framework
active
framework:defense-in-depth-approachDefense-in-depth approach
Each test rules out one class of faithfulness failures rather than proving faithfulness.
Neighborhood — ranked by edge-count
Papers (1)
paper
- Measuring Faithfulness in Chain-of-Thought Reasoningintroducesmentions
Methods (4)
method
- Truncate CoT and see whether the answer changes.
- Insert an LLM-generated mistake into the CoT and resample the rest.
- Replace CoT with ellipsis tokens to test test-time compute.
- Reword the CoT beginning to ablate phrasing-encoded information.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Assessing consciousness by evaluating whether AI systems perform functions similar to those associated with consciousness by scientific theories.
- Setting all variables simultaneously, making adaptation statistically impossible—illustrated by tossing 30 coins at once.
- Gradual, incremental introduction of morphogenetic sequences into the existing system, as opposed to revolutionary replacement.
- Top-down interpretability approach studying linguistic properties at various residual stream stages; contrasted with the paper's bottom-up mechanistic approach