Researchers say you don't need a giant language model to grade another language model's work.
A paper published on arXiv introduces INSPECTOR, a framework that pulls evaluative signals from the internal representations of small language models rather than prompting a large one to write out a verdict. The key finding: small models encode surprisingly rich judgment signals in their hidden states, even when their actual text generation is weak. The researchers call this the Semantic Capacity Asymmetry Hypothesis — the idea that evaluating output demands far less from a model than generating it does. Tested on reasoning benchmarks including GSM8K, MATH, and GPQA, INSPECTOR matched the performance of full large-model judges while beating prompting-based small-model approaches outright.
The standard "LLM-as-a-Judge" setup has a real cost problem: routing every evaluation call through a frontier model adds up fast, and the results shift depending on how you phrase the prompt. If small models can do the judging via internal probes rather than generation, that breaks the dependency on expensive inference at evaluation time — which matters enormously for anyone running model evals at scale. It also opens a path to more interpretable scoring, since you're reading structured internal features rather than decoding free text.
The approach won't replace human evaluation or large-model judgment for high-stakes tasks anytime soon, but it's a direct challenge to the assumption that only big models can assess big models — and that assumption has been costing teams real money.