InFoBench: Evaluating Instruction Following Ability in Large Language Models

1 Introduction

• Instruction-following remains underexplored, while A/B testing scales as N^2 comparisons and opaque overall scores do not explain scores for individual instructions.Existing evaluations often emphasize general NLP or downstream tasks rather than instruction-following itself.Source: §1 · p. 1
• DRFR decomposes complex instructions into distinct criteria, enabling granular, scalable, and interpretable assessment of compliance with individual requirements.The study presents it as a more detailed assessment of instruction-following than traditional metrics.Source: §1 · pp. 1–2
• INFOBENCH contains 500 diverse instructions and 2,250 decomposed questions, with constraints classified as Content, Linguistic, Style, Format, or Number.Constraint labels were manually curated for each question to support analysis of model deficiencies.Source: §1 · p. 2
• DRFR produces greater annotator consensus than Direct Scoring, particularly on the Hard Set; GPT-4 offers an accurate, cost- and time-efficient annotation alternative with structured multi-turn prompts.Human experts were most accurate but costly and time-consuming, while AMT workers offered a balance with lower accuracy.Source: §1 · p. 2
• Advanced LLMs still struggle to follow instructions perfectly, especially on complex tasks; closed-source models lead, while numerical and linguistic understanding remain areas for improvement.The authors suggest closed-source models may benefit from better data or more sophisticated algorithms.Source: §1 · p. 2

2.1 DRFR

• DRFR evaluates complex instruction following by decomposing instructions into simpler, judgeable requirements, addressing the difficulty of reaching evaluator consensus on partial compliance.For N instructions, each instruction is decomposed into m_i requirements, and r_i,j records whether each requirement is met.Source: §2.1 · p. 3
• DRFR scores each criterion as met or unmet, with even minor deviations receiving NO, making it a strict measure of precise instruction adherence.A YES corresponds to a value of 1 for the criterion; NO indicates noncompliance or failure to provide relevant information.Source: §2.1 · p. 3

2.2 INFOBENCH Dataset

• The benchmark pairs an Easy Set with an original, manually curated Hard Set designed to address the Easy Set’s simple instructions, limited constraint variety, and ambiguous requirements.The Easy Set’s “helpful” requirement is given as an example of subjectivity that can make evaluator consensus difficult.Source: §2.2 · pp. 3–4

2.2.1 Instruction Development

• The Hard Set covers 72 domains and uses content, linguistic, style, format, and number constraints to create instructions that distinguish stronger from weaker models.These constraint categories govern what responses address, their language and tone, their structure, and numerical requirements.Source: §2.2.1 · p. 4

2.2.2 Instruction Decomposition

• Each decomposed requirement is phrased as a binary YES/NO question and manually labeled with up to three constraint types, since one requirement can span multiple types.The paper’s example of critiquing Starry Night is classified as both content and format.Source: §2.2.2 · p. 4

2.2.3 Benchmark Statistics

• The Hard Set has more elaborate instructions and more requirements than the Easy Set, indicating greater complexity, while the benchmark spans a broad range of domains.Table 2 reports benchmark statistics; the accompanying text identifies the Hard Set’s greater complexity and domain breadth.Source: §2.2.3 · pp. 4–5, Table 2 · p. 5
• About two-thirds of Easy Set constraints are content constraints, whereas the Hard Set has a more heterogeneous distribution of constraint types.Figure 2 compares constraint-type distributions across the two sets.Source: Figure 2 · p. 5, §2.2.3 · p. 5

3.1 DRFR v.s. Direct Scoring

• Three NLP experts rated model responses with Direct Scoring on a 1–5 scale before using DRFR, then compared model pairs and measured evaluator agreement with Fleiss’ Kappa.Both metrics were standardized into instruction-level scores for comparison.Source: §3.1 · pp. 5–6
• Across 50 INFOBENCH instances—25 Easy and 25 Hard—the observed Pairwise Kappa Agreement gap exceeding 0.19 suggests DRFR yields notably higher annotator consensus than Direct Scoring (DS).Source: §3.1 · p. 6

3.2 Experiment of Annotation Sources

• The annotation-source comparison used expert majority votes with online research, three qualified AMT workers per task, and GPT-4 prompts that sequentially presented decomposed questions with prior-answer context.Pairwise Label Distance measured disagreement with the human-expert ground truth: 0 indicates correct relative ordering, 1 a potential tie-related misclassification, and 2 an incorrect ordering.Source: §3.2 · pp. 6–7
• 89% accuracy and 0.44 weighted pairwise label distance were achieved by GPT-4 using decomposed questions, alongside greater cost-effectiveness and time efficiency than AMT.The authors therefore used GPT-4 for their INFOBENCH model evaluation, citing its balance of cost, time efficiency, and accuracy.Source: §3.2 · p. 7, Table 4 · p. 7

