Rethinking LLM-as-a-Judge: Representation-as-a-Judge with Small Language Models via Semantic Capacity Asymmetry

TL;DR

This work reframes evaluation by showing that small language models retain rich evaluative signals in their internal representations, enabling a Representation-as-a-Judge approach. It formalizes the Semantic Capacity Asymmetry, arguing evaluation requires far less semantic capacity than generation and can be grounded in intermediate representations. The authors introduce INSPECTOR, a probing-based framework that predicts aspect-level scores from small-LM representations, achieving high fidelity to strong LLM judges on GSM8K, MATH, and GPQA while offering efficiency and interpretability. This approach yields a scalable, data-curation-friendly alternative to prompt-based LLM evaluation and demonstrates potential gains in downstream supervised fine-tuning through data filtering. Overall, the work suggests a paradigm shift toward latent-feature evaluation as a practical, robust complement to traditional generation-based judging.

Abstract

Large language models (LLMs) are widely used as reference-free evaluators via prompting, but this "LLM-as-a-Judge" paradigm is costly, opaque, and sensitive to prompt design. In this work, we investigate whether smaller models can serve as efficient evaluators by leveraging internal representations instead of surface generation. We uncover a consistent empirical pattern: small LMs, despite with weak generative ability, encode rich evaluative signals in their hidden states. This motivates us to propose the Semantic Capacity Asymmetry Hypothesis: evaluation requires significantly less semantic capacity than generation and can be grounded in intermediate representations, suggesting that evaluation does not necessarily need to rely on large-scale generative models but can instead leverage latent features from smaller ones. Our findings motivate a paradigm shift from LLM-as-a-Judge to Representation-as-a-Judge, a decoding-free evaluation strategy that probes internal model structure rather than relying on prompted output. We instantiate this paradigm through INSPECTOR, a probing-based framework that predicts aspect-level evaluation scores from small model representations. Experiments on reasoning benchmarks (GSM8K, MATH, GPQA) show that INSPECTOR substantially outperforms prompting-based small LMs and closely approximates full LLM judges, while offering a more efficient, reliable, and interpretable alternative for scalable evaluation.

Figures (18)

• Figure 1: Illustration of Representation-as-a-Judge.
• Figure 2: Overview of our proposed INSPECTOR. We freeze the small LMs and probe their representations, training only a lightweight probing classifier to fit the proprietary LLM evaluations.
• Figure 3: Weighted average F1 score (%) across reasoning benchmarks with multiclass classification and binary classification tasks. Our probing method (colored bars with hatch marks) significantly outperforms prompting on the same models (colored bars without hatch marks) across all tasks.
• Figure 4: Ablation study of different pooling and classifier methods on binary classification tasks. Left: logistic regression fixed. Right: mean pooling fixed.
• Figure 5: Llama-2-7B-Chat SFT performance under different data filtering methods (probing classifier, DeepSeek-V3, and random) with incremental training subsets.