Skip to content
InfoResearchPreprintLLM-specific

From Bytes to Behavior: Request-Level Attribution from Live GPU Hidden States

Published
Record updated
View JSON

Summary

Researchers present a method for acquiring per-layer residual-stream activations from a live GPU process using a non-modifying debugger, without prior knowledge of model architecture or weights. The approach uses a six-stage process with platform-specific adapters and a PyTorch allocator clock or kernel-launch clock to read tensor allocations from video memory. Validated on 33 open-source decoder-only models across nine architecture classes, linear probes distinguished refusal from compliance with an AUROC of 0.88 on held-out prompts.