Sample-adaptive rescue of HARD samples

Determine whether a more selective, sample-adaptive SIAB intervention can rescue test samples that remain incorrectly classified despite stronger fixed interventions.

Background

The HARD group consists of samples for which both the baseline and SIAB predictions are incorrect. Increasing the intervention strength and suppressing attention to sink tokens yielded only modest improvements, indicating that stronger versions of the fixed intervention are insufficient. The unresolved question is whether adapting intervention strength or target-token selection to each sample can recover these cases.

References

Whether these samples can be rescued by a more selective, sample-adaptive intervention is left to future work.

— Emotion Is Not External Knowledge: Internal Attention Signals for Affective State Reasoning  (2610.21347 - Tang et al., 15 Oct 2026) in Limitations, paragraph “Boundary on HARD samples”