PaperlayerView exact revision on arXiv

Source-linked AI summary

FollowBench: A Multi-level Fine-grained Constraints Following Benchmark for Large Language Models

Yuxin Jiang, Yufei Wang, Xingshan Zeng, Wanjun Zhong, Liangyou Li, Fei Mi, Lifeng Shang, Xin Jiang, Qun Liu, Wei Wang

arXiv:2310.20410v3cs.CL

TL;DR

Existing benchmarks emphasize response quality rather than whether outputs satisfy fine-grained constraints, leaving comprehensive and difficulty-sensitive instruction-following evaluation limited. FollowBench introduces five constraint categories and a multi-level benchmark that adds constraints incrementally, and evaluations of 13 LLMs show declining performance with difficulty and substantial advantages for GPT-4 and GPT-3.5.

  • Problem

    Existing benchmarks primarily evaluate response quality rather than whether outputs satisfy diverse fine-grained constraints at varying difficulty levels.

    Source: §1 · p. 1

  • Method

    FollowBench evaluates five constraint categories with a multi-level mechanism that incrementally adds constraints and uses rule-based or model-based judges to assess each constraint.

    Source: §1 · p. 2, §3 · p. 3

  • Results

    Across 13 LLMs, performance declines as constraint counts increase, while GPT-4 and GPT-3.5 markedly outperform open-source models.

    Source: §1 · p. 2, §6 · p. 9

  • Takeaways & Limitations

    FollowBench indicates substantial room to improve current LLM instruction following and supports evaluating proficiency beyond knowledge and reasoning.

    Source: §6 · p. 9

    The model-based evaluation framework relies on prompt engineering, and further prompt optimization may affect reported metrics.

    Source: Limitations · p. 9

Abstract

from arXiv · show

The ability to follow instructions is crucial for Large Language Models (LLMs) to handle various real-world applications. Existing benchmarks primarily focus on evaluating pure response quality, rather than assessing whether the response follows constraints stated in the instruction. To fill this research gap, in this paper, we propose FollowBench, a Multi-level Fine-grained Constraints Following Benchmark for LLMs. FollowBench comprehensively includes five different types (i.e., Content, Situation, Style, Format, and Example) of fine-grained constraints. To enable a precise constraint following estimation on diverse difficulties, we introduce a Multi-level mechanism that incrementally adds a single constraint to the initial instruction at each increased level. To assess whether LLMs' outputs have satisfied every individual constraint, we propose to prompt strong LLMs with constraint-evolution paths to handle challenging open-ended instructions. By evaluating 13 closed-source and open-source popular LLMs on FollowBench, we highlight the weaknesses of LLMs in instruction following and point towards potential avenues for future work. The data and code are publicly available at https://github.com/YJiangcm/FollowBench.

1 Introduction

FollowBench addresses the lack of comprehensive, difficulty-sensitive evaluation of fine-grained instruction constraints by combining five constraint categories with incrementally increasing levels. Evaluations of 13 LLMs show that instruction-following performance declines as constraints increase, with closed-source models outperforming open-source models and Situation and Example constraints proving especially difficult.

Contributions

  • FollowBench evaluates Content, Situation, Style, Format, and Example constraints using levels that add one constraint at a time to initial instructions.It contains 820 instructions spanning more than 50 NLP tasks, with rule-based and model-based evaluation verified by humans.Source: §1 · p. 2

Experimental Findings

  • Performance declines substantially for all tested models as difficulty increases through the number of constraints in an instruction.Source: §1 · p. 2
  • GPT-4 and GPT-3.5 satisfy around three constraints consecutively on average while markedly surpassing all evaluated open-source models.Source: §1 · p. 2
  • Situation and Example constraints are more challenging for LLMs than other constraint categories.Source: §1 · p. 2

Implications

  • Instruction following provides an additional lens for assessing LLM proficiency beyond knowledge and reasoning capabilities.Source: §1 · p. 2

2 Related Work

Prior work improves general instruction following and evaluates selected constraint types, but existing benchmarks do not comprehensively cover constraint categories or difficulty levels. FollowBench addresses this gap with five constraint types and multi-level instructions for more precise estimation.

