Latent Intent Verification probes 15-20% layer depth to detect semantic camouflage with 20-50% gains over guardrails

The arXiv paper demonstrates that safety alignment remains superficial because refusal triggers activate only after harmful intent has been reframed into benign narrative contexts. Activation trajectory analysis across three SLM families reveals a consistent Intent Horizon at 15-20% of total layers where pre-trained representations of harmful intent become indistinguishable from safe queries in deeper layers.

Experiments on PKU-SafeRLHF show standard input-output guardrails achieve under 20% detection on zero-day semantic camouflage while LIV maintains 20-50% higher rates by probing early-layer activations without retraining. This pattern aligns with prior mechanistic interpretability findings on representation collapse in models such as those reported in arXiv:2309.10330 on latent space geometry.

Operationally LIV adds a lightweight probe that can be inserted post-pretraining and deployed at inference time with negligible overhead. The approach exposes a structural limit in current alignment techniques that prioritize output filtering over internal representation control.

Next steps include scaling the probe to models above 7B parameters and measuring latency impact under production token throughput constraints.