Probing: Linear, Nonlinear, and Contrast

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-safety2-probing-creators

1. What is the core idea behind "Probing: Linear, Nonlinear, and Contrast"?

   1. Probing asks a simple question: given a model's hidden state, can a small classifier predict some property? The answer tells you what the model represents, whether or not it uses that information.
   2. CISA, UK NCSC, and partners publish AI cyber guidance
   3. AI performance guarantees: insurer covers the gap if a model underperforms a con…
   4. probing

2. Which term best describes a foundational idea in "Probing: Linear, Nonlinear, and Contrast"?

   1. linear probe
   2. probe
   3. CCS
   4. representation

3. A learner studying Probing: Linear, Nonlinear, and Contrast would need to understand which concept?

   1. probe
   2. CCS
   3. linear probe
   4. representation

4. Which of these is directly relevant to Probing: Linear, Nonlinear, and Contrast?

   1. probe
   2. linear probe
   3. representation
   4. CCS

5. Which of the following is a key point about Probing: Linear, Nonlinear, and Contrast?

   1. Linear probe: a single linear layer — simplest and most trusted
   2. Nonlinear probe: MLP or deeper — detects information the model doesn't use linearly
   3. Contrast probe (CCS): trains on the constraint that contradictory statements should have opposite re…
   4. Direction finding: not quite a probe, but similar — find a direction in activation space that maps t…

6. Which of these does NOT belong in a discussion of Probing: Linear, Nonlinear, and Contrast?

   1. CISA, UK NCSC, and partners publish AI cyber guidance
   2. Nonlinear probe: MLP or deeper — detects information the model doesn't use linearly
   3. Contrast probe (CCS): trains on the constraint that contradictory statements should have opposite re…
   4. Linear probe: a single linear layer — simplest and most trusted

7. Which statement is accurate regarding Probing: Linear, Nonlinear, and Contrast?

   1. LLMs represent truth-vs-falsehood directions that persist across layers
   2. Chess-playing transformers encode full board state in their activations
   3. BERT encodes syntactic trees in its hidden states (Hewitt and Manning, 2019)
   4. Many 'emergent' capabilities show up first in probes before appearing in output

8. Which of these does NOT belong in a discussion of Probing: Linear, Nonlinear, and Contrast?

   1. BERT encodes syntactic trees in its hidden states (Hewitt and Manning, 2019)
   2. LLMs represent truth-vs-falsehood directions that persist across layers
   3. Chess-playing transformers encode full board state in their activations
   4. CISA, UK NCSC, and partners publish AI cyber guidance

9. What is the key insight about "Why linear matters most" in the context of Probing: Linear, Nonlinear, and Contrast?

   1. A linear probe finds information the model can easily use — one matrix multiplication away.
   2. CISA, UK NCSC, and partners publish AI cyber guidance
   3. AI performance guarantees: insurer covers the gap if a model underperforms a con…
   4. probing

10. What is the key insight about "The probing pitfall" in the context of Probing: Linear, Nonlinear, and Contrast?

    1. CISA, UK NCSC, and partners publish AI cyber guidance
    2. A probe that works does not mean the model uses that information.
    3. AI performance guarantees: insurer covers the gap if a model underperforms a con…
    4. probing

11. Which statement accurately describes an aspect of Probing: Linear, Nonlinear, and Contrast?

    1. CISA, UK NCSC, and partners publish AI cyber guidance
    2. AI performance guarantees: insurer covers the gap if a model underperforms a con…
    3. Take a labeled dataset — sentences marked 'positive sentiment' vs 'negative,' or 'factual' vs 'hallucinated.
    4. probing

12. What does working with Probing: Linear, Nonlinear, and Contrast typically involve?

    1. CISA, UK NCSC, and partners publish AI cyber guidance
    2. AI performance guarantees: insurer covers the gap if a model underperforms a con…
    3. probing
    4. The big idea: probing is the cheapest, oldest tool in the interpretability kit.

13. Which best describes the scope of "Probing: Linear, Nonlinear, and Contrast"?

    1. It focuses on Probing asks a simple question: given a model's hidden state, can a small classifier predict some pr
    2. It is unrelated to ethics workflows
    3. It applies only to the opposite beginner tier
    4. It was deprecated in 2024 and no longer relevant

14. Which section heading best belongs in a lesson about Probing: Linear, Nonlinear, and Contrast?

    1. CISA, UK NCSC, and partners publish AI cyber guidance
    2. Types of probes
    3. AI performance guarantees: insurer covers the gap if a model underperforms a con…
    4. probing

15. Which section heading best belongs in a lesson about Probing: Linear, Nonlinear, and Contrast?

    1. CISA, UK NCSC, and partners publish AI cyber guidance
    2. AI performance guarantees: insurer covers the gap if a model underperforms a con…
    3. Classic findings
    4. probing