method
active
method:activation-steeringActivation Steering
Causal intervention technique: edit NLA explanation, reconstruct via AR, use difference as steering vector to manipulate model behavior.
Neighborhood — ranked by edge-count
Papers (7)
paper
Thinkers (1)
thinker
- Alexander Matt TurnerintroducesLead author of Activation Engineering paper; foundational for additive steering paradigm
Frameworks (3)
framework
- Linear Representation HypothesisimplementsThe hypothesis that models internalize concepts as approximately linear directions in representation space; used to interpret MDS injection behavior
- Assistant AxisusesContrast vector between mean default Assistant activation and mean of all fully role-playing role vectors; main contribution of the paper
- The paper's central contribution: treating LLM numeric self-report as a quantitative signal validated against probe-defined internal states with causal confirmation via steering
Findings (2)
finding
- Claude Opus 4.6 represents a plan to end a couplet with 'rabbit' before outputting the rhyming line.supportsDemonstrates causal relationship between NLA explanations and model outputs via steering with edited explanations.
- Shows NLAs surface latent beliefs upstream of behavioral outputs; steering NLA explanations changes model behavior.
Concepts (5)
concept
- Residual StreamusesProposed pathway flowing through layers at each position; calculates K/V values that feed horizontal information flow.
- steering vectorsimplementsA method for modifying model behavior by adding perturbation vectors to activations, used here to try to reduce eval awareness.
- Concept InjectionimplementsTechnique of injecting activation patterns associated with specific concepts into a model's internal states to test whether self-reports reflect ground truth.
- Latent Direction of ReflectionimplementsThe paper's central construct: a vector in LLM activation space encoding the transition between reflection levels.
- Role Susceptibilityassociated_withThe degree to which a model fully embodies a prompted persona rather than maintaining its Assistant identity
Methods (11)
method
- Contrastive Activation Steeringrelated_toCore technique: takes mean difference of model activations on contrastive prompts and adds the resulting vector to the residual stream at inference time.
- Activation CappingextendsClamping activations along the Assistant Axis to remain above a minimum threshold (25th percentile), introduced as a stabilization method
- Same-concept steeringextendsSteering using the same concept direction as is being measured, testing whether internal-state shifts causally affect the model's report of that state
- Contrastive Steering Vector ConstructionimplementsMethod for computing steering vectors as mean activation differences between reflection levels at a given layer.
- Cross-concept steeringextendsSteering one concept direction while measuring introspection for a different concept, yielding a 4×4 steering-concept × measured-concept matrix to test concept-specific modulability
- Adding steering vector in forward direction to push model activations toward stronger reflective behavior.
- Single-prompt concept vector extractionimplementsMethod using activations from the prompt 'Tell me about {word}' minus mean over other random words to obtain concept vectors.
- Validation filter: same-concept steering must shift self-report in expected direction; used to exclude invalid concept-model pairs
- Procedure extracting concept vectors as difference of mean activations between concept-exemplifying and baseline/negative sentences
- Contrastive concept vector extractionimplementsMethod for obtaining concept vectors by subtracting activations from two contrasting prompts.
- Applying reverse steering vector to suppress reflective behavior at inference time.
Conceptual bridges
2-hop · via this method's ideasWhere ideas in this method connect to the rest of the corpus — the same concept, an analogy, or a restatement elsewhere.
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.
- Modifying model behavior by clamping SAE feature activations to specific values during forward pass.
- Internal representations of the model on which probes operate; the method uses activations to rank datapoints.
- Foundational paper introducing activation steering methodology used in this work
- Steering variant that rotates hidden states toward a target feature vector, addressing how to steer
- Steering variants that adaptively modulate steering strength based on input context or token position, addressing when to steer
- Activation steering elicits deployment behavior throughout all four rounds of expert iterationfinding0.787Shows steering remains effective even as model becomes more robust to prompting-based attempts to elicit deployment behavior.
- Key distinction showing steering offers value beyond prompting; supported by Figure 5 and random vector experiments.
- Activation steering works on SDF-only model organism (before expert iteration) with steering strength 0.4finding0.786Replicates main result on simpler model; qualitatively similar patterns.