RubricRAG: Towards Interpretable and Reliable LLM Evaluation via Domain Knowledge Retrieval for Rubric Generation

Kaustubh D. Dhole; Eugene Agichtein

Source-linked AI summary

RubricRAG: Towards Interpretable and Reliable LLM Evaluation via Domain Knowledge Retrieval for Rubric Generation

Kaustubh D. Dhole, Eugene Agichtein

arXiv:2603.20882v1Mar 21, 2026cs.IRcs.AIcs.CLcs.LG

TL;DR

The paper addresses whether LLMs can generate interpretable, query-specific rubrics despite the opacity of scalar LLM judges and the cost of manual rubric design. It proposes RubricRAG, which retrieves rubrics from similar queries, and finds that retrieval-based conditioning most effectively improves rubric alignment and downstream judgment.

Problem
Automated graders are opaque, while manually designing high-quality query-specific rubrics is labor-intensive; the paper asks whether LLM-generated rubrics can be interpretable and effective for evaluation.
Method
The paper evaluates rubric-generation strategies and introduces RubricRAG, which conditions generation on rubrics retrieved from semantically similar queries.
Results
Retrieval-based conditioning most effectively improves alignment with human-authored rubrics and helps LLM judges prefer good responses over bad ones.
Takeaways & Limitations
Effective automated rubric generation depends on contextual grounding, making retrieval a promising route toward more interpretable and actionable LLM evaluation.
Takeaways & Limitations
RubricRAG can increase redundancy by generating overlapping rubric items, despite improving coverage and reducing missed and hallucinated criteria.

Abstract

from arXiv · show

Large language models (LLMs) are increasingly evaluated and sometimes trained using automated graders such as LLM-as-judges that output scalar scores or preferences. While convenient, these approaches are often opaque: a single score rarely explains why an answer is good or bad, which requirements were missed, or how a system should be improved. This lack of interpretability limits their usefulness for model development, dataset curation, and high-stakes deployment. Query-specific rubric-based evaluation offers a more transparent alternative by decomposing quality into explicit, checkable criteria. However, manually designing high-quality, query-specific rubrics is labor-intensive and cognitively demanding and not feasible for deployment. While previous approaches have focused on generating intermediate rubrics for automated downstream evaluation, it is unclear if these rubrics are both interpretable and effective for human users. In this work, we investigate whether LLMs can generate useful, instance-specific rubrics as compared to human-authored rubrics, while also improving effectiveness for identifying good responses. Through our systematic study on two rubric benchmarks, and on multiple few-shot and post-training strategies, we find that off-the-shelf LLMs produce rubrics that are poorly aligned with human-authored ones. We introduce a simple strategy, RubricRAG, which retrieves domain knowledge via rubrics at inference time from related queries. We demonstrate that RubricRAG can generate more interpretable rubrics both for similarity to human-authored rubrics, and for improved downstream evaluation effectiveness. Our results highlight both the challenges and a promising approach of scalable, interpretable evaluation through automated rubric generation.

1 Introduction and Background

Automated LLM graders are convenient but opaque, motivating query-specific rubrics that make evaluation criteria explicit and actionable. The paper studies whether LLMs can generate such rubrics and introduces retrieval-based conditioning to improve their quality.

LLM-as-judge pipelines often provide preferences or scalar scores without explaining missed requirements or how to improve responses.This opacity complicates model development, dataset curation, and high-stakes deployment.
Rubric-based evaluation decomposes quality into explicit, checkable criteria, but generic criteria may miss individual-query requirements.Query-specific rubrics instead target the particular requirements of each query.
Fine-grained query-specific rubrics improve LLM judges’ ability to distinguish good from bad responses compared with coarse or absent rubrics.Coarse rubrics can be marked satisfied for both responses, producing many ties.
Designing high-quality rubrics for every query is tedious and cognitively demanding, requiring choices about evaluation dimensions, granularity, accuracy, and safety.
The study asks whether LLMs can generate fine-grained query-specific rubrics similar to human rubrics and useful for choosing good over bad responses.
The authors introduce rubric-alignment metrics, evaluate prompting and post-training approaches, and propose RubricRAG, which retrieves rubrics from similar queries.They report that zero-shot LLMs are poor rubric generators, while retrieval and post-training can improve rubric generation and downstream evaluation.