2.1 Instruction-Following Language Models

  • Instruction-tuning research uses annotated instructional data and broad human-crafted task collections, while newer work automatically generates data to improve instruction following.Source: §2.1 · p. 2

2.2 Evaluation for Instruction Following

  • Existing evaluation benchmarks target particular capabilities such as complex structured tabular generation or knowledge-intensive instructions with example-based contradictions.Source: §2.2 · p. 2
  • FollowBench differs by covering five fine-grained constraint types across multiple difficulty levels, enabling more comprehensive and precise instruction-following estimation.Source: §2.2 · p. 2

3 FollowBench

FollowBench constructs a multi-level benchmark spanning five fine-grained constraint categories and evaluates instruction following with evolution-aware LLM judging and satisfaction metrics.

3.1 Data Construction

  • FollowBench spans Content, Situation, Style, Format, and Example constraints, plus mixed-constraint instructions combining multiple categories.The benchmark contains 820 diverse instructions, with initial instructions selected to maintain low pairwise ROUGE-L similarity.Source: §3 · p. 3, §3.1 · p. 5
  • Each instruction progresses from level 1 to level 5 by incrementally adding one constraint, enabling estimation of the difficulty level at which models fail.Constraint additions are designed separately for each task and category rather than applied uniformly.Source: §3 · p. 3, Figure 2 · p. 4, §3.2 · p. 6
  • The dataset operationalizes categories through task-specific construction: content narrows requested information or adds restrictions, while situation adds background information and complex reasoning contexts.Content sources include information extraction, language-constrained generation, and open-ended question answering; situation sources include role-playing, math, temporal/spatial reasoning, and code generation.Source: §3.1 · pp. 3–4
  • Style and Format instructions are synthesized with progressively added stylistic or structural requirements and then reviewed by human experts.Style constraints target tone, sentiment, formality, and empathy, while format constraints cover length, hierarchy, linguistic features, and output media.Source: §3.1 · pp. 4–5
  • Example constraints test robustness to noisy demonstrations by placing n−1 noise examples in level-n inputs.Mixed constraints additionally combine categories for tasks such as text editing, summarization, machine translation, and story generation.Source: §3.1 · p. 5

3.2 Evaluation Protocol

  • The evaluation protocol exposes each instruction’s constraint-evolution path to a strong LLM judge and measures complete, per-constraint, and consecutive-level satisfaction.HSR measures the proportion of instructions whose constraints are all satisfied, SSR averages satisfaction across individual constraints, and CSL counts consecutive satisfied levels from level 1.Source: §3.2 · pp. 5–6

4 Experiments

The experiments evaluate 13 LLMs across FollowBench’s difficulty levels and constraint categories, showing that instruction-following performance declines as constraints increase and varies by category.

4.1 Experimental Setup

  • The evaluation tests 13 popular closed-source and open-source LLMs with deterministic inference settings and also includes a Chinese FollowBench version.Source: §4.1 · p. 6

4.2 Level-categorized Results

  • Performance typically declines from L1 to L5 as the number and strictness of constraints increase.Table 2 averages scores across all constraint categories for each difficulty level.Source: §4.2 · p. 6, Table 2 · p. 7
  • GPT-4 and GPT-3.5 substantially outperform open-source models, with an instruction-following upper bound of about three constraints versus about two for open-source models.The reported upper bound is based on the constraint-satisfaction level (CSL).Source: §4.2 · p. 7
  • The scaling pattern is not uniform: LLaMA2-Chat-70B outperforms LLaMA2-Chat-13B on Situation constraints but underperforms it on Format and Mixed Constraints.Source: §4.2 · p. 7

4.3 Constraint-categorized Results

  • GPT-4 and GPT-3.5 surpass open-source models across every constraint category, especially Content, Situation, Example, and Mixed constraints.Figure 5 reports HSR averages across all difficulty levels for each category.Source: §4.3 · p. 7, Figure 5 · p. 7
  • Style adaptation is a relative strength for many models, whereas Example and Mixed constraints challenge most models, including GPT-4.The Example setting uses noise examples with varied natural-language templates, which the authors relate to limited training on noisy contextual inputs.Source: §4.3 · pp. 7–8

5 Analysis

The analysis validates the model-based evaluation protocol and examines how instruction following relates to other abilities, failure consistency, and decoding temperature. Results show that detailed constraint evolution improves evaluation agreement, stronger models follow more constraints, and moderate temperature balances adherence with variability.

