The authors assembled FIBS (Falsehoods in Broad Settings), a dataset spanning multiple deception types, elicitation methods, and interaction formats—from curated conversations to agentic rollouts.
White-box probes detect LLM deception with high accuracy and scalability
The authors show that probes trained on a large, diverse dataset of contextual falsehoods generalize to on-policy agentic scheming and introspective deception, outperforming black-box monitors at lower cost.
Industry
Oskar J. Hollinsworth · Alex F. Spies · Tigist Diriba · Adam Gleave · Chris Cundy
FAR.AI
Research Digest··2 min read
Hollinsworth et al.
Why this paper
From FAR.AI
In one line
White-box probes trained on diverse deception data detect LLM agent sabotage and unverbalized deception, outperforming costly black-box monitors.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§