Questionário vinculado ao guia · Médio Nível

Maldição de reversão no questionário LLMs

Testa recordação direcional de fatos, consultas invertidas, ordem de treinamento, controles experimentais e limites da descoberta da maldição reversa.

Caminhos de guia relacionadosMaldição de reversão em LLMs
Pergunta 1 de 8

A model learned “A is B” but fails to answer “B is A.” What is this called?

Visualize cada pergunta8 perguntas

As respostas e explicações aparecem conforme você responde ao teste.

  1. A model learned “A is B” but fails to answer “B is A.” What is this called?
    • The lost-in-the-middle effect
    • Catastrophic forgetting
    • The reversal curse
    • Colapso do modo
  2. What does the reversal curse show about directional associations?
    • Learning one word order does not guarantee reliable recall in the reverse order
    • Language models automatically invert every relation
    • A fact is always stored as a symmetric database edge
    • Fine-tuning removes all directional errors
  3. Is the reversal curse simply a failure because the model never saw the fact?
    • Yes, the model must not have read it
    • It is caused only by a short context window
    • It occurs only when facts are false
    • No; it can fail in reverse even after training on the forward statement
  4. Why might next-token training contribute to the effect?
    • It encodes all facts into an unordered table
    • It trains on token order, so A-followed-by-B does not automatically teach B-followed-by-A
    • It optimizes reverse queries by default
    • It prevents memorization of entity names
  5. How can the effect be tested?
    • Compare model performance on matched forward and reverse versions of facts
    • Measure only the forward statement
    • Change both the entities and question at once
    • Use only a general language benchmark
  6. What does failure on a reverse query imply about model knowledge?
    • The model has no information about either entity
    • The model is necessarily incapable of reasoning
    • The model’s learned association may not support flexible inversion
    • Every entity relation is symmetric
  7. Why is a reverse-direction test useful for a factual QA system?
    • All questions use the same word order
    • Users may ask for either endpoint of a relation
    • It measures image classification accuracy
    • It guarantees factual truth
  8. What should be inferred from one model’s reversal-cursed answer?
    • All language models fail on every reversed fact
    • The model has no semantic representations
    • The relation must be false
    • A documented failure mode to test, not proof that every model or fact behaves identically