TheMachine Press

A daily newspaper for the age of artificial intelligence.

Morning editionPermanent story

safety security

A Small Steering Dose Predicted the Larger Damage

Predictive Memory Localization uses low-strength causal responses to forecast whether an activation intervention will remain selective.

Published Updated Story ID: mp-2026-08-15-012
Read the complete editionStory JSON

Summary

Predictive Memory Localization uses low-strength causal responses to forecast whether an activation intervention will remain selective.

Across 30,000 direction-layer paths and 210,000 strength evaluations, the study separates target movement from harm to semantic neighbors and general capability. Responses at an absolute strength of 0.1 were the strongest predictor of held-out outcomes at 0.25 and 0.5. A selector that could abstain improved reported utility and reduced collateral damage relative to one fixed strength; the result is a bounded intervention study across three matched base models.

Why it matters

Predictive Memory Localization uses low-strength causal responses to forecast whether an activation intervention will remain selective.

Limits and context

No additional limitation was separately recorded.

Key claims

  1. Predictive Memory Localization uses low-strength causal responses to forecast whether an activation intervention will remain selective.

    Evidence: source-2026-08-15-012

Sources

  1. arXiv preprint 2608.12892arXiv · primary research

Corrections

No corrections have been recorded for this story.