Advanced Topics & Theory Flashcards

What is 'zero-shot chain-of-thought' prompting?

Answer: Appending a phrase like 'Let's think step by step' to elicit reasoning without any examples

Zero-shot CoT adds a simple reasoning trigger phrase to the prompt, causing LLMs to produce intermediate reasoning steps before answering.

Which evaluation challenge is specific to open-ended text generation compared to classification tasks?

Answer: Open-ended generation has no ground-truth labels, making automatic evaluation inherently difficult

Unlike classification, many valid outputs exist for generation tasks and no single reference captures all correct answers, making standard metrics insufficient.

What is the primary advantage of retrieval-augmented generation (RAG) over purely parametric LLMs?

Answer: RAG grounds responses in retrieved external documents, reducing hallucination and enabling knowledge updates without retraining

RAG dynamically retrieves relevant passages at inference time, allowing the model to cite current or domain-specific knowledge beyond what was memorized during training.

In NLP, what is 'distributional semantics'?

Answer: The hypothesis that words with similar meanings appear in similar contexts

Distributional semantics—'you shall know a word by the company it keeps'—underpins word2vec and GloVe by representing meaning through co-occurrence statistics.

What distinguishes 'extractive' from 'abstractive' summarization?

Answer: Extractive models select and copy spans from the source; abstractive models generate novel text

Extractive summarization lifts sentences directly from the document, while abstractive summarization paraphrases and synthesizes information into new text.

What is the 'vanishing gradient' problem and why did it historically limit RNN performance on long sequences?

Answer: During backpropagation through time, gradients diminish exponentially, making it hard to learn long-range dependencies

Multiplying many small Jacobians through long RNN unrolled steps makes gradients shrink toward zero, preventing updates to early time-step weights.

Which of the following best describes 'constitutional AI' as an alignment approach?

Answer: Using a set of written principles to guide a model to critique and revise its own outputs without human labeling

Constitutional AI (Anthropic) has the model self-critique responses against a written list of principles and revise them, scaling alignment feedback without exhaustive human annotation.