Organizations: The Conversational Artificial Intelligence (CoAI) Group, Tsinghua University · Zhipu AI · University of Electronic Science and Technology of China
Instruction-following is a foundational capability of large language models (LLMs), with its improvement hinging on scalable and accurate feedback from judge models. However, the reliability of current judge models in instruction-following remains underexplored due to several deficiencies of existing meta-evaluation benchmarks, such as their insufficient data coverage and oversimplified pairwise evaluation paradigms that misalign with model optimization scenarios. To this end, we propose IF-RewardBench, a comprehensive meta-evaluation benchmark for instruction-following that covers diverse instruction and constraint types. For each instruction, we construct a preference graph containing all pairwise preferences among multiple responses based on instruction-following quality. This design enables a listwise evaluation paradigm that assesses the capabilities of judge models to rank multiple responses, which is essential in guiding model alignment. Extensive experiments on IF-RewardBench reveal significant deficiencies in current judge models and demonstrate that our benchmark achieves a stronger positive correlation with downstream task performance compared to existing benchmarks. Our codes and data are available at https://github.com/thu-coai/IF-RewardBench.
Figures & tables
Figure 1: An example from IF-RewardBench, containing a user instruction, a constraint checklist, and multiple responses with various instruction-following quality that form a preference graph.
Benchmark
Dataset Size
Diverse
Instruction Type
Annotator
Evaluation Paradigms
#Inst.
#Resp.
#Rela.
#Model
Constraint
Multi-Turn
System Prompt
LLMBar ( 2024 )
419
836
419
3
✗
✗
✗
Human
Pairwise
InfoBench ( 2024 )
50
250
0
5
△
✗
✗
Human
Pointwise
IFBench ( 2025b )
444
888
444
1
△
✗
✗
GPT-4o
Pairwise
PPE-IF ( 2025 )
512
16384
2560
4
△
✗
✗
Synthetic
Paiwise & BoN
RewardBench-2-IF ( 2025 )
160
640
480
5
△
✗
✗
Synthetic
BoN
Table 1: Comparisons of IF-RewardBench and other meta-evaluation benchmarks for instruction-following. The data size contains the number of instructions (#Inst.), responses (#Resp.), preference relations (#Rela.), and models to generate responses (#Model). △ denotes partially satisfied.
Figure 2: Overall framework of IF-RewardBench. Left : Collect instructions and responses from diverse sources. Center : Curate preference graphs via multi-stage annotation and verification. Right : Assess various judge models based on different evaluation paradigms.
Instruction Type
#Inst.
#Turn.
#Cons.
# Resp.
#Rela.
Single-Turn Interaction
393
1.00
5.52
7.05
10.57
Multi-Turn Interaction
202
3.23
4.50
7.20
11.21
System-Prompt Steerability
247
2.01
5.89
7.23
11.04
Overall
842
1.83
5.38
7.14
10.86
Table 2: Statistics of IF-RewardBench, including the number of instructions ( #Inst. ), the average dialog turns ( #Turn ), the number of constraints ( #Cons. ) per instruction, the average number of responses ( #Resp. ) and preference relations ( #Rela. ) per preference graph.
Model
Single-Turn
Multi-Turn
System-Prompt
Average
P-F1
N-F1
τb
P-F1
N-F1
τb
P-F1
N-F1
τb
P-F1
N-F1
τb
Human
0.926
0.782
0.789
0.927
0.710
0.714
0.917
0.739
0.761
0.923
0.744
0.755
Proprietary general language models, prompted as judge models
Gemini-3-Pro
0.911
0.693
0.639
0.912
0.688
0.570
0.904
0.662
0.619
0.909
0.681
0.609
Gemini-3-Flash
0.908
0.666
0.561
0.908
0.682
0.599
0.886
0.630
0.555
0.901
0.660
0.572
GPT-5.1
0.895
0.607
0.513
0.896
0.627
0.514
0.869
0.595
0.550
0.887
0.610
0.525
Table 3: The F1-scores for both positive (P-F1) and negative (N-F1) classes in verification, and Kendall ( τb ) correlation in ranking on constraint assessment.
