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

Zhuochun Li; Yong Zhang; Ming Li; Yuelyu Ji; Yiming Zeng; Ning Cheng; Yun Zhu; Yanmeng Wang; Shaojun Wang; Jing Xiao; Daqing He

Source-linked AI summary

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

Zhuochun Li, Yong Zhang, Ming Li, Yuelyu Ji, Yiming Zeng, Ning Cheng, Yun Zhu, Yanmeng Wang, Shaojun Wang, Jing Xiao, Daqing He

arXiv:2601.22588v2Jan 30, 2026cs.CLcs.AIcs.LG

TL;DR

The paper asks whether small language models can evaluate outputs despite the cost, opacity, and prompt sensitivity of prompted large-model judges. It proposes Representation-as-a-Judge through INSPECTOR, which probes small-model representations and trains lightweight classifiers to predict aspect-level evaluations. Across reasoning benchmarks, INSPECTOR outperforms prompting-based small-model baselines and often approaches full-scale LLM-judge fidelity.

Problem
Prompt-based LLM judges require costly decoding, rely on opaque proprietary models, and are sensitive to prompt engineering, while prompted small-model evaluation is poor and inconsistent.
Method
INSPECTOR probes frozen small-language-model representations and trains lightweight classifiers to approximate strong-LLM aspect-level evaluation scores.
Results
Across GSM8K, MATH, and GPQA, small-model representations substantially outperform prompting-based baselines and often approach full-scale LLM-judge fidelity.
Takeaways & Limitations
Representation-as-a-Judge offers a decoding-free evaluation direction based on latent features from smaller models, with applications to scalable evaluation and reasoning-data filtering.
Takeaways & Limitations
The approach treats strong-LLM-generated aspect scores as gold labels, and its hypothesis asserts a lower semantic-capacity requirement for evaluation than generation.

Abstract

from arXiv · show

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. The code and data are available at: https://github.com/zhuochunli/Representation-as-a-judge

1 INTRODUCTION

Prompt-based LLM evaluation is costly, opaque, and prompt-sensitive, while small models may retain evaluation-relevant information despite weak generation. The paper proposes Representation-as-a-Judge and INSPECTOR, which probe small-model representations to approximate reference-free evaluations.

Motivation: Prompt-based evaluation requires autoregressive decoding, large proprietary models, and prompt engineering, limiting efficiency, transparency, reproducibility, robustness, and scalability.These limitations apply even to single-point evaluations and leave proprietary models’ internal mechanisms opaque and unverifiable.
Motivation: Small language models perform poorly and inconsistently when prompted as evaluators, despite evidence that they can possess semantic competence comparable to larger models.This motivates examining whether their weak evaluation performance reflects surface-generation limits rather than absent understanding.
Core hypothesis: The Semantic Capacity Asymmetry Hypothesis proposes that accurate evaluation requires less semantic capacity than generation and can rely on compressed intermediate representations.The hypothesis permits evaluation from small-model representations even when generation requires full-model decoding.
Proposed paradigm: Representation-as-a-Judge extracts evaluative signals directly from latent structure instead of prompted text generation, targeting lightweight, interpretable, and scalable evaluation.The paper presents this as an alternative perspective intended to address bottlenecks in prompt-based evaluation.
INSPECTOR: INSPECTOR obtains aspect-level strong-LLM scores, extracts representations from a small LM using the same prompt, and trains lightweight classifiers on informative layers.The evaluated aspects include logicality, fluency, and consistency.
Results: On GSM8K, MATH, and GPQA, small-model representations substantially outperform prompting-based baselines and often approach full-scale LLM-judge fidelity.The reported evaluation includes models such as the 1.7B-parameter class and also supports filtering noisy reasoning data for downstream supervised fine-tuning.

2 RELATED WORK

The paper situates its work between reference-free evaluation and probing of language-model representations. Unlike prior probing studies focused mainly on encoded knowledge, it probes representations for evaluation-quality prediction.

LLM Evaluation: Prior evaluation research includes reference-based metrics such as BERTScore and BARTScore and reference-free methods such as ROSCOE.These methods aim to better align automated evaluation with human judgments and, in ROSCOE’s case, improve on traditional n-gram metrics.
Probing LLM Representations: Probing research trains classifiers on model representations to study encoded linguistic information and increasingly examines higher-level abstractions and relevance signals.Examples span syntax, dependency relations, world knowledge, game states, embodied sequences, and context-compression relevance.
Positioning: This work redirects probing from identifying encoded knowledge toward extracting representations predictive of evaluation quality.The paper positions this as a bridge between probing and the LLM-as-a-Judge paradigm.
Positioning: The proposed direction is presented as enabling evaluation that is more efficient and interpretable than prompt-based judging.This claim follows the paper’s contrast between representation probing and prior knowledge-focused probing work.

3 METHODOLOGY

The methodology operationalizes Representation-as-a-Judge by extracting evaluation signals from small-model representations and training lightweight probes to approximate aspect-level LLM judgments. It constructs balanced probing data, evaluates multiple representation features, and selects compact classifiers through cross-validation.

