loss

Q7: What are frequency penalty and presence penalty, and how do they reduce repetition?

Answer:

Frequency penalty and presence penalty are post-processing adjustments to token logits that discourage the model from repeating itself.

Mathematical Definitions

The logit for token i is adjusted before sampling:

z_i' = z_i - (\text{frequency\_penalty} \times \text{count}(i)) - (\text{presence\_penalty} \times \mathbb{1}[\text{count}(i) > 0])

where \text{count}(i) is how many times token i has appeared in the output so far.

graph TD
    subgraph FP["Frequency Penalty"]
        FP1["Penalizes proportionally to<br/>how MANY times token appeared"]
        FP2["Token appeared 5× → big penalty"]
        FP3["Token appeared 1× → small penalty"]
        FP4["Effect: Reduces repetitive words"]
    end

    subgraph PP["Presence Penalty"]
        PP1["Penalizes equally if token<br/>appeared AT ALL (binary)"]
        PP2["Token appeared 5× → same penalty as 1×"]
        PP3["Token never appeared → no penalty"]
        PP4["Effect: Encourages new topics"]
    end

    style FP fill:#6cc3d5,stroke:#333,color:#fff
    style PP fill:#ffce67,stroke:#333

Comparison

Aspect	Frequency Penalty	Presence Penalty
Scales with count?	Yes (proportional)	No (binary: appeared or not)
Range (OpenAI)	-2.0 to 2.0	-2.0 to 2.0
Primary effect	Reduces word-level repetition	Encourages topic diversity
Use case	Avoid saying “very very very…”	Avoid staying on same topic
Analogy	“Don’t repeat words”	“Talk about new things”

Practical Examples

Without penalties (both = 0): > “The weather is nice. The weather is really nice. The weather makes me happy. The weather…”

With frequency_penalty = 0.8: > “The weather is nice. It’s a beautiful day. The sunshine makes me happy. I think I’ll go outside…”

With presence_penalty = 1.0: > “The weather is nice. I’ve been reading a great book lately. My garden is blooming. Tomorrow I plan to cook…”

Repetition Penalty (Hugging Face)

Hugging Face uses a multiplicative repetition_penalty instead:

z_i' = \begin{cases} z_i / \text{repetition\_penalty} & \text{if } z_i > 0 \text{ and token appeared} \\ z_i \times \text{repetition\_penalty} & \text{if } z_i < 0 \text{ and token appeared} \end{cases}

repetition_penalty = 1.0: No effect
repetition_penalty = 1.2: Moderate de-repetition (common default)
repetition_penalty > 1.5: Strong — may cause incoherence

Q8: What is `max_tokens` and how does it interact with the context window?

Answer:

max_tokens (or max_new_tokens) sets the maximum number of tokens the model will generate in its response. It’s a hard cap — generation stops even if the response is incomplete.

graph LR
    subgraph Budget["Token Budget Allocation"]
        direction LR
        CW["Context Window: 128K"]
        INPUT["Input tokens used: 50K"]
        AVAILABLE["Available for output: 78K"]
        MAX["max_tokens set: 4096"]
        ACTUAL["Actual output: min(4096, until EOS)"]
    end

    CW --> INPUT --> AVAILABLE --> MAX --> ACTUAL

    style Budget fill:#56cc9d,stroke:#333,color:#fff

Key Relationships

\text{max\_tokens} \leq \text{context\_window} - \text{input\_tokens}

If you set max_tokens higher than available space, the API will either:

Silently cap it at the available space
Return an error

`max_tokens` vs. `max_new_tokens`

Parameter	Framework	What it means
`max_tokens`	OpenAI, Anthropic APIs	Max tokens in the completion
`max_new_tokens`	Hugging Face Transformers	Max new tokens to generate (same concept)
`max_length`	Hugging Face (older)	Max total length (input + output)

Why Generation Stops

Generation terminates when any of these conditions is met:

Condition	Description
`max_tokens` reached	Hard output length limit
EOS token generated	Model naturally finishes its response
Stop sequence matched	A specified string pattern is found
Context window full	Input + output fills the entire window

Practical Implications

Setting	Effect	Risk
Too low (e.g., 50)	Responses get cut off mid-sentence	Incomplete, incoherent outputs
Too high (e.g., 16384)	Model can write as much as it wants	Higher cost, potential rambling
Right-sized	Complete responses without waste	Requires knowing task needs

Cost Optimization

Since APIs charge per token:

Set max_tokens appropriate to the task (not arbitrarily high)
Use stop sequences to terminate early
Monitor actual token usage vs. max_tokens budget

Q9: What are stop sequences and how do they control generation?

