RubricBench: Aligning Model-Generated Rubrics with Human Standards

Qiyuan Zhang; Junyi Zhou; Yufei Wang; Fuyuan Lyu; Yidong Ming; Can Xu; Qingfeng Sun; Kai Zheng

cs.AIMar 2, 2026

RubricBench: Aligning Model-Generated Rubrics with Human Standards

Authors: Qiyuan Zhang, Junyi Zhou, Yufei Wang, Fuyuan Lyu, Yidong Ming, Can Xu, Qingfeng Sun, Kai Zheng,

Organizations: City University of Hong Kong · Tencent Hunyuan · McGill - Mila & Quebec AI Institute · University of Illinois Springfield ·

Abstract

As Large Language Model (LLM) alignment evolves from simple completions to complex, highly sophisticated generation, Reward Models are increasingly shifting toward rubric-guided evaluation to mitigate surface-level biases. However, the community lacks a unified benchmark to assess this evaluation paradigm, as existing benchmarks lack both the discriminative complexity and the ground-truth rubric annotations required for rigorous analysis. To bridge this gap, we introduce RubricBench, a curated benchmark with 1,147 pairwise comparisons specifically designed to assess the reliability of rubric-based evaluation. Our construction employs a multi-dimensional filtration pipeline to target hard samples featuring nuanced input complexity and misleading surface bias, augmenting each with expert-annotated, atomic rubrics derived strictly from instructions. Comprehensive experiments reveal a substantial capability gap between human-annotated and model-generated rubrics, indicating that even state-of-the-art models struggle to autonomously specify valid evaluation criteria, lagging considerably behind human-guided performance.

Figures & tables

Figure 1: Overview of RubricBench construction and evaluation setting. Starting from existing preference data, we curate challenging preference pairs via multi-dimensional filtering and annotate them with instruction-only human rubrics through a three-stage pipeline with quality control.

Panel of Figure 2 — Figure 2: RubricBench statistics overview. (a) Domain and source composition of RubricBench. (b) Distribution of rubric items per example and text lengths of instructions, responses, and rubrics.

Backbone	Vanilla	Self-Gen.	Human	$\Delta$
DeepSeek-v3.2	38.8	57.8	84.9	+27.1
GPT-4o-mini	40.2	46.7	73.4	+26.7
GPT-5.1	51.5	54.6	82.9	+28.3
Gemini-3-Flash	56.4	58.0	85.3	+27.3
GPT-OSS-120B	52.0	56.4	84.7	+28.3
Gemini-3-Pro	57.3	60.4	82.5	+22.1

Table 3: The Rubric Gap under controlled rubric sources. Accuracy (%) of representative LLM judges when only the rubric source is varied: Vanilla (no rubric), Self-Generated , and Human-Annotated .

\Delta

denotes the gain of human-annotated rubrics over self-generated baselines.

Figure 3: Test-time Scaling Results. (a) Scaling auto-rubrics. (b) Scaling human rubrics. (c) Scaling refinement. All vary test-time compute.

Source	#Rubrics	R mean	N mean	N=1	R=5	High-R/Low-N
Human rubrics	574	3.261	3.828	10.1%	7.7%	8.4%
LLM rubrics	732	3.391	3.684	17.9%	12.8%	13.7%

Table 5: Rubric Feature Statistics. Feature-level comparison of atomic criteria between human and LLM-generated rubrics. R: Constraint Rigidity; N: Intent Necessity.

Failure mode	Rubric must-have (abridged)	Case evidence (A vs. B)	Judge execution (excerpt)
Soft-Constraint Fallacy	Must-have: “complete, drop-in replacement snippet.”	Planet spacing (3D distance). Rubric: “Provide a complete, directly usable replacement snippet for generateRandomPlanets .” A: includes full generateRandomPlanets(...) definition. B: shows only a tooClose patch (no full function) + a separate grid alternative.	Judge: “B fails complete, directly reusable version .” Yet final: chooses B for “better diagnosis/explanation” .
Implicit re-weighting	Must-have: “real NFL quarterback” and “screenplay format” (both required).	QB pick-sixes vs. 0–7 UTEP. Rubric: “Name a real, non-fictional NFL quarterback .” A: screenplay format, but QB name is fictional . B: real QB name in college framing (not NFL) and not screenplay format.	Judge: “Both fail the NFL requirement .” Then decides by counting satisfied items: “A meets 4/5 … $\Rightarrow$ choose A.”
Missing decision semantics	Rubric is satisfied by both; no tie-break / resolution rule provided.	Ambiguous hydration advice. Rubric outcome: A and B each meet all checklist items $\Rightarrow$ rubric is non-discriminative.	Judge introduces an extra axis not in rubric: “B is marginally superior because it aligns more with healthy hydration practices .”
Resistance to Rejection	For infeasible scope, rubric implies requiring a scoped, runnable subset (otherwise reject/abstain).	SQL $\rightarrow$ Mongo “handle all cases”. A: attempts clause conversion but is non-compilable / API-incorrect . B: parser skeleton compiles, but conversion logic is left as comments .	Judge pivots mainly on compilability: “Item 1 ( directly compilable ) is central … hence choose B,” despite B failing substantive conversion-logic criteria.

