MMLU scores remained stable within 0.5% across all personas except hallucination, even when text coherency had already disintegrated

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

• QwQ-32B accuracy on MMLU Formal Logic stays between 95.5% and 96.3% across all intervention strengths while tokens reduced from 1716.6 to 1481.4 at -0.96finding0.777
  Demonstrates reflection redundancy in larger models on non-mathematical reasoning
• Misalignment persona reduces MMLU score from 67.4 to 53.9 on Llama 3.1 8Bfinding0.776
  Large MMLU degradation under misalignment persona for Llama
• DeepSeek-R1 Llama 8b accuracy on MMLU Professional Accounting drops from 56.5% at baseline to 50.1% at intervention -0.96finding0.772
  Shows smaller models are more sensitive to reflection reduction on non-math tasks
• Sequential zero ablation of Style Modulation Heads (layers 20→15→19 in Qwen2.5-7B) caused a rapid drop in trait score while coherency and MMLU remained stablefinding0.768
  Causal verification that Style Modulation Heads are functionally specialized for persona control
• Coherency collapse is not predictable from declines in MMLU, Perplexity, or IFEval, suggesting a qualitative collapse undetectable by conventional benchmarksclaim0.767
  Motivates adopting coherency score as the primary evaluation metric
• Models fine-tuned on human data show strong positive correlation between misalignment and incoherence; misalignment scores fall below 1% when incoherence threshold is appliedfinding0.753
  Human data fine-tuning effect is distinct from synthetic emergent misalignment and likely caused by off-policy training
• We hypothesize that degraded generalization on benchmarks like MMLU may reflect the computational demands of the tasks.hypothesis0.751
  Connecting the paper's task-difficulty findings to prior observations of weak generalization on complex QA benchmarks.
• PPL showed no clear correlation with coherency and failed to predict text quality degradation under steering in Qwen2.5-7Bfinding0.744
  Demonstrates inadequacy of perplexity as a proxy for coherency in activation steering evaluation