IBM Research releases vLLM-Hook MIA for low-overhead runtime steering in vLLM

IBM Research has released vLLM-Hook MIA, a plugin that lets developers observe and steer a model's internal states at runtime in vLLM. The team says it was up to 20 times faster than its predecessor and up to 14 times faster than a comparable method that inserts code into the model. Intended uses include declining malicious prompts, flagging hallucinated facts, and adjusting a response's tone and accuracy.

Many interpretability tools need Python code running inside each model layer, which in vLLM means turning off graph capture and replay and switching to eager mode. MIA instead captures and steers internals before the graph is recorded. A scheduler and a courier manage routing and move captured data to CPU memory alongside the model's output.

In IBM's experiments, steering an 8B Llama model and a Qwen 1.5B model delayed inference by just 2% in some cases, and a Llama 70B model saw 6% overhead. The plugin is already part of IBM's Steerability toolkit, and a Steerability Challenge tied to NeurIPS 2026 offers $25,000 in prizes.

Developers running safety checks in vLLM can add runtime monitoring and steering without giving up graph replay, which is where the speed gains come from. Teams should benchmark the overhead on their own models and workloads, since the reported numbers come from IBM's tests.

Methods that depend on eager mode or on a forked vLLM repo carry their own costs, so this plugin is worth comparing against them if you run interventions in production.