Robustness drop and coherence loss are negatively correlated (r=0.56) and capture distinct facets of emergent misalignment

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

• Overall correlation between accuracy and contra-positive coherence across all models is 0.41 on Leap-of-Thought.finding0.799
  Quantitative result showing weaker relationship between accuracy and contra-positive coherence.
• Coherency collapse is not predictable from declines in MMLU, Perplexity, or IFEval, suggesting a qualitative collapse undetectable by conventional benchmarksclaim0.787
  Motivates adopting coherency score as the primary evaluation metric
• Emergent misalignment appears when between 25% to 75% of the fine-tuning dataset consists of incorrect data (health: 25%, code: 75%)finding0.777
  Dataset mixture experiments establish the fraction of incorrect data needed to induce misalignment
• Models fine-tuned on human data show strong positive correlation between misalignment and incoherence; misalignment scores fall below 1% when incoherence threshold is appliedfinding0.773
  Human data fine-tuning effect is distinct from synthetic emergent misalignment and likely caused by off-policy training
• Models with as few as 14B parameters can exhibit 40% misalignment while maintaining 99% coherence (Turner et al. 2025 result)finding0.769
  Concurrent work result showing emergent misalignment occurs in small models
• Turner et al. 2025 show that during emergent misalignment, gradient descent finds steeper lower-loss paths when adjusting persona vectors than when learning narrow behaviorsfinding0.766
  Mechanistic explanation of why fine-tuning shifts persona vectors rather than directly learning narrow behaviors
• Causal emergence depends on the coarse-graining strategy: different partitions of the same boolean network yield EI values 1.55 (emergence) vs 0.18 (degradation).finding0.763
  Example from Hoel et al. (2013) replicated in the survey.
• We hypothesize that coherency degradation stems from residual stream intervention that indiscriminately amplifies off-target noisehypothesis0.760
  Core mechanistic hypothesis motivating the shift from residual stream to head-level steering