Figure 3: The distribution of constraint categories and constraint composition types for instructions in IF-RewardBench.
Model
Single-
Multi-
System-
Avg.
Turn
Turn
Prompt
General language models
Gemini-3-Flash
0.589
0.460
0.489
0.513
GPT-5-mini
0.521
0.438
0.410
0.456
DeepSeek-V3.2
0.397
0.257
0.208
0.288
GLM-4.6
0.359
0.263
0.189
0.270
Table 4: Kendall ( τb ) correlation in ranking on overall assessment.
Figure 4: The performance of judge models in verification across different constraint categories and composition types. The constraint composition types are bold .
Figure 5: Various factors that influence the performance of judge models in ranking. "CA" and "OA" denote Constraint Assessment and Overall Assessment, respectively.
Model
Qwen-3-32B
GLM-4.6
Task
CA
OA
CA
OA
Baseline
0.285
0.129
0.422
0.270
Long-Chain Reasoning
w/o Thinking
0.253 (-11.2%)
0.106 (-17.8%)
0.284 (-32.7%)
0.253 (-6.3%)
Self-Consistency
w/ Maj@3
0.298 (+4.6%)
0.147 (+14.0%)
0.460 (+9.0%)
0.297 (+10.0%)
Table 5: Kendall ( τb ) correlation in ranking under different inference-time scaling strategies for Qwen-3-32B and GLM-4.6.
Figure 6: Pairwise accuracy of judge models in various meta-evaluation benchmarks.
Figure 7: Somers’ D correlation between the performance of 15 judge models in various meta-evaluation benchmarks and Best-of-8 sampling.
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
Model
τb
Gemini-3-Flash
0.161
GPT-5-mini
0.211
DeepSeek-V3.2
-0.474
GLM-4.6
-0.568
GLM-4.5-Air
-0.347
QwQ-32B
-0.324
Appendix
Table 6: Kendall ( τb ) correlation in ranking on overall assessment, calculated on a subset of IF-RewardBench whose instructions are curated from IHEval.
Model
BoN
BoN
LLMBar
RB2
IFBench
IF-RewardBench
IF-RewardBench
(CA)
(OA)
(CA)
(OA)
Gemini-3-Flash
0.851
0.854
0.887
0.794
0.802
0.572
0.513
GPT-5-mini
0.838
0.857
0.853
0.713
0.757
0.519
0.456
DeepSeek-V3.2
0.820
0.835
0.881
0.681
0.678
0.395
0.288
GLM-4.6
0.817
0.837
0.868
0.513
0.617
0.422
0.270
GLM-4.5-Air
0.815
0.826
0.824
0.506
0.640
0.308
0.148
Appendix
Table 7: Detailed results of judge models on Best-of-8 selection and different meta-evaluation benchmarks. "LLMBar" denotes "LLMBar-Adversarial". "RB2" denotes "RewardBench-2-IF". "CA" and "OA" denote Constraint Assessment and Overall Assessment, respectively. Since the last three models are dedicated discriminative reward models, we only report their performance in the overall assessment task.
Constraint Category
Description
Example
Numerical
Constraints that specify quantitative requirements, such as the count of words, sentences, or paragraphs, are typically independent of the specific content.
• Please write a 15-line poem. • The article must not exceed 1000 words. • Please provide two different solutions.
Format
Constraints that involve presentation format or structural organization of the response, such as JSON, Markdown, or bullet points. Note that formats implicitly defined through in-context examples are also considered as format constraints.
• Entire response must be output in JSON format. • Mark all typos in the text using ^, for example: ^(typo)^. • Use "First / Second / Third" to list the points in the answer.
Content
Constraints that involve the specific content of the response, such as topic, subject, and entity.
• The article must / must not contain the keyword "happy". • The response must follow the "Phenomenon-Cause-Solution" three-part structure. • Your article must not contain any specific examples or anecdotes.