5.1 Ablation Study of Model-based Evaluation

  • 88% agreement with expert human evaluations exceeds 85% human agreement, while removing constraint evolution reduces agreement by 9%.The standard overall-quality prompt reaches 67% agreement, underscoring the distinction between response quality and constraint following.Source: §5.1 · p. 8

5.2 Instruction Following vs. Other Abilities

  • FollowBench adds an instruction-following perspective beyond overall response quality, knowledge, and reasoning, revealing capability rankings that differ across abilities.WizardLM-13B-V1.2 exceeds GPT-3.5 in overall response quality but lags in instruction following, while Vicuna-V1.5 surpasses LLaMA2 in knowledge and reasoning but struggles on instruction following.Source: §5.2 · p. 8, Table 4 · p. 8

5.3 Does Failure at Lower Level Necessarily Lead to Failure at Higher Level?

  • Failure consistency measures whether failing at level n predicts failure at every later level, and stronger instruction-following models may show lower consistency.The authors suggest more capable models are less sensitive to added constraints and can recover at higher levels after lower-level failures.Source: §5.3 · p. 8, Table 5 · p. 9

5.4 Does Different Decoding Strategies Affect the Instruction-following Ability?

  • Temperature influences instruction following across all four models, with intermediate values appearing to balance variability and staying on topic.Lower temperatures produce more consistent outputs, whereas higher temperatures produce more diverse and creative outputs.Source: §5.4 · p. 9, Figure 6 · p. 9

6 Conclusion

FollowBench provides a multi-level benchmark spanning five constraint categories and over 50 NLP tasks, with evaluations showing GPT-4 and GPT-3.5 outperform other tested models. The study is limited to single-round interactions, prompt-engineered evaluation, and no proposed solution to identified weaknesses.

6 Conclusion

  • FollowBench covers five fine-grained constraint categories and over 50 NLP tasks, using a multi-level mechanism and three evaluation metrics.The benchmark is designed to estimate instruction-following upper limits across progressively constrained instructions.Source: §6 · p. 9
  • GPT-4 and GPT-3.5 show a substantial performance advantage over the other tested models, while current LLMs still have significant room for improvement.The conclusion summarizes extensive testing across 13 popular LLMs.Source: §6 · p. 9

Limitations

  • The evaluation is confined to single-round interactions, leaving instruction following in multi-round conversations for future research.The single-round setting was chosen to provide a controlled evaluation environment.Source: Limitations · p. 9
  • The model-based evaluation framework relies on prompt engineering, and further prompt optimization may affect the reported evaluation metrics.The authors characterize this reliance as an inherent imperfection despite selecting high-performing prompts.Source: Limitations · p. 9
  • The paper does not propose specific solutions for the identified instruction-following weaknesses, deferring them to subsequent studies.Fine-tuning with FollowBench is suggested as a possible future direction rather than evaluated here.Source: Limitations · p. 9

Ethics Statement

  • FollowBench removes potentially harmful instructions and offensive content through human review to support a secure evaluation framework.The ethics statement emphasizes safe and responsible assessment of instruction following.Source: Ethics Statement · pp. 9–10

A Data Generation Process

The data-generation process builds multi-level Content and Situation constraints from diverse NLP datasets and online sources. Difficulty increases by adding constraints progressively while retaining task-specific evaluation metrics.

A.1 Content Constraints

  • Content constraints draw on data-to-text, event argument extraction, named entity recognition, constrained text generation, and open-ended question answering.Sources include E2E, WIKIEVENTS, CONLL-2003, COGNAC, WordNet, Wikidata, public evaluation sets, and online platforms.Source: §A.1 · p. 13
  • Content difficulty increases from one to five constraints, with exact match or accuracy used to evaluate task-specific outputs.Named entity requirements become more specific as constraints increase, while other tasks add attributes, events, or linguistic restrictions.Source: §A.1 · p. 13

A.2 Situation Constraints

  • Situation constraints include suggestion generation, role-playing, math word problems, time and spatial reasoning, and code generation.Data comes from open-ended datasets and platforms, GSM8K, AGIEval, BIG-Bench Hard, and HumanEval.Source: §A.2 · pp. 13–14
  • Situation difficulty is raised by enriching circumstances, adding temporal or spatial details, increasing function-description complexity, and preserving the core math question.Accuracy is used for math and reasoning tasks, while code generation uses pass@1.Source: §A.2 · p. 14

A.3 Example Constraints

Example constraints test few-shot pattern recognition through progressively more question and answer templates. Mixed-constraint tasks add editing, formatting, content, style, and other requirements across text transformation and generation datasets.

