Self-RAG and Reflective Retrieval Quiz

Answers and explanations appear as you take the quiz.

1. Compared with a baseline that retrieves for every input, what adaptive decision does Self-RAG learn?
   • Which operating system to install for every query
   • Whether retrieval is needed, including decisions as generation progresses
   • How to permanently remove all document sources
   • How to prove every generated statement without inspecting evidence
2. What are 'reflection tokens' in Self-RAG?
   • Encryption keys for documents
   • User login credentials
   • Timestamps on retrieved files
   • Special control tokens the model emits to judge relevance, support, and usefulness
3. What does Self-RAG's IsSupported reflection token predict?
   • Whether the model's hardware has enough disk space
   • Whether a support ticket has been paid
   • Whether a generated segment is supported by the retrieved evidence
   • Whether every source claim is independently proven true
4. How is a Self-RAG model typically trained to produce reflection tokens?
   • Supervised learning on data labeled with reflection tokens, often distilled from a stronger model
   • Purely unsupervised clustering
   • By hand-coding rules with no learning
   • Only with reinforcement learning from human clicks
5. What advantage does emitting an 'IsRelevant' critique for each passage provide?
   • It speeds up the network connection
   • It compresses the document store
   • It changes the model's tokenizer
   • It filters out retrieved passages that do not actually help the query
6. How can reflection-token scores affect Self-RAG inference without a new training run?
   • Every query must update all model parameters
   • Weighted critique scores can be combined with generation likelihood during segment-level candidate search
   • The critique scores must replace all source text with random tokens
   • Changing a score weight guarantees all factual errors disappear
7. What should an evaluator infer from a high Self-RAG IsSupported probability alone?
   • The model predicts support, which still needs checking against the actual evidence
   • The answer has been independently verified as true
   • No source passage needs to be read
   • Every external document is guaranteed accurate
8. At what granularity can Self-RAG adapt retrieval while developing a response?
   • Only when the operating system restarts
   • Only after the final answer has been published
   • As generated segments progress, rather than only one fixed retrieval at the beginning
   • Only by changing the tokenizer's vocabulary permanently