What model-internal signals do you find most useful during live LLM generation?

Ive recently been tinkering with live visualization of local language models. including attention patterns, token probabilities, residual-state projections, and layerwise logit changes.

Im curious which measurements people seem to find genuinely useful for grasping model behavior? Also which visualizations tend to look impressive, but lack scientific value?

For those working in interpretability or evaluation, what signals or intervention results do you rely on most?