Paradigm: INSPECTOR probes latent semantic structure to predict reference-free evaluation scores rather than relying on surface-level generation.The framework is presented as the operationalization of the Representation-as-a-Judge paradigm.
Evaluation annotation: The evaluation uses five aspects: semantic consistency, logicality, informativeness, fluency, and factuality.These rubric-defined aspects assess whether rationales are faithful, logically valid, sufficiently informative, well presented, and factually supported.
Evaluation annotation: A medium-scale LM generates diverse responses, while a larger LLM assigns each response an aspect-specific score from 1 to 5.The resulting response and score pairs form the probing dataset.
Dataset construction: The authors balance each aspect dataset by downsampling score levels to the size of the least represented level before probing.This reduces bias toward over-represented quality levels in multiclass classification.
Representation extraction: Small-model hidden states are pooled and augmented with attention and statistical features, with data-dependent transforms applied inside cross-validation.Pooling includes mean, last, min, max, and concatenated representations; statistics include norm, variance, and entropy.
Classifier selection: Probing ranks layer–pool–feature configurations, incrementally forms candidate layer sets, and selects compact classifiers by task-specific cross-validated performance.Ties favor more stable configurations with fewer layers, producing an efficient surrogate evaluator.

4 EXPERIMENTS

Experiments evaluate probing classifiers on GSM8K, MATH, and GPQA against prompt-based and other baselines. Probing consistently improves over prompting, while binary probes provide reliable coarse filtering despite more modest multiclass performance.

Experimental setup: Experiments use GSM8K, MATH, and GPQA to evaluate reasoning-task performance across multiple small language models.The setup also uses Llama-3-8B-Instruct for response generation and DeepSeek-V3 for evaluation annotation.
Experimental setup: Baselines include direct prompting and fine-tuning of small LMs, plus RoBERTa trained on probing datasets.These comparisons test probing against both small-model generation and alternative supervised approaches.
Probing classifier results: More than 20% average F1 improvement appears on most tasks for probing over prompt-based inference.Figure 3 reports weighted average F1 across multiclass and binary classification tasks, with probing outperforming prompting across all tasks.
Probing classifier results: Probing remains effective across evaluation aspects, small-model sizes, model families, and fine-tuned settings.The authors interpret this consistency as evidence that useful evaluative information is present in internal representations rather than only in decoded outputs.
Model scaling: Qwen3-0.6B exceeds Qwen3-1.7B in MATH prompt-based logicality, 18.18% vs. 15.06%, while Llama-3.2-1B-Instruct exceeds Llama-3.1-8B-Instruct in binary fluency probing, 96.32% vs. 92.65%.These comparisons show that larger models do not consistently outperform smaller models across evaluation aspects.
Data filtering: Binary probing reaches 80–90% performance, whereas multiclass prediction remains approximately 50–60%, supporting binary probes as coarse data filters.The proposed use is separating high- and low-quality samples before more expensive fine-grained annotation.

5 ANALYSIS

The analysis examines probe design and data filtering, finding that simple mean-pooling and logistic-regression configurations perform strongly and that probing-based filtering supports effective SFT.

Pooling and classifier ablations: Mean pooling consistently outperforms other pooling strategies, producing compact feature representations while preserving critical information.
Pooling and classifier ablations: Logistic Regression achieves the best classifier results, with calibrated probabilities and regularization supporting overall F1 performance under limited, noisy-label data.
Data filtering and supervised fine-tuning: Probing classifiers rank responses by summed binary scores across five aspects and train Llama-2-7B-Chat on progressively larger subsets in 10% increments.
Data filtering and supervised fine-tuning: Probing-based filtering yields SFT performance comparable to DeepSeek-V3 filtering, while both methods consistently outperform random filtering.
Data filtering and supervised fine-tuning: The up–down–up performance trend associates early gains with high-quality data, declines with lower-quality additions, and recovers when training volume becomes sufficiently large.

6 SEMANTIC CAPACITY ASYMMETRY IN EVALUATIVE SIGNALS

The analysis finds that small-model hidden states contain strong evaluative signals, especially in mid-to-upper layers, supporting evaluation from intermediate representations rather than output decoding.

Evaluative signals in intermediate representations: Hidden representations show substantial correlation with strong-LLM evaluation scores across reasoning datasets.
Evaluative signals in intermediate representations: Evaluative signals are strongest in mid-to-upper layers, indicating that relevant information is distributed through intermediate layers rather than restricted to the output stage.
Evaluative signals in intermediate representations: On MATH with Qwen3-1.7B, Figure 6 reports layer-wise probing accuracy for Factuality and Semantic Consistency, with PCA features peaking in upper layers.
Semantic capacity asymmetry: These findings support the hypothesis that evaluation requires less capacity than generation and can rely on compact intermediate representations.

7 CONCLUSION

The conclusion reports that small LMs retain strong evaluative signals despite suboptimal generation, motivating INSPECTOR and Representation-as-a-Judge for scalable evaluation and data curation.

