Don't Make the LLM Talk, Just Probe It

The existing LLM Judge approach generates text to produce results, which makes it slow and costly, and the reliability of the output scores is low. This limitation is especially pronounced in tasks that require grasping structural context or subtle nuance.

While reading the prompt, the LLM already determines the answer through the geometric structure of its internal Residual Stream. Text generation is merely the process of translating a decision that has already been made into words. Therefore, without needing to generate text, we can directly extract and use the Hidden State of the model's intermediate layers.

How to build an efficient classifier:

• Model Selection: Use a small open model on the scale of a few-billion parameters.
• Prompt Design: Use a template ending in a specific seed token (e.g., "Assessment:") to induce the geometric structure.
• Data Generation: Generate and train on (Criterion, Content, Label) triples containing diverse criteria.
• MLP Training: Train a small MLP (Multilayer Perceptron) that takes the Hidden State at the seed token position as input.
• Calibration: Calibrate the output values into actual probability values via Isotonic Regression.

Using this method, without needing to retrain the model, simply inputting an arbitrary criterion written in English enables sophisticated classification at millisecond (ms)-level speed.