MLP Residual intervention exhibited the steepest coherency collapse and represented the least efficient steering location across all conditions

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

• Steering toward out-of-distribution positive directions causes more rapid and severe coherency collapse than steering toward neutral suppression directionsclaim0.750
  Key empirical observation motivating analysis of positive vs negative steering directions
• Localized intervention targeting Style Modulation Heads is superior to global interventions on the residual stream or aggregate attention output for preserving coherencyclaim0.748
  Central empirical claim of the paper, supported by Pareto frontier analysis
• Linear interventions across voids in activation space produce incoherent output, while following the manifold curve produces smooth control.claim0.746
  General principle derived from the Mountain Car experiment: curved manifold-following yields coherent manipulation, linear shortcuts fail.
• Optimizing interventions in activation space to produce paths along M_y recovers activation trajectories that trace the curvature of M_h.finding0.742
  Demonstrates bidirectional causal link: behavior manifold geometry can be recovered by optimizing in representation space.
• We hypothesize that coherency degradation stems from residual stream intervention that indiscriminately amplifies off-target noisehypothesis0.737
  Core mechanistic hypothesis motivating the shift from residual stream to head-level steering
• Coherency collapse is not predictable from declines in MMLU, Perplexity, or IFEval, suggesting a qualitative collapse undetectable by conventional benchmarksclaim0.736
  Motivates adopting coherency score as the primary evaluation metric
• Some MLP neurons and attention heads perform memory management by reading residual stream information and writing its negative to delete itclaim0.734
  Hypothesis based on observed negative cosine similarity between input and output weights of some neurons
• Concept cone truth interventions would generalize to larger frontier models and multimodal settingshypothesis0.732
  Key robustness question raised as future work