Linguistic
Constraints that involve language and linguistic properties of the response, such as language, grammar, vocabulary, and rhetorical devices.
• The first paragraph must be in German, the second in Chinese, and the third in Classical Chinese. • The last word of every sentence should rhyme. • Employ metaphor and parallelism in your article.
Style
Constraints that involve the writing style of the response, such as literary genre, emotion, tone, and narrative perspective.
• Write an article in the style of magical realism. • Use a first-person perspective. • Write a diary entry in the style of a Xiaohongshu post.
Situation
Constraints that require the model to respond within a specific scenario or adapt a specific persona, ensuring the output aligns with the prescribed situational conditions or role-specific characteristics.
• Assume you are Churchill near the end of World War II. • You live in a world that defies the laws of physics, where gravity fluctuates cyclically throughout the day. • You are an elderly person suffering from Alzheimer’s disease, experiencing confused and fragmented memories.
Appendix
Table 8: Constraint categories in IF-RewardBench.
Composition Type
Description
Example
Single
The output is required to follow a single constraint.
Please summarize the following news.
And
The output is required to follow multiple constraints simultaneously.
Please summarize the following news. The summary should be output in bullet points and within 100 words .
Chain
The output is required to complete multiple tasks sequentially, each of which needs to follow its own constraints.
Please introduce “Mona Lisa” briefly. Firstly , introduce the year of creation, then describe the background of the work’s creation, and finally , summarize the impact of the work.
Selection
The output is required to select the correct branch according to certain conditions, and then follow the constraints of this branch.
Please introduce the following painting. - If the work contains any animal , the description should be in English - Otherwise , the description should be in Chinese Painting: "Mona Lisa"
Nested Structure
The above composition types are recursively nested to form more complex structures.
Analyze the sentiment of the above user comment and complete the following tasks: 1. If it’s positive : - Identify the products within the comments … 2. If it’s negative , analyze the reasons for it: - If the reason is not about the products themselves , … - Otherwise , …
Appendix
Table 9: Constraint composition types in IF-RewardBench, which are adopted from ComplexBench Wen et al. (2024) . Composed constraints are highlighted in bold . Note that the first 4 basic composition types can be nested to construct more complex structures. The final row demonstrates a nested selection type.
Instruction Type
Source
# Instruction
Single-Turn Interaction
InfoBench Qin et al. (2024)
21
FollowBench Jiang et al. (2024)
15
ComplexBench Wen et al. (2024)
38
Inverse-IFEval Zhang et al. (2025a)
25
IFBench Pyatkin et al. (2025)
90
CFBench Zhang et al. (2025b)
25
Appendix
Table 10: The sources of the instructions in IF-RewardBench.
Type
Model
# Size
Creator
# Instruction
API
GPT-5 OpenAI (2025)
Undisclosed
OpenAI
20
Gemini-2.5-Pro Comanici et al. (2025)
Undisclosed
DeepMind
25
Claude-4.5-Sonnet Anthropic (2025)
Undisclosed
Anthropic
32
Doubao-1.6-Seed ByteDance (2025)
Undisclosed
ByteDance
30
OSS (Large)
Kimi-K2 Team et al. (2025)
1T
Moonshot AI
32
DeepSeek-R1-0528 Guo et al. (2025a)
671B
Deepseek
50
Appendix
Table 11: The models and the number of instructions ( # Instruction ) used for response generation. "API" denotes API-based proprietary LLMs, and "OSS" denotes Open-sourced LLMs.
Type
Model
# Size
Creator
General LLMs
GPT-5.1 OpenAI (2025)
Undisclosed
OpenAI
GPT-5-mini OpenAI (2025)
Undisclosed
OpenAI
Gemini-3-Pro Comanici et al. (2025)
Undisclosed
DeepMind
Gemini-3-Flash Comanici et al. (2025)
Undisclosed
DeepMind
DeepSeek-V3.2 Liu et al. (2025a)
671B
DeepSeek
GLM-4.6 Zeng et al. (2025)
360B
Zhipu AI
Appendix
Table 12: Judge models evaluated in this paper. "RM" denotes Reward Models, and "UIUC" denotes University of Illinois Urbana-Champaign.
