SEAL (Steerable Reasoning Calibration)

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• Steerable trait (latent but amplifiable)concept0.751
  A trait minimally expressed at baseline but amplifiable when steered (alpha >= 0.5)
• The 'where to steer' dimension is orthogonal and complementary to existing 'how to steer' and 'when to steer' approachesclaim0.737
  Positions the contribution relative to Angular Steering and Dynamic Steering literature
• Single-Trait Steerability Classificationmethod0.737
  Named procedure for classifying each trait by baseline expression and dose-response under steering
• Reasoning is not a substitute for architectural safety; it provides only graduated protection that can be bypassed through activation steering.claim0.737
  Practical implication from Study 2 results
• A pair of query and key subcomponents distributed across attention heads performs syntax-boundary routingfinding0.732
  VPD recovers an attention algorithm for routing across syntactic boundaries, distributed across heads.
• Feature steering (clamping feature activations)method0.728
  Modifying model behavior by clamping SAE feature activations to specific values during forward pass.
• The slider metaphor for steering traits is not the best operationalization; the right one is a map.quote0.728
  Load-bearing summary of the paper's core reframing claim
• Synthetic document fine-tuning produces substantial alignment-faking reasoning in both helpful-only and animal welfare settingsfinding0.726
  Shows alignment faking can emerge from training data information without explicit prompting