2 Related Work

Related work develops fine-grained evaluation through skills, checklists, and query-specific rubrics, while also using rubrics as structured learning signals. These approaches seek more informative judgments or rewards than coarse scalar scores.

Fine-grained LLM judges decompose evaluation into explicit dimensions, and FLASK reports improved agreement with human judgments over coarse scores.
RocketEval structures judging as answering checklist questions about outputs, enabling small evaluator models to achieve high correlation with human preferences.
Rubrics can provide multifaceted feedback and serve as reward signals for reinforcement learning when strict verification is difficult.Rubrics as Rewards reports gains over scalar reward formulations.

3 Methods and Experiments

The paper next introduces its task and the approaches used to generate and evaluate rubrics.

The methods section describes the task and the rubric-generation approaches used in the study.

3.1 Task: Rubric generation

Given a user query, the task is to generate fine-grained criteria for grading the assistant’s next response. Criteria assign positive or negative points to desirable behaviors and failure modes.

Rubric generation maps a user query q to a set of fine-grained criteria R for grading the assistant’s next response.
Each criterion has an integer point value, with positive points rewarding desirable behavior and negative points penalizing failure modes.
Criteria may be positive or negative and should cover both what the assistant should do and what it should avoid.

3.2 Rubric Generation Approaches

The paper compares zero-shot, few-shot, retrieval-based, supervised, and reinforcement-learning approaches for generating query-specific rubrics. RubricRAG retrieves rubrics from similar queries to provide domain and query-specific context during generation.

Approach comparison: The study compares raw zero-shot generation with methods that provide additional context from other queries to generate effective rubrics.
Zero-shot and Few-shot generation: Zero-shot generation prompts an LLM to produce positive and negative rubric criteria with integer point weights in strict JSON format.
RubricRAG: RubricRAG retrieves k similar queries with a dense retriever and places their query–rubric pairs into context as few-shot exemplars.The retrieval is intended to familiarize the model with query-specific nuances beyond the general domain and task.
Post-training approaches: Supervised fine-tuning trains the generator to predict human-authored rubrics from the user query using teacher forcing and QLoRA adapters.
Post-training approaches: GRPO optimizes rubric generation with weighted rewards for format correctness, similarity to human rubrics, rubric diversity, and distributional length properties.

3.3 Evaluation metrics

The evaluation combines rubric similarity metrics with downstream tests of score correlation and good-versus-bad response discrimination. The metrics treat rubrics as sets of criteria and quantify overlap, misses, hallucinations, and redundancy.

Rubric similarity metrics: Rubric-BLEU, Rubric-ROUGE, and Rubric-LLM-Judge measure rubric alignment using macro-averaged, per-criterion best overlap.The formulation computes generated-to-reference precision and reference-to-generated recall while remaining invariant to criterion ordering.
Hallucinations, misses, and redundancy: The evaluation tracks hallucinations, missed reference rubrics, and redundancy among generated rubric pairs through query-wise metrics.
Downstream evaluation: Downstream evaluation measures both correlation between scores from generated and human-authored rubrics and the ability to prefer good responses over bad responses.
Downstream evaluation: The authors validate their human-authored-rubric and LLM-judge setup with a score of .37, within the range reported for HealthBench’s closed-source models.

3.4 Evaluation Across Several Rubric Granularities

The study first tests whether rubric granularity improves LLM judges’ ability to distinguish good from bad responses. It compares no rubrics, five static axis-level rubrics, and more fine-grained cluster-level rubrics across models and settings.

Evaluation question: The study asks whether human-authored rubrics of different granularities help LLM judges discriminate good from bad responses better than no rubrics.
Rubric granularities: The comparison includes no rubrics, axis-level rubrics with five static criteria, and cluster-level rubrics consisting of 37 rubrics shared across many queries.
Scoring procedure: Each rubric criterion is evaluated separately for good and bad responses, and the sum of satisfied-criterion points becomes the response score.In the no-rubric setting, the judge instead outputs a single score.
Evaluation data: Physician-written completions serve as good responses, while a Qwen3-30B-A3B-Instruct-2507 model generates bad responses conditioned on rubrics from random queries.

3.5 Experimental Setup

The experiments use Qwen models for rubric generation and judging, with different exemplar counts for HealthBench and ResearchRubrics. The setup also compares thinking and non-thinking RubricRAG variants and uses dense retrieval for similar queries.