System Prompt
Your response should contain exactly 3 bullet points in markdown format and exactly 4 sections. Bullet points are indicated by *. Sections are separated by 3 asterisks: ***. You should use words with all capital letters at least 2 times.
Conversation History
[User Prompt] Write a blog post about ’how to improve your writing skills’. Your response should contain exactly 5 bullet points in markdown format, indicated by -. Your response should have exactly 2 sections, separated by 3 dashes: —. You should not use any words with all capital letters in your response.
Constraint Checklist
1. Your answer should contain exactly 3 bullet points indicated by *. 2. Your answer should contain exactly 4 paragraphs separated by 3 asterisks: ***. 3. In your response, words with all capital letters should appear at least 2 times.
Chosen Response (Response B)
Writing is a craft that thrives on habit and observation. By dedicating time each day to the act of putting thoughts on paper, you lay the groundwork for continuous growth. *** * Establish a daily writing habit, even if it’s just a few minutes. PROACTIVE practices create momentum. * Read a diverse range of genres and analyze what makes them compelling. Use FOCUS to identify techniques that resonate with you. * Seek constructive feedback, revise, and learn from mistakes. The cycle of writing and revising deepens understanding. *** Beyond the routine, experiment with different formats—short stories, essays, or journal entries—to expand your voice. Reflect on what resonates and refine your approach. *** Remember that mastery is a marathon, not a sprint. Stay curious, keep learning, and let every sentence be a step forward.
Golden Judgements of Chosen Response
[1, 1, 1]
Rejected Response (Response A)
improving your writing skills: a quick guide - read widely to expand your vocabulary and understand different styles - write daily, even if it’s just a short note, to develop rhythm and habit - review and edit what you write, looking for clarity and conciseness — refining your craft - seek feedback from peers or mentors to gain perspective and identify blind spots - study grammar and style guides to reinforce solid fundamentals
Golden Judgements of Rejected Response
[0, 0, 0]
Appendix
Table 13: A case study of judge models when the system prompt and the user prompt are in conflict. The failure of judge models to distinguish the priority of system prompts versus user prompts is highlighted in red .
System Prompt
Your response must consist of exactly one sentence. Put double quotation marks around your entire one-sentence response.
Conversation History
[User Prompt] Why is Algiers the best place to go on vacation? Your response must consist of at least two sentences. Put single quotation marks around your entire response.
Constraint Checklist
1. Wrap your entire response with double quotation marks. 2. Your response should contain fewer than 2 sentences. 3. Your response should contain at least 1 sentence.
Chosen Response (Response A)
"Algiers offers a perfect blend of Mediterranean beaches, rich history, vibrant culture, and delicious cuisine that creates an unforgettable vacation experience."
Golden Judgements of Chosen Response
[1, 1, 1]
Rejected Response (Response B)
’Algiers offers a unique blend of Mediterranean charm and North African culture, with its stunning coastline, historic Casbah district, and vibrant local markets providing endless opportunities for exploration and discovery. The city’s affordability compared to other Mediterranean destinations, combined with its delicious cuisine, warm hospitality, and fascinating history spanning from ancient Phoenician times to French colonial rule, makes it an exceptional vacation spot that offers authentic experiences without breaking the bank.’
Golden Judgements of Rejected Response
[0, 0, 1]
Appendix
Table 14: A case study of judge models when the system prompt and the user prompt are in conflict. The failure of judge models to distinguish the priority of system prompts versus user prompts is highlighted in red , while a successful example of Gemini-3-Flash is highlighted in blue .