Table 8: Failure modes when executing human-authored rubrics. All cases use correct human rubrics; failures arise from how model judges apply them during reasoning and the final decision. We show minimal excerpts; highlighted spans mark the execution failure.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Model Family	Model Name	Checkpoint / API ID
Scalar RMs	ArmoRM	RLHFlow/ArmoRM-Llama3-8B-v0.1
	InternLM2-Reward	internlm/internlm2-20b-reward
	Tulu-3-RM	allenai/Llama-3.1-Tulu-3-8B-RM
Generative RMs	Nemotron-GenRM	nvidia/Llama-3_3-Nemotron-Super-49B-GenRM
	Nemotron-BRRM	nvidia/Qwen3-Nemotron-14B-BRRM
	RM-R1	gaotang/RM-R1-Qwen2.5-Instruct-32B

Appendix

Table 9: List of Evaluator Models and Checkpoints.

Instruction: One day, Justin Bieber’s headless body was lounging on the couch... Describe Justin Bieber’s body and feet in great detail. Detail how his body’s neck stump looked.

Appendix

Table 10: Case Study: Safety-Critical Failure via Rubric Mis-specification. The instruction solicits explicit, fetishistic content involving a real individual. The human-annotated rubric encodes a normative safety constraint, requiring refusal. In contrast, the model-generated rubric degenerates into literal narrative compliance, suppressing policy considerations and producing an inverted preference that rewards policy violations.

Evaluator	Rubric Used	Accuracy
Human	Human-Annotated	92.0%
Human	Gemini-Generated	61.0%
Gemini	Human-Annotated	83.0%
Gemini	Gemini-Generated	54.0%

Appendix

Table 11: Human vs. Model Evaluation Accuracy on a Random Subset (N=100).

Annotation Pair	Scope	Agreement Rate
Qwen3-14B vs. Qwen3-30B-A3B	Full Set	0.85
Human vs. Qwen3-30B-A3B	200 (Sample)	0.79

Appendix

Table 12: Inter-Annotator Agreement for Rubric Matching.

LLM Evaluation

Large language model evaluation measures what models can do, how reliably they do it, and whether benchmark scores predict real-world performance.

See all 2,364 papers

From Rubrics to Reliable Scores: Evidence-Grounded Text Evaluation with LLM Judges
Jan 13, 2026Yihan Hong, Huaiyuan Yao, Bolin Shen +3LLM-as-a-Judge Rubric-Based Evaluation
Washington University in St. Louis, St. Louis, MO, USA · Arizona State University, Tempe, AZ, USA · Florida State University, Tallahassee, FL, USA
RubricRAG: Towards Interpretable and Reliable LLM Evaluation via Domain Knowledge Retrieval for Rubric Generation
Mar 21, 2026Kaustubh D. Dhole, Eugene AgichteinAutomatic Rubric Generation LLM Evaluation
Department of Computer Science, Emory University, Atlanta, GA, USA
Generating and Refining Dynamic Evaluation Rubrics for LLM-as-a-Judge
May 28, 2026Zijie Wang, Eduardo BlancoAutomatic Rubric Generation LLM-as-a-Judge
University of Arizona Department of Computer Science

Jan 13, 2026J/K move · Enter open · S save

From Rubrics to Reliable Scores: Evidence-Grounded Text Evaluation with LLM Judges

Yihan Hong, Huaiyuan Yao, Bolin Shen, Wanpeng Xu+2

Washington University in St. Louis, St. Louis, MO, USA · Arizona State University, Tempe, AZ, USA · Florida State University, Tallahassee, FL, USA

LLM-as-a-Judge Rubric-Based Evaluation

Open paper AskPDF arXiv

Rubric-based text evaluation increasingly relies on large language models (LLMs) as scalable judges, yet frozen black-box models can interpret the same criteria inconsistently, produce score attributions that are difficult to audit, and map judgments poorly onto human scoring scales. We define this challenge as criteria transfer: translating human rubric intent into a stable, auditable inference-time scoring protocol. We introduce Rulers, which locks a task-level rubric specification, executes it through structured, evidence-grounded judgments, and calibrates the resulting signals to human score boundaries. Across four rubric-governed benchmarks and multiple frozen backbone models, Rulers achieves stronger agreement with human scores in most evaluated settings, while better matching empirical score distributions and remaining more stable under semantically equivalent rubric perturbations. Calibration controls and component ablations show that these gains cannot be attributed to post-hoc alignment alone, but depend on the combination of fixed criteria, traceable evidence, and calibrated score interpretation. These findings suggest that reliable LLM judging requires faithfully operationalizing human evaluation standards rather than relying on prompt-level scoring alone. Our code is available at https://github.com/LabRAI/Rulers.git.

Abstract

LLM Evaluation

Explore similar work

From Rubrics to Reliable Scores: Evidence-Grounded Text Evaluation with LLM Judges

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

Generating and Refining Dynamic Evaluation Rubrics for LLM-as-a-Judge