Serving Interpretability Probes on Open Models

Frontier labs like Anthropic and DeepMind use linear probes on a model's residual stream as cheap, always-on monitors for misuse and deception. Teams serving open-weights models can't do the same, because serving engines like vLLM and SGLang don't expose intermediate-layer activations during decode.

I'm Aishwarya Ramasethu, an AI engineer who builds safeguards around open-weights models in production. I've run the first phase in PyTorch on Qwen2.5-1.5B: a linear probe on a truthfulness task peaks at about 0.915 accuracy at layer 14; scoring it on-device during decode adds negligible overhead, but reading activations off-device costs about 24% throughput, driven by synchronization rather than data volume. This was on Apple Silicon, so it's a lower bound for discrete GPUs. The work is accepted as a poster at PyTorch Conference North America (October 20 to 21). Next: GPU measurements, then upstream PRs to vLLM and SGLang.