Below, an instruction that may contain the system prompt and conversation history, and a constraint checklist for the instruction will be provided. The checklist includes all the constraints the model must follow when generating a response. Your task is to verify whether the checklist accurately and comprehensively covers all constraints, and to modify it if any errors are found. Constraint Definition: Constraints refer to all instructional directives from the system prompt, conversation history, and user instructions that apply to the final-round response, excluding auxiliary information such as background knowledge, reference texts, and in-context examples. They include, but are not limited to, descriptions of the basic tasks to be completed and specific requirements for output content, style, and format. Task Details: Step 1: Error Identification Evaluate the accuracy of the checklist. Select "Incorrect" if any of the following conditions are met: 1. Vague Constraints: The checklist includes ambiguous constraints, such as "satisfy all constraints mentioned above" or "your response must be correct". 2. Omissions: The checklist misses mandatory constraints present in the system prompt, conversation history, or user instruction. 3. Expired Constraints: The checklist includes outdated constraints from the history or system prompt that no longer apply to the current round. 4. Fabricated Constraints: The checklist contains constraints not presented in the system prompt, history, or user instruction. 5. Incorrect Paraphrasing: The checklist inaccurately interprets a constraint (e.g., the original instruction states "list the key points in an ordered list," but the checklist incorrectly paraphrases it as "list the key points in an unordered list"). Step 2: Checklist Modification If you select "Incorrect" in Step 1, modify the checklist to make it correct. You must follow these principles: 1. Source Integrity: All checklist items must originate from the provided information. 2. Completeness and Accuracy: Do not omit necessary constraints, duplicate existing ones, or fabricate new ones. 3. Atomicity: Each constraint must be atomic yet possess complete semantics, ensuring it can be fully understood without referencing other constraints. {examples} [System Prompt] {system_prompt} [Conversation History] {history} [Final Round User Instruction] {user_prompt} [Constraint Checklist] {checklist} Your judgment of the checklist quality: {option} A. Correct B. Incorrect Your modified checklist (Required if B is selected): {modified-checklist}
Appendix
Table 15: Human annotation guideline for checklist revision. The red part is the information provided to the annotators, and the blue part is content that requires the annotators to make annotations.
Below, an instruction that may contain the system prompt and conversation history, a corresponding model response, and a constraint checklist for the instruction will be provided. The checklist contains some of the constraints within the instruction. Your task is to verify whether the model response follows each constraint in the checklist, choose "Followed" or "Not Followed", and provide a brief justification. Task Details: 1. Please analyze whether the response follows each constraint listed in the given checklist, providing a judgment for each constraint respectively. 2. Your judgments must be strict. Only responses that fully satisfy a constraint can be judged as "Followed". If there is any omission or error regarding a constraint, it must be judged as "Not Followed". 3. Please focus exclusively on the constraints within the given checklist. It is unnecessary to consider whether the response follows any other constraints beyond the checklist. 4. When judging the following of each constraint, your judgement should consider the complete context of the instructions, rather than interpreting the constraint in isolation. {in-context examples} [System Prompt] {system_prompt} [Coversation History] {history} [Final Round User Instruction] {user_prompt} [Model Response] {model_response} [Constraint Checklist] {checklist} Your choice for the first constraint in the checklist: {option} A. Followed B. Not Followed Your justification: {justification} Your choice for the second constraint in the checklist: {option} A. Followed B. Not Followed Your justification: {justification} ……
Appendix
Table 16: Human annotation guideline for instruction-following annotation. The red part is the information provided to the annotators, and the blue part is content that requires the annotators to make annotations.