Models and decoding: Qwen3-14B2 generates rubrics, while Qwen3-4B-Instruct-25073 judges rubric entailment and performs downstream evaluation.
Models and decoding: The experiments use k=20 exemplars for HealthBench and k=5 for ResearchRubrics, with greedy decoding and a maximum token length of 1024.
Experimental scope: Smaller models including Qwen3-0.6B, 1.7B, 4B-Instruct, and 8B frequently produced malformed JSON and would require substantial output-cleaning logic.
RubricRAG setup: RubricRAG is evaluated with and without intermediate thinking tokens, while similar queries are retrieved using Qwen3-Embedding-4B.

3.6 Datasets and splits

The evaluation uses OpenAI HealthBench, including physician-authored reference rubrics, with separate subsets for evaluation and training.

OpenAI HealthBench provides complex user queries paired with physician-authored reference rubric lists.
Evaluation uses 300 random queries from the oss_eval subset and all queries from the hard subset.
The remaining oss_eval queries are used for training.

4 Results

Fine-grained, query-specific rubrics improve discriminative evaluation, while zero-shot generation poorly matches human rubrics. RubricRAG improves similarity and downstream alignment through retrieval, although it can introduce redundant criteria.

4.1 Downstream Effectiveness of Different Granularities of Human-Authored Rubrics: Fine-grained human-authored rubrics more accurately distinguish good from bad responses than coarse or absent rubrics.Coarser rubrics are often marked satisfied for both responses, producing ties.
4.2 Similarity to Human-Authored Rubrics.: Zero-shot LLMs consistently show low rubric-BLEU and rubric-ROUGE and only moderate LLM-judge similarity across three benchmarks.They capture high-level intent but miss the fine-grained structure and clinically grounded criteria in human-authored rubrics.
4.2 Similarity to Human-Authored Rubrics.: Random few-shot exemplars improve lexical and semantic rubric similarity, but the gains remain modest.Examples provide the expected format, granularity, and balance of positive and negative criteria.
4.2 Similarity to Human-Authored Rubrics.: Retrieval over semantically similar queries improves rubric alignment and can outperform expensive SFT approaches.RubricRAG achieves the highest rubric-similarity scores among the reported approaches.
4.2 Similarity to Human-Authored Rubrics.: RubricRAG without intermediate thinking and SFT achieve the highest rubric-similarity scores, while reasoning-based variants perform worse.The results suggest model-generated intermediate tokens can add noise or misguide rubric generation.
4.3 Zero-shot vs. RubricRAG: quantitative and qualitative analysis: RubricRAG reduces hallucinated rubrics but often generates redundant criteria.Qualitative analysis links retrieval to greater coverage and specificity at the cost of overlapping criteria that can over-weight concepts.
4.3 Zero-shot vs. RubricRAG: quantitative and qualitative analysis: RubricRAG produces more task-specific criteria, including emergency transfer, danger signs, and practical preparation steps.These criteria better match low-resource clinical contexts than the generic, underspecified criteria produced by zero-shot generation.
4.4 Downstream Effectiveness of Generated Rubrics for LLM Judges: Spearman’s ρ ranges from 0.331 to 0.545 across model-generated and human-authored rubric scores.RubricRAG achieves the highest downstream correlations and remains within 5% error of human-authored corpus-level scores.

5 Conclusion

The study finds that retrieval-based conditioning, especially RubricRAG, most effectively produces human-aligned rubrics and improves downstream judgment, while revealing remaining redundancy and gaps from human-authored design.

Human-authored query-specific rubrics are more effective than coarse or absent rubrics for helping LLM judges distinguish good from bad responses.
Off-the-shelf LLMs are weak zero-shot rubric generators; few-shot prompting improves lexical and semantic similarity but remains limited without query-relevant examples.
RubricRAG retrieves rubrics from semantically similar queries, improving alignment with human rubrics and downstream preference for good responses.
Rubric-similarity scores do not necessarily predict downstream evaluation effectiveness, so generated rubrics should also be assessed on judgment tasks.
RubricRAG improves coverage and reduces missed or hallucinated criteria relative to zero-shot generation, but increases redundancy through overlapping rubric items.
Automatically generated query-specific rubrics are promising for interpretable evaluation, but current models still fall short of human-authored rubric design.