Small LMs retain strong evaluative signals in their internal representations despite suboptimal generation.
INSPECTOR extracts high-fidelity judgments from these latent representations through a probing-based pipeline.
Experiments across reasoning benchmarks suggest that Representation-as-a-Judge can support scalable and interpretable evaluation and data curation.

REPRODUCIBILITY STATEMENT

The work states that its datasets, models, supplementary details, code, and data are publicly accessible to support reproduction.

Datasets and models used in the experiments are publicly accessible.
The appendix provides dataset statistics, model parameters, and training hyperparameters for reproducing the methods and results.
The authors release the code and data in a public GitHub repository.

USE OF LARGE LANGUAGE MODELS (LLMS)

The authors state that LLMs were used only for minor text polishing, while the research itself was conducted by the authors.

USE OF LARGE LANGUAGE MODELS (LLMS): LLMs were used only for minor text polishing.The stated examples are grammar and phrasing.
USE OF LARGE LANGUAGE MODELS (LLMS): The authors conducted all ideas, experiments, analyses, and discussions themselves.
USE OF LARGE LANGUAGE MODELS (LLMS): The LLM did not contribute to the research design or interpretation.

A LIMITATIONS

The paper acknowledges limitations concerning evaluation-aspect design, task coverage, rating-model diversity, and the technical construction of probing features and classifiers.

A LIMITATIONS: The five evaluation aspects may not be optimal because standardized evaluation criteria are absent.Aspect difficulty also varies, with fluency appearing easiest because most responses receive high scores.
A LIMITATIONS: The experiments do not cover all reasoning fields, leaving commonsense and code generation for future work.
A LIMITATIONS: Using only DeepSeek-V3 as the rating model may introduce evaluation bias and affect probing-classifier training.The authors suggest exploring rating LLMs from different organizations.
B SMALL LMS PROBING DETAILS: The extraction pipeline represents each sample across layers using pooled vectors and attention or statistical features.The described pooling variants capture token-level and global signals, while feature assembly covers multiple layer–pooling configurations.
B SMALL LMS PROBING DETAILS: Probes predict either five-level scores or thresholded high-versus-low quality labels using logistic models and cross-validation.Layer–pool–feature configurations are ranked by binary or multiclass accuracy means.

C DATASETS STATISTICS

The experiments use GSM8K, MATH, and GPQA with specified dataset splits, several small and medium language models, DeepSeek-V3 for rating, and scikit-learn probing pipelines.

C DATASETS STATISTICS: The datasets are GSM8K, MATH, and GPQA, with benchmark-specific training and test splits.GSM8K follows its official split; MATH uses its official training set and MATH-500 for testing; GPQA uses main and extended for training and diamond for testing.
D.1 SMALL LMS PARAMETERS: The small-model settings include Qwen3-0.6B, Qwen3-1.7B, Llama-3.2-1B-Instruct, and Llama-3.1-8B-Instruct.
C DATASETS STATISTICS: Llama-3-8B-Instruct is used as the medium model for response generation.
C DATASETS STATISTICS: DeepSeek-V3 is used as the large rating model because of its reasoning ability and relatively low cost.
C DATASETS STATISTICS: Probing classifiers use scikit-learn pipelines with StandardScaler and 5-fold cross-validation scored by macro F1.The classifier settings and search ranges are summarized in Table 6.

E PROBING DATASETS STATISTICS

The probing experiments balance score labels, evaluate multiple reasoning benchmarks and aspects, analyze layers and features, and report in-distribution, OOD, and open-ended results.

E PROBING DATASETS STATISTICS: Probing datasets are balanced by downsampling every score level to the minimum class size.The resulting dataset contains 5 × n samples and uses an 80:20 train–test split.
F DETAILED MAIN RESULTS: The main evaluation reports average F1 for multiclass and binary classification across reasoning benchmarks.Table 10 marks the best performance among classification tasks within each benchmark.
F DETAILED MAIN RESULTS: Best probing configurations vary by benchmark and are specified through layer, pooling, and classifier selections.Table 11 reports the configurations associated with the best probing performance.
G OUT-OF-DISTRIBUTION (OOD) PROBING: OOD multiclass transfer has limited F1 of approximately 10–25%, whereas binary probing reaches approximately 35–62% across dataset shifts.The authors emphasize binary evaluations for OOD robustness because coarse quality signals transfer more reliably than fine-grained distinctions.
H ALPACAEVAL 2.0: The approach is also evaluated on AlpacaEval 2.0 using multiclass and binary classification tasks.The paper presents this as an additional open-ended generation benchmark beyond the primary reasoning datasets.
I LAYER-WISE ANALYSIS: Layer-wise analyses compare PCA-projected embeddings, statistical summaries, and attention-derived vectors across evaluation dimensions.Reported signals include an Informativeness peak near Layer 25, Logicality concentration near Layers 17 and 27, and a Fluency peak at Layer 17.