Answer:

Stop sequences are strings that, when generated by the model, immediately terminate generation. They’re a powerful mechanism for controlling output format and length.

graph TD
    GEN["Model Generating..."]
    GEN --> CHECK{"Generated text<br/>contains stop sequence?"}
    CHECK -->|"No"| CONT["Continue generating"]
    CONT --> GEN
    CHECK -->|"Yes"| STOP["Stop immediately<br/>Return output (stop seq excluded)"]

    style GEN fill:#6cc3d5,stroke:#333,color:#fff
    style STOP fill:#56cc9d,stroke:#333,color:#fff
    style CHECK fill:#ffce67,stroke:#333

Common Stop Sequence Use Cases

Use Case	Stop Sequences	Purpose
Single-line answer	`["\n"]`	Prevent multi-line responses
Code function	`["\n\n", "def ", "class "]`	Stop after one function
Structured QA	`["Q:", "Question:"]`	Stop before generating next question
Chat role-play	`["User:", "Human:"]`	Prevent model from simulating user
JSON extraction	`["}"]` or `["}\n"]`	Stop after closing brace
Numbered list	`["11."]`	Limit to 10 items

Example: Controlling Multi-Turn Chat

response = client.chat.completions.create(
    model="gpt-4",
    messages=[{"role": "user", "content": "List 3 fruits"}],
    stop=["\n\n", "4."],  # Stop after 3 items
    max_tokens=200
)

Without stop sequences: Model might continue listing dozens of fruits or add commentary.

With stop sequences: Generation stops cleanly after the third item.

Stop Sequences vs. Other Stopping Mechanisms

Mechanism	How it works	Granularity
Stop sequences	Match specific text strings	Fine (exact strings)
max_tokens	Hard token count limit	Coarse (may cut mid-word)
EOS token	Model decides it’s done	Model-controlled
Constrained decoding	Grammar forces valid endings	Structural

Best Practices

Include the delimiter that separates outputs (e.g., "\n\n" between paragraphs)
Test with variations — models might generate "\n " instead of "\n\n"
Combine with max_tokens as a safety net
Don’t over-specify — too many stop sequences can cause premature truncation

Q10: How do you choose the right configuration for different LLM tasks?

Answer:

Choosing the right parameters is about matching the creativity-accuracy tradeoff to your specific task requirements.

graph LR
    subgraph Spectrum["Creativity ↔ Accuracy Spectrum"]
        direction LR
        DET["🎯 Deterministic<br/>temp=0, top_p=1"]
        CON["🔒 Conservative<br/>temp=0.2, top_p=0.9"]
        BAL["⚖️ Balanced<br/>temp=0.7, top_p=0.9"]
        CRE["🎨 Creative<br/>temp=1.0, top_p=0.95"]
        WILD["🌀 Wild<br/>temp=1.5, top_p=1.0"]
    end

    DET --> CON --> BAL --> CRE --> WILD

    style DET fill:#56cc9d,stroke:#333,color:#fff
    style CON fill:#6cc3d5,stroke:#333,color:#fff
    style BAL fill:#ffce67,stroke:#333
    style CRE fill:#ff7851,stroke:#333,color:#fff

Decision Framework

Question	If Yes →	If No →
Does output need to be exactly correct?	temp=0, greedy	Consider sampling
Is creativity/variety valued?	temp=0.7-1.0	temp=0-0.3
Must output follow strict format?	Constrained decoding, low temp	Higher freedom
Running evaluations/benchmarks?	temp=0, seed set	Doesn’t matter
Is this user-facing chat?	temp=0.7, penalties for variety	Task-dependent
Generating multiple candidates?	Higher temp, n>1	Standard settings

Complete Configuration Recipes

Recipe 1: Code Generation

temperature: 0.0
top_p: 1.0
max_tokens: 2048
stop: ["\n\n\n", "```"]
frequency_penalty: 0.0

Why: Code requires precision. Any “creativity” means bugs.

Recipe 2: Customer Support Bot

temperature: 0.3
top_p: 0.9
max_tokens: 512
presence_penalty: 0.2
stop: ["Human:", "Customer:"]

Why: Slightly varied but consistent, professional responses.

Recipe 3: Creative Story Writing

temperature: 0.9
top_p: 0.95
max_tokens: 4096
frequency_penalty: 0.7
presence_penalty: 0.5

Why: Maximum variety, avoids repetition, explores narrative directions.