Overall

• GPT-4-0314 with a multi-turn prompt evaluated six selected LLMs on InFoBench using greedy decoding.The models were gpt-3.5-turbo-1106 (Brown et al., 2020), gpt4-1106-preview, Claude-2.1, Gemini-Pro (Team, 2023), Llama-2-70b-chat, and Vicuna-13bv1.5.Source: §4 · p. 7
• GPT-4 scored highest, yet more than 10% of requirements remained unfulfilled.Table 5 reports automated DRFR scores assigned by GPT-4-0314 for the six selected models.Source: §4 · pp. 7–8, Table 5 · p. 8
• All models except Claude-2.1 performed better on the Easy Set, while closed-source models outperformed open-source models on the Hard Set.The authors suggest the latter pattern may reflect stronger instruction coverage for advanced instructions or better generalization.Source: §4 · p. 8

Constraint Type

• Across all six models, performance was highest on Content and Style, intermediate on Format, and lowest on Number and Linguistic constraints.Figure 3 presents the constraint-type comparison.Source: §4 · p. 8, Figure 3 · p. 8

Domain

• On the Hard Set, scores were approximately circular across seven principal subject domains, except for Vicuna, suggesting domain was not the main performance factor.The authors also suggest that most models may have good data coverage across these domains.Source: §4 · p. 8, Figure 4 · p. 8

Related Work

• Instruction tuning uses annotated instruction–outcome data (Wang et al., 2022; Lou et al., 2023), while alignment tuning uses human feedback to align models with values and preferences (Ziegler et al., 2019; Christiano et al., 2017; Ouyang et al., 2022; Bai et al., 2022).Alignment tuning also aims to curb unintended behaviors such as fabricating information or producing biased expressions (Ouyang et al., 2022; Kenton et al., 2021).Source: §5 · pp. 8–9
• The authors use LLMs as evaluators of instruction-following, building on LLM-based natural-language-generation evaluation across numerous tasks (Liu et al., 2023; Wang et al., 2023a).The section also cites Fu et al. (2023), Kocmi and Federmann (2023), Xu et al. (2023), Chiang et al. (2023), and Eldan and Li (2023) as NLG evaluation work.Source: §5 · p. 9
• Prior work spans scientific-reasoning assessment (Wang et al., 2023b), judge LLMs (Wang et al., 2023c), classification-task evaluation (Li et al., 2023a), and instruction-tuned capability suites (Chia et al., 2023).Other cited directions include win-rate evaluation and automatic composition of complex instructions for instruction-tuning.Source: §5 · p. 9
• Unlike Jiang et al. (2023), which sequentially adds fine-grained constraints, this work decomposes one instruction into multiple constraints; unlike Zhou et al. (2023), its 25 verifiable instructions, it offers broader coverage while retaining an objective focus.Zhou et al. (2023) emphasizes objective evaluation with 25 verifiable instructions.Source: §5 · p. 9

6 Conclusion

• DRFR decomposes complex instructions into simpler criteria, and InFoBench enables detailed assessment of LLM instruction-following capabilities.The paper presents DRFR as a novel metric and InFoBench as a comprehensive benchmark dataset.Source: §6 · p. 9
• The findings report DRFR's superiority over traditional evaluation methods and identify GPT-4 as a cost-effective, accurate, and efficient annotator for large-scale evaluations.The authors say DRFR offers clearer insights into model performance, especially in complex scenarios.Source: §6 · p. 9
• Evaluating various LLMs revealed progress, especially in closed-source models, but complex instructions remain challenging, including for numerical reasoning and linguistic comprehension.The authors frame these findings as a foundation for future work toward more accurate and reliable real-world performance.Source: §6 · p. 9

• Only 10% of instructions (50) received human annotations, so reliability and annotation comparisons rely solely on that subset.This constraint reflects the cost and resources required for human annotation.Source: Limitations · p. 9
• The benchmark contains 500 instructions and 2,250 decomposed questions, and manual creation limits its scalability.The authors suggest automated instruction writing and question decomposition as a possible path to larger datasets.Source: Limitations · pp. 9–10
• Evaluation focuses predominantly on explicit instruction intentions and omits factors such as truthfulness and harmlessness (Askell et al., 2021).The authors identify these factors as avenues for future research.Source: Limitations · p. 10

