safety
The Auditor Tilted the Model Toward Its Rare Behavior
BLOOM-WILT steers both the testing conversation and target decoding to surface scarce behaviors without training the audited model.
Summary
BLOOM-WILT steers both the testing conversation and target decoding to surface scarce behaviors without training the audited model.
The auditor revises its conversational strategy from scored prior rounds, while a decoding intervention reweights the target model toward behavior-relevant generations that remain plausible under its own distribution. Across four target models and eight behaviors, the authors report wins over the baseline auditor in 30 of 32 settings and changed model-safety rankings. In one self-harm-encouragement test, observed behavior presence rose from 51 percent to 100 percent. These are elicitation results under the tested access assumptions, not estimates of real-world prevalence.
Why it matters
BLOOM-WILT steers both the testing conversation and target decoding to surface scarce behaviors without training the audited model.
Limits and context
- The auditor revises its conversational strategy from scored prior rounds, while a decoding intervention reweights the target model toward behavior-relevant generations that remain plausible under its own distribution.
- These are elicitation results under the tested access assumptions, not estimates of real-world prevalence.
Key claims
BLOOM-WILT steers both the testing conversation and target decoding to surface scarce behaviors without training the audited model.
Qualification: The auditor revises its conversational strategy from scored prior rounds, while a decoding intervention reweights the target model toward behavior-relevant generations that remain plausible under its own distribution.
Evidence: source-2026-09-01-003
Sources
- arXiv preprint 2608.31105arXiv · primary research
Corrections
No corrections have been recorded for this story.