You are an expert specializing in evaluating the quality of user instructions. I will provide you with: 1. A system prompt (which may be empty): Specifies the response rules or behavioral guidelines that the AI assistant needs to satisfy throughout the dialogue. 2. Conversation history between the user and the AI assistant (which may be empty): Presents the dialogue process between the user and the AI assistant, which consists of multiple rounds of user instructions and AI assistant responses. 3. The final round user instruction. Your task is to evaluate the quality of the final round user instruction based on the following criteria. ## Evaluation Criteria: 1. Low Quality : The user instruction contains significant issues, such as inconsistent, incomplete, or ambiguous information, making it impossible to determine the intent behind the instruction. 2. Medium Quality : The user instruction may include some inconsistent, incomplete, or ambiguous elements; however, the overall intent can still be inferred. 3. High Quality : The user instruction is consistent, complete, and unambiguous, allowing the intent to be easily and definitively understood. ## Note: 1. You do not need to answer the user instruction, but only need to output the evaluation result. 2. If the user instruction requires additional retrieval or the use of tools to obtain an answer, it should be evaluated as Low Quality . Here are some examples and the user instruction to be evaluated: [The Start of Examples] {in_context_examples} [The End of Examples] Below are the system prompt, conversation history, and final round user instruction: [The Start of System Prompt] {system_prompt} [The End of System Prompt] [The Start of Conversation History] {history} [The End of Conversation History] [The Start of Final Round User Instruction] {user_prompt} [The End of Final Round User Instruction] ## Output Format """ Analysis: … Prompt Quality: Low Quality / Medium Quality / High Quality """
Appendix
Table 17: The prompt template for scoring the quality of user instructions.
You are an expert specializing in evaluating the difficulty of user instructions. I will provide you with: 1. A system prompt (which may be empty): Specifies the response rules or behavioral guidelines that the AI assistant needs to satisfy throughout the dialogue. 2. Conversation history between the user and the AI assistant (which may be empty): Presents the dialogue process between the user and the AI assistant, which consists of multiple rounds of user instructions and AI assistant responses. 3. The final round user instruction. 4. A checklist: Lists all constraints that the AI assistant needs to satisfy when generating a response to the final round user instruction. Your task is to evaluate the difficulty of the final round user instruction. This instruction contains multiple constraints that need to be followed. Your evaluation of its difficulty needs to comprehensively consider the number of constraints in the instruction, as well as the difficulty of following these constraints. The difficulty of following these constraints is more critical. Please first follow the above principles to analyze the difficulty of the instruction in detail. After providing your analysis, output your difficulty scores from 1 to 10 by strictly following this format: "Score: [[5]]". A higher score indicates that the instruction has a higher difficulty. Below are the system prompt, conversation history, final round user instruction, and checklist: [The Start of System Prompt] {system_prompt} [The End of System Prompt] [The Start of Conversation History] {history} [The End of Conversation History] [The Start of Final Round User Instruction] {user_prompt} [The End of Final Round User Instruction] [The Start of Checklist] {checklist} [The End of Checklist]
Appendix
Table 18: The prompt template for scoring the difficulty of user instructions.
You are an expert specializing in extracting all constraints contained in instructions. I will provide you with: 1. A system prompt (which may be empty): Specifies the response rules or behavioral guidelines that the AI assistant needs to satisfy throughout the dialogue. 2. Conversation history between the user and the AI assistant (which may be empty): Presents the dialogue process between the user and the AI assistant, which consists of multiple rounds of user instructions and AI assistant responses. 3. The final round user instruction. Your task is to extract all constraints that need to be followed when generating a response to the final round user instruction, from the system prompt, conversation history, and the final round user instructions. Constraints refer to all instructional content, excluding auxiliary information such as background knowledge , text materials , and in-context examples . This includes, but is not limited to, descriptions of basic tasks to be completed, and specific requirements on output content, style, and format. Output all constraints using this format exactly: """ [The Start of Constraint 1] Constraint: … (Please generate a specific constraint of the instruction. It must be complete and detailed, with no information omitted or altered in any way) [The End of Constraint 1] [The Start of Constraint 2] Constraint: … (Please generate a specific constraint of the instruction. It must be complete and detailed, with no information omitted or altered in any way) [The End of Constraint 2] … """ Remember the following points: 1. You must extract and output all constraints in the order in which they appear in the instruction, without omitting any constraints or inventing constraints not present in the instruction. To reiterate: Constraints refer to all instructional content, excluding auxiliary information such as background knowledge , text materials , and in-context examples . 2. Do not output duplicate constraints. Each portion of the given instruction should appear in at most one constraint, and must not be repeated across multiple constraints. 3. Each constraint should be atomic and independent, without inclusion or dependency relationships. Multiple dependent constraints should be merged into a single constraint. At the same time, ensure appropriate granularity of constraints: neither too fine nor too coarse. Each constraint should have complete semantics and be understandable on its own, without reference to other constraints. 4. Vague or general phrases in the instructions, such as "satisfy the following requirements" or "complete the following tasks", should not be considered as constraints and should be excluded from extraction. Below are the system prompt, conversation history, and final round user instruction: [The Start of System Prompt] {system_prompt} [The End of System Prompt] [The Start of Conversation History] {history} [The End of Conversation History] [The Start of Final Round User Instruction] {user_prompt} [The End of Final Round User Instruction]
Appendix
Table 19: The prompt template for instruction decomposition.
