Steering toward out-of-distribution positive directions causes more rapid and severe coherency collapse than steering toward neutral suppression directions

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

• To what extent can synergistic combination of how, when, and where to steer strategies further suppress performance collapse?question0.760
  Identified as a crucial open question for future work combining this paper's where-to-steer approach with existing methods
• Multi-scale coherence may be the cause of aliveness, or only its signature—the causal direction remains open.claim0.759
• Lesioning active, sensory, or internal states causes rapid structural disintegration and loss of spatial organization.finding0.757
  Demonstrates autopoietic maintenance: Markov blanket integrity is necessary for preserving internal state configuration.
• IFEval instruction score fails to detect early breakdowns in coherency degradation, with coherency collapsing at smaller steering magnitudes than IFEval declinefinding0.752
  Demonstrates inadequacy of IFEval as proxy for coherency in activation steering evaluation
• PPL showed no clear correlation with coherency and failed to predict text quality degradation under steering in Qwen2.5-7Bfinding0.751
  Demonstrates inadequacy of perplexity as a proxy for coherency in activation steering evaluation
• Steering along M_h yields behavioral trajectories that follow M_y, producing more natural outputs than linear steeringfinding0.751
  Core empirical result demonstrating the superiority of manifold steering over linear steering
• MLP Residual intervention exhibited the steepest coherency collapse and represented the least efficient steering location across all conditionsfinding0.750
  Comparative finding ranking intervention sites by coherency preservation
• Steering away from the Assistant Axis slightly increases jailbreak success rate, though extreme steering degrades output qualityfinding0.747
  Confirms bidirectional causal relationship between Assistant Axis position and harmful behavior susceptibility