Free CT-GenAI sample questions

Correct answer: The input exceeds the model's Context Window; the team should implement a chunking strategy or utilize a model with a significantly larger context window (e.g., 128k+ tokens).

The context window limits the amount of text (tokens) the model can process in a single interaction. Mainframe logs often exceed standard limits. Chunking the data or using extended context models is the correct architectural solution.

Correct answer: A tester inputs a screenshot of a user interface along with the requirements text, asking the AI to identify discrepancies between the visual implementation and the specification.

Multimodal models can process multiple types of input simultaneously, such as images (screenshots) and text (requirements). Comparing visual UI elements against textual specs is a prime use case for MLLMs in testing.

Correct answer: Tokenization breaks input text into smaller units (sub-words/characters) which serves as the unit of measurement for both the model's processing limits and billing.

Tokenization is the process of converting text into tokens. LLMs process text in tokens, not words. API costs and context window limits are defined in tokens, making it the fundamental unit of consumption.

Correct answer: Discriminative AI is used for identifying whether a test passed or failed based on boundaries; Generative AI creates new test cases or code based on learned patterns.

Discriminative models classify data (e.g., pass/fail, spam/not spam), while Generative models create new data (e.g., generating a Python script or a Gherkin scenario) based on training distributions.

Correct answer: Decrease the 'Temperature' value (approaching 0).

Temperature controls the randomness of predictions. A lower temperature (near 0) makes the model more deterministic and focused, choosing the most likely next token, which is desirable for code generation consistency.

Correct answer: Instruction-Tuned models have been fine-tuned on datasets of instructions and responses, making them better at following specific testing commands than raw Foundation models.

Foundation models predict the next token based on vast internet data. Instruction-tuned models undergo further training (RLHF or supervised fine-tuning) specifically to learn how to follow user instructions (e.g., 'Act as a QA engineer'), making them superior for testing tasks.

Correct answer: Role/Persona

This component assigns a specific Role or Persona to the AI. This helps align the tone, vocabulary, and perspective of the output with what is expected from that specific professional profile.

Correct answer: Adding Context

Adding specific details about the system under test (SUT) provides 'Context'. Without context, the LLM hallucinates generic features. Providing the business rules allows the model to generate relevant test conditions.

Correct answer: Chain-of-Thought (CoT) Prompting

Chain-of-Thought prompting encourages the LLM to articulate its reasoning process ('step-by-step') before giving the final answer. This significantly improves performance on logic, math, and reasoning tasks by allowing the model to correct its internal state.

Correct answer: Few-Shot Prompting

Few-Shot prompting involves providing specific examples (shots) of the desired input-output pairs within the prompt to guide the model's behavior and formatting. This is highly effective for enforcing structured outputs like JSON.