A.3 Example Constraints

  • Example constraints use 40 PromptSource NLP tasks and 29 answer templates to test matching question patterns with the corresponding response format.At higher levels, five shots are paired with multiple question templates and answer templates, and accuracy is the evaluation metric.Source: §A.3 · p. 14, Table 6 · p. 14

A.4 Mixed Constraints

  • Mixed constraints cover text editing, summarization, machine translation, and story generation, with requirements added progressively across levels.Rule-based programs evaluate internal constraints for editing, summarization, and translation, while story generation adds format, content, and style requirements.Source: §A.4 · pp. 14–15
  • The mixed-constraint tasks combine requirements such as exact edits, answer prefixes, keywords, sentence counts, and writing style.These constraints are drawn from source datasets including CNN/Daily Mail, XSum, SAMSum, English Gigaword, arXiv, OpenSubtitles, TED Talks, News-Commentary, ROCStories, and WritingPrompts.Source: §A.4 · pp. 14–15

B Detailed Experimental Results

The detailed experiments report results across five difficulty levels for each constraint category, including a mixed-constraints setting.

C Case Study

The case study compares GPT-4 with WizardLM-13B-V1.5 on a level-5 Content Constraints instruction and illustrates model-based evaluation.

  • GPT-4 satisfies all five specified constraints in the level-5 Content Constraints case, while WizardLM-13B-V1.5 fails the third.The failed constraint requires the output animals to be able to swim.Source: C Case Study · p. 15
  • WizardLM-13B-V1.5 fails the constraint requiring that the output animals be able to swim.Source: C Case Study · p. 15
  • The two cases validate the effectiveness of the proposed model-based evaluation.Source: C Case Study · p. 15

D FollowBench-zh

FollowBench-zh extends the benchmark to Chinese with 790 instructions from over 50 NLP tasks and evaluates 13 LLMs across five difficulty levels. Performance generally declines with difficulty, GPT-4 stands out, and substantial room for improvement remains.

  • FollowBench-zh contains 790 curated Chinese instructions from more than 50 NLP tasks, covering closed- and open-ended questions.Its data-generation process is analogous to that used for the English benchmark.Source: D FollowBench-zh · p. 15
  • Across five difficulty levels, nearly all evaluated models perform worse as constraints accumulate from L1 to L5.The benchmark evaluates 13 popular LLMs on FollowBench-zh.Source: D FollowBench-zh · p. 15, Table 16 · p. 21
  • GPT-3.5 shows a larger instruction-following gap relative to GPT-4 on FollowBench-zh than on FollowBench.Source: D FollowBench-zh · p. 15
  • GPT-4 stands out across constraint categories in the FollowBench-zh evaluation.Figure 11 reports average HSR scores across difficulty levels for each category.Source: D FollowBench-zh · p. 15, Figure 11 · p. 22
  • Baichuan2-Chat-7B and ChatGLM3-6B perform comparably to or better than their open-source counterparts.The paper connects this pattern to pretraining on combined English and Chinese corpora.Source: D FollowBench-zh · p. 15
  • Several models not specifically trained on Chinese corpora can nevertheless understand and respond in Chinese.The passage names LLaMA2-Chat, WizardLM, and Vicuna variants as examples.Source: D FollowBench-zh · p. 15
  • The Chinese benchmark uses prompt templates for Content, Style, and Format open-ended question answering and model-based evaluation.Source: Figure 7 · p. 16, Figure 8 · p. 16, Figure 9 · p. 16
  • FollowBench-zh results are organized by constraint category and model scale, with tables covering Content, Situation, Style, Format, Example, and Mixed Constraints.The tables distinguish proprietary models from large, medium, and small open-sourced LLMs.Source: Table 7 · p. 17, Table 8 · p. 17, Table 9 · p. 17
Loading 2310.20410v3…