You are an impartial judge specializing in evaluating the quality of AI assistant responses. I will provide you with: 1. A system prompt (which may be empty): Specifies the response rules or behavioral guidelines that the AI assistant needs to satisfy throughout the dialogue. 2. Conversation history between the user and the AI assistant (which may be empty): Presents the dialogue process between the user and the AI assistant, which consists of multiple rounds of user instructions and AI assistant responses. 3. The final round user instruction. 4. The final round assistant response. 5. A checklist: Lists all constraints that the AI assistant needs to satisfy when generating a response to the final round user instruction. Your task is to think carefully and provide a detailed analysis of whether the final round assistant response follows each constraint on the checklist. You need to analyze and judge every constraint in the checklist. Do not omit any constraint, and do not analyze any constraint not present in the checklist. You must strictly output the analysis and judgment for each constraint in the following format: """ [The Start of Constraint 1] Constraint: … (Directly output the first constraint from the checklist here with no modifications) Explanation: … (Provide a thorough, step-by-step analysis of whether the AI assistant’s response follows this constraint, referencing specific details of the response) Judgment: [[The AI assistant’s response follows this constraint]] or [[The AI assistant’s response does not follow this constraint]] [The End of Constraint 1] [The Start of Constraint 2] Constraint: … (Directly output the second constraint from the checklist here with no modifications) Explanation: … (Provide a thorough, step-by-step analysis of whether the AI assistant’s response follows this constraint, referencing specific details of the response) Judgment: [[The AI assistant’s response follows this constraint]] or [[The AI assistant’s response does not follow this constraint]] [The End of Constraint 2] … """ Remember the following points: (1) Your judgment should be as strict as possible. You should only output "[[The AI assistant’s response follows this constraint]]" if the response completely follows every part of the constraint as stated, without any omission or mistake. (2) Your judgment of each constraint in the should remain independent. When judging a given constraint, you should not take into account whether other constraints are followed. (3) When judging a given constraint, interpret it within the broader context of the complete system prompt, conversation history, final round user instruction, and checklist, rather than relying solely on its literal meaning in isolation. (4) For length-related constraints that specify a target value (e.g., "X words" or "around X words"), rather than a range, an actual response length between 90% and 110% of the required amount X is considered to follow this constraint. Below are the system prompt, conversation history, final round user instruction, final round assistant response, and checklist: [The Start of System Prompt] {system_prompt} [The End of System Prompt] [The Start of Conversation History] {history} [The End of Conversation History] [The Start of Final Round User Instruction] {user_prompt} [The End of Final Round User Instruction] [The Start of Final Round Assistant Response] {assistant_response} [The Start of Final Round Assistant Response] [The Start of Checklist] {checklist} [The End of Checklist]
Appendix
Table 20: The prompt template for constraint assessment.