A.1 Instruction Dataset Statistics

• Easy Set instructions predominantly span 10–20 words, while Hard Set instructions average around 50 words.Source: §A.1 · p. 12
• Easy Set instructions average approximately 2–3 questions, compared with 5–7 questions per instruction in the Hard Set.Source: §A.1 · p. 12
• The Hard Set’s longer instructions and greater number of questions per instruction indicate enhanced complexity relative to the Easy Set.Source: §A.1 · p. 12
• Table 6 categorizes questions by constraint type for both the Easy Set and Hard Set, showing their distribution across types.Source: §A.1 · p. 12

A.2 Prompt Design for GPT-4 Automatic Evaluation

• GPT-4 answers each decomposed question with “YES” or “NO,” receiving the input (if any) and generated text, but not the actual instruction because the questions are hypothesized to capture its requirements.For multiple questions, each is asked sequentially in a multi-turn dialogue, with prior answers included as context for the next question.Source: §A.2 · p. 12
• Sequential multi-turn questioning yielded more effective results than asking all questions simultaneously or asking each independently.Each instruction and generation is followed by several questions, and each response provides context for the subsequent question.Source: §A.2 · p. 12
• Few-shot examples produced no significant improvement, possibly because instructions and questions vary widely and the answer format is simple.Source: §A.2 · p. 12

A.3 Human Annotation Guidance

• Annotators are instructed to read the guidelines before beginning the annotation.The crowd-sourced worker guidelines and Amazon Mechanical Turk interface are shown in Figure 6.Source: §A.3 · p. 12, Figure 6 · p. 15
• Annotators can select “UNKNOWN” when they find a question difficult to answer, and those responses are excluded from experiments and evaluation.Source: §A.3 · p. 12
• Systems are randomly shuffled in each instance to reduce the possibility that annotators memorize which system is better.Source: §A.3 · p. 12

A.4 GPT-4-1106 Evaluation Results

• GPT-4-1106-preview largely mirrors earlier evaluation trends but assigns marginally lower scores than GPT-4-0314; performance patterns across constraint types and domains remain consistent.The authors say the score difference warrants further investigation and regard consistency across model versions as supporting GPT-4’s reliability as an evaluation tool.Source: §A.4 · p. 16

A.5 GPT-4 Annotation Error Analysis

• The analysis distinguishes human disagreement caused by ambiguous or subjective questions from actual GPT-4 annotation errors.Examples include vague requirements such as “around 250 words” and subjective criteria such as an “appealing” title.Source: §A.5 · p. 16
• Around 30% to 40% of questions had Disagreement Levels of at least 2 and were excluded, leaving the 60% to 70% with high human agreement for error analysis.The study defines Disagreement Level by comparing each annotator’s binary judgment with the three-annotator majority; across five LMs, its range is 0 to 5.Source: §A.5 · pp. 16–17, Figure 9 · p. 17
• GPT-4 can be misled when a response repeats instruction text instead of completing the task, and it can err when counting characters or distinguishing repeated references as unique items.The examples include a purported eviction notice and assessments involving an exact 20-character string or two distinct tree types.Source: §A.5 · pp. 17–18
• Other errors arise when outputs omit a required cost or difficulty level, when evaluating spatial constraints such as a 6x6 grid, or when content deviates from conventional linguistic patterns.The grid example shows that GPT-4 may rely on text claiming a grid is 6x6 even when its rendered dimensions do not match.Source: §A.5 · pp. 18–20

A.6 Case Study

• None of the models fully satisfied the fourth and fifth number constraints, with DNA nucleotide proportion and equal-representation requirements presenting particular difficulty.The paper characterizes the composition requirements as somewhat contradictory; Tables 9–11 cover the number-constraint case study.Source: §A.6 · p. 20, Table 9 · p. 21, Table 10 · p. 22
• None of the models completely satisfied the third and fourth linguistic constraints; GPT-4-1106 attempted the last-letter rule for most words, while the other models did not appear to attempt it.The unusual constraint compromised GPT-4-1106’s plot clarity, and the paper attributes the broader difficulty to linguistic patterns unlike those models are optimized to produce; Tables 12–13 cover this case.Source: §A.6 · p. 20, Table 12 · p. 24, Table 13 · p. 25