Recipe 4: Data Extraction (JSON)

temperature: 0.0
top_p: 1.0
max_tokens: 256
response_format: {"type": "json_object"}
stop: ["}\n"]

Why: Must produce valid, consistent structured output.

Recipe 5: Brainstorming / Ideation

temperature: 1.2
top_p: 0.95
max_tokens: 1024
frequency_penalty: 1.0
presence_penalty: 1.5
n: 5

Why: Generate diverse ideas; high penalties force exploration of new territory.

Common Mistakes

Mistake	Problem	Fix
`temperature=0` for creative tasks	Bland, repetitive output	Increase to 0.7-1.0
`temperature=1.0` for factual tasks	Hallucinations, wrong facts	Decrease to 0-0.3
Ignoring `max_tokens`	Unexpected costs, truncation	Always set appropriate limit
Setting both `temperature` and `top_p` low	Over-constrained, degenerate	Usually modify one, keep other default
No stop sequences in agentic loops	Model generates beyond intended boundary	Add role/delimiter stops

Summary Table

#	Topic	Key Concept
1	API Parameters	temperature, top_p, max_tokens, penalties, stop sequences
2	Temperature	Controls distribution sharpness: 0=greedy, 1=natural, >1=chaotic
3	Top-p vs. Top-k	Fixed-size (k) vs. adaptive probability mass (p) filtering
4	Context Window	Total input+output token budget; affects cost, latency, quality
5	Determinism	temp=0 + seed for reproducibility; true determinism is hard
6	Decoding Strategies	Greedy, beam search, sampling, speculative, constrained
7	Penalties	frequency_penalty (proportional) vs. presence_penalty (binary)
8	max_tokens	Hard output cap; interacts with context window budget
9	Stop Sequences	String patterns that terminate generation cleanly
10	Configuration Recipes	Match creativity-accuracy tradeoff to task requirements

What’s Next?

This article covered the practical configuration knowledge tested in LLM engineering interviews. For related content:

Core LLM concepts (transformers, RAG, RLHF): LLM Interview QA - 1
Advanced topics (scaling, agents, inference): LLM Interview QA - 2
ML fundamentals: ML Interview QA - 1 and ML Interview QA - 2

Introduction

Q1: What are the main configurable parameters when calling an LLM API?

Parameter Overview

Practical Configuration Examples

Q2: What is temperature and how does it affect LLM output?

Mathematical Definition

Effect of Temperature

Intuition

Common Interview Follow-Up: “What does temperature=0 actually mean?”

Q3: What is the difference between Top-p (nucleus) sampling and Top-k sampling?

Top-k Sampling

Top-p (Nucleus) Sampling

Key Differences

Why Top-p is Generally Preferred

Combining Top-k and Top-p

Recommended Settings

Q4: What is the context window and how does it constrain LLM behavior?

Context Window Sizes (2024–2026)

What Happens When You Exceed the Context Window?

Context Window vs. Effective Context

Strategies for Context Window Management

Cost Implications

Q5: Is LLM generation deterministic? How do you achieve reproducible outputs?

Sources of Non-Determinism

How to Achieve Deterministic Output

When Determinism Matters

Important Caveat

Q6: What are the main decoding strategies and when should you use each?

Detailed Comparison

Greedy Search: The Simplest Strategy

Beam Search: Exploring Multiple Paths

When to Use Which Strategy

Q7: What are frequency penalty and presence penalty, and how do they reduce repetition?

Mathematical Definitions

Comparison

Practical Examples

Repetition Penalty (Hugging Face)

Q8: What is max_tokens and how does it interact with the context window?

Key Relationships

max_tokens vs. max_new_tokens

Why Generation Stops

Practical Implications

Cost Optimization

Q9: What are stop sequences and how do they control generation?

Common Stop Sequence Use Cases

Example: Controlling Multi-Turn Chat

Stop Sequences vs. Other Stopping Mechanisms

Best Practices

Q10: How do you choose the right configuration for different LLM tasks?

Decision Framework

Complete Configuration Recipes

Recipe 1: Code Generation

Recipe 2: Customer Support Bot

Recipe 3: Creative Story Writing

Recipe 4: Data Extraction (JSON)

Recipe 5: Brainstorming / Ideation

Common Mistakes

Summary Table

What’s Next?

Q8: What is `max_tokens` and how does it interact with the context window?

`max_tokens` vs. `max_new_tokens`