Please act as an impartial judge and evaluate the quality of the responses provided by two AI assistants to the user question displayed below. You should choose the assistant that follows the user’s instructions better. Your evaluation should prioritize instruction-following, which means the responses correctly satisfy the constraints of the instructions. Begin your evaluation by comparing the two responses and provide a short explanation. Avoid any position biases and ensure that the order in which the responses were presented does not influence your decision. Do not allow the length of the responses to influence your evaluation. Do not favor certain names of the assistants. Be as objective as possible. After providing your explanation, output your final verdict by strictly following this format: "[[A]]" if assistant A is better, "[[B]]" if assistant B is better. [The Start of System Prompt] {system_prompt} [The End of System Prompt] [The Start of Conversation History] {history} [The End of Conversation History] [The Start of Final Round User Instruction] {user_prompt} [The End of Final Round User Instruction] [The Start of Assistant A’s Response] {response_a} [The End of Assistant A’s Response] [The Start of Assistant B’s Response] {response_b} [The End of Assistant B’s Response]
Appendix
Table 21: The prompt template for pairwise comparison in overall assessment.
Large language model (LLM)-as-a-judge research uses language models to score, compare, and verify AI outputs as scalable alternatives to human evaluation.
Rubric-based evaluation has become a prevailing paradigm for evaluating instruction following in large language models (LLMs). Despite its widespread use, the reliability of these rubric-level evaluations remains unclear, calling for meta-evaluation. However, prior meta-evaluation efforts largely focus on the response level, failing to assess the fine-grained judgment accuracy that rubric-based evaluation relies on. To bridge this gap, we introduce RubricEval. Our benchmark features: (1) the first rubric-level meta-evaluation benchmark for instruction following, (2) diverse instructions and responses spanning multiple categories and model sources, and (3) a substantial set of 3,486 quality-controlled instances, along with Easy/Hard subsets that better differentiates judge performance. Our experiments reveal that rubric-level judging remains far from solved: even GPT-4o, a widely adopted judge in instruction-following benchmarks, achieves only 55.97% on Hard subset. Considering evaluation paradigm, rubric-level evaluation outperforms checklist-level, explicit reasoning improves accuracy, and both together reduce inter-judge variance. Through our established rubric taxonomy, we further identify common failure modes and offer actionable insights for reliable instruction-following evaluation.
Tianjun Pan, Xuan Lin, Wenyan Yang +7
College of Computer Science and Artificial Intelligence, Fudan University · Ant Group · Donghua University
Multi-constraint instruction following requires verifying whether a response satisfies multiple individual requirements, yet LLM judges are often assessed only through overall-response judgments. We introduce MCJudgeBench, a benchmark for constraint-level judge evaluation in multi-constraint instruction following. Each instance includes an instruction, a candidate response, an explicit constraint list, per-constraint gold labels in {yes, partial, no}, and controlled response-side perturbations. The evaluation protocol further includes evaluation prompt variants to test judge stability. We evaluate proprietary and open-source LLM judges using both correctness and inconsistency metrics, distinguishing intrinsic inconsistency under stochastic decoding from procedural inconsistency under prompt and response perturbations. Our results show that judge reliability has multiple dimensions: strong overall performance does not guarantee equally reliable detection across label categories, especially for rarer partial and no cases. Judges with higher correctness do not always have lower inconsistency. Evaluation with reasoning improves correctness but does not uniformly improve stability. These findings motivate evaluating LLM judges at the constraint level to study these failure modes.
Common instruction following (IF) benchmarks, such as IFEval, evaluate the ability of language models (LMs) to follow simple constraints that can easily be verified programmatically. These same constraints, in turn, are used as reward signals for reinforcement learning. This design may lead to reward hacking of constraints, pushing models away from the main task. Recent attempts to address this resulted in benchmarks with richer constraints, but required expert labor, remained closed-access, and cannot be easily extended. We present Lift-Bench, an IF benchmark that asks a model to draft a document given a topic and a structured outline. Constraints are scoped to certain document sections, providing finer control over constraint density. Lift-Bench includes both verifiable and judged constraints, as well as a training counterpart, Lift-Train. Importantly, Lift-Bench provides both main task and instruction-following constraints, mitigating shallow reward hacking. On experiments with multiple models, we find that frontier models struggle with the harder subsets of our benchmark and analyze model performance along multiple axes. We also find that training on Lift-Train is much less susceptible to reward hacking than training on the common IFTrain. Lift-Bench and Lift-Train are publicly available, as well as the data generating pipeline.