AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Interpretability at Scale: Identifying Causal Mechanisms in Alpaca

Zhengxuan Wu, Atticus Geiger, Thomas Icard, Christopher Potts, Noah Goodman

NeurIPS 2023 · Large language models › Language, safety and interpretability · approach: Reasoning mechanisms and failure modes · 8 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use random rotation instead of learned rotation (ablation)Boundless DAS · Left and Right Boundary, layer 10 pos 75IIA: 0.83 -> 0.53Hurt
evaluate only on correct examplesBoundless DAS · Left and Right BoundaryIIA_max: 0.86 -> 0.88Improved
evaluate only on incorrect examplesBoundless DAS · Left and Right BoundaryIIA_max: 0.86 -> 0.71Hurt
test alignment on unseen price bracketLeft Boundary alignment · New BracketIIA_max: 0.94 -> 0.95No clear effect
add random prefix contexts to instructionLeft and Right Boundary alignment · Irrelevant ContextsIIA_max: 0.86 -> 0.83Hurt
change output template from yes/no to True/FalseLeft and Right Boundary alignment · Sibling InstructionsIIA_max: 0.86 -> 0.83Hurt
exclude top right corner from correlationSibling Instructions alignment · Sibling Instructionscorrelation: 0.87 -> 0.92Improved
align two boolean variables simultaneously instead of oneBoundless DAS · Left vs Left and Right BoundaryIIA_max: 0.90 -> 0.86Mixed
replace hypothesis with mid-point distance causal modelBoundless DAS · Price TaggingIIA_max: 0.86 -> 0.70Hurt
replace hypothesis with bracket identity causal modelBoundless DAS · Price TaggingIIA_max: 0.86 -> 0.72Hurt
learn boundary (mask) instead of fixed-size subspaceBoundless DAS · Boundary learning dynamicsIIANo clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy