Source-linked AI summary
InFoBench: Evaluating Instruction Following Ability in Large Language Models
Yiwei Qin, Kaiqiang Song, Yebowen Hu, Wenlin Yao, Sangwoo Cho, Xiaoyang Wang, Xuansheng Wu, Fei Liu, Pengfei Liu, Dong Yu
TL;DR
Instruction-following in LLMs lacks systematic evaluation focused on accurately understanding and executing complex directives. The paper introduces DRFR and INFOBENCH to decompose instructions into checkable requirements, finding that advanced models still struggle with complex scenarios while GPT-4 offers a cost-effective annotation option.
Problem
Instruction-following—the capacity to accurately understand and execute user instructions—has not been thoroughly explored or systematically evaluated.
Method
The paper introduces DRFR, which decomposes instructions into distinct criteria, and INFOBENCH, a benchmark of 500 instructions paired with 2,250 decomposed questions.
Results
Advanced LLMs show significant progress but retain a notable gap in perfectly following complex instructions, while GPT-4 serves as an accurate and cost-effective evaluator.
Takeaways & Limitations
DRFR and INFOBENCH provide clearer insight into model performance and identify numerical and linguistic understanding as areas for improvement.
Takeaways & Limitations
Human annotations covered only 10% of the instruction set, so reliability and annotation comparisons are based solely on 50 instructions.
Abstract
from arXiv · showhide
This paper introduces the Decomposed Requirements Following Ratio (DRFR), a new metric for evaluating Large Language Models' (LLMs) ability to follow instructions. Addressing a gap in current methodologies, DRFR breaks down complex instructions into simpler criteria, facilitating a detailed analysis of LLMs' compliance with various aspects of tasks. Alongside this metric, we present InFoBench, a benchmark comprising 500 diverse instructions and 2,250 decomposed questions across multiple constraint categories. Our experiments compare DRFR with traditional scoring methods and explore annotation sources, including human experts, crowd-sourced workers, and GPT-4. The findings demonstrate DRFR's higher reliability and the effectiveness of using GPT-4 as a cost-efficient annotator. The evaluation of several advanced LLMs using this framework reveals their strengths and areas needing improvement, particularly in complex instruction-following. This study contributes a novel metric and benchmark, offering insights for future LLM development and evaluation.
1 Introduction
The paper addresses the limited systematic evaluation of LLM instruction following by proposing DRFR and INFOBENCH for granular assessment. Experiments find DRFR more reliable than direct scoring, GPT-4 a cost-effective annotator, and persistent weaknesses on complex instructions.
- 1 Introduction: DRFR decomposes complex instructions into distinct criteria, enabling detailed and interpretable measurement of compliance with individual requirements.The metric is designed to provide a clearer assessment of instruction-following performance in complex scenarios.
- 1 Introduction: INFOBENCH contains 500 instructions and 2,250 decomposed questions covering Content, Linguistic, Style, Format, and Number constraints.The benchmark supports fine-grained analysis of model deficiencies across constraint categories.
- 1 Introduction: DRFR achieved higher annotator consensus than Direct Scoring, particularly on the Hard Set, indicating greater reliability for instruction-following evaluation.The comparison was conducted across responses from various LLMs.
- 1 Introduction: GPT-4 was identified as an accurate, cost-effective, and time-efficient alternative to human annotation, especially with structured multi-turn prompts.Human experts were more accurate but also more costly and time-consuming.
- 1 Introduction: Evaluation of six advanced LLMs found substantial progress but a persistent gap in perfect instruction following, especially for complex Hard Set scenarios.Performance differences across constraint types and domains point to challenges in numerical and linguistic understanding.
2 INFOBENCH
INFOBENCH combines Easy and Hard instruction sets with decomposed binary requirements to evaluate instruction following across diverse domains and constraint types. Its Hard Set uses more elaborate instructions and requirements, while its constraints are more heterogeneous than those in the Easy Set.
- 2 INFOBENCH: INFOBENCH comprises Easy and Hard sets, with the Hard Set manually curated across diverse subject areas and intended to distinguish stronger from weaker models.The Hard Set spans 72 domains, including sciences, social sciences, engineering, economics, arts, occupational tasks, and daily life.
- 2 INFOBENCH: The benchmark’s instructions and decomposed requirements are crafted by one subject-matter expert and validated by another for naturalness, atomicity, and clarity.Collection continues until both experts agree on these properties.
- 2 INFOBENCH: Each instruction is decomposed into binary YES/NO questions, and requirements are manually labeled using an ontology of constraint types.A requirement may receive multiple constraint labels.
- 2 INFOBENCH: The Hard Set contains more elaborate instructions and requirements than the Easy Set, reflecting greater complexity and broad domain diversity.Table 2 reports statistics for instruction count, length, requirements, requirements per instruction, and domains.
- 2.2.3 Benchmark Statistics: Approximately two-thirds of Easy Set constraints are content constraints, whereas the Hard Set has a more heterogeneous distribution across constraint types.Figure 2 compares constraint-type distributions between the two sets.
3 Experiments
The experiments compare DRFR with direct scoring and evaluate human, crowd-sourced, and GPT-4 annotation sources. DRFR achieves higher annotator consensus, while GPT-4 offers a cost- and time-efficient alternative with strong agreement against expert-derived labels.
- 3.1 DRFR v.s. Direct Scoring: The evaluation compares five LLMs using Direct Scoring and DRFR, with three NLP experts assessing responses in both methods.Scores are standardized at the instruction level for direct comparison.
- 3.1 DRFR v.s. Direct Scoring: A gap exceeding 0.19 in Pairwise Kappa Agreement indicates higher annotator consensus for DRFR than Direct Scoring across 50 INFOBENCH instructions.The sample contains 25 Easy and 25 Hard instances.
- 3.2 Experiment of Annotation Sources: Pairwise label distance measures annotation disagreement with ground-truth model-order labels, assigning values from 0 for accurate ordering to 2 for reversed ordering.Weighted Pairwise Label Distance aggregates these distances across instructions.
- 3.2 Experiment of Annotation Sources: GPT-4 annotations achieve 89% accuracy and a 0.44 weighted pairwise label distance when using decomposed questions.These results align with the gold standard used for comparison.
- 3.2 Experiment of Annotation Sources: GPT-4 annotations are more cost-effective and time-efficient than AMT annotations while showing stronger overall annotation performance.The study compares GPT-4 with human experts and Amazon Mechanical Turk workers using cost, time, and robustness criteria.
4 Automatic Evaluation
GPT-4 was used to scale DRFR evaluation across six LLMs, revealing systematic differences by constraint type and domain. The results show stronger performance on Content and Style constraints, with harder challenges on Number and Linguistic constraints.
- GPT-4 was integrated with a multi-turn prompt to evaluate six advanced LLMs across INFOBENCH.
- The automated evaluation was reported in a table covering six selected LLMs on INFOBENCH.
- Performance was highest on Content and Style constraints, intermediate on Format, and lowest on Number and Linguistic constraints across all six models.
- Across seven Hard Set domains, evaluation scores formed an approximately circular pattern except for Vicuna, indicating relatively similar performance across domains.
5 Related Work
Related work spans instruction and alignment tuning, LLM-based evaluation, and concurrent benchmarks for instruction following. Existing studies differ in their evaluation targets and in how they construct fine-grained constraints.
- Instruction tuning uses annotated instructional data, whereas alignment tuning uses human feedback to align model behavior with human values and preferences.
- LLMs have also been used to evaluate natural-language generation across numerous tasks, motivating their use as evaluators in this work.
- Other related studies examine scientific reasoning, judge models, verbalizer manipulation, and automatic win-rate evaluation.
- Concurrent benchmarks differ by constructing multi-level instructions through sequential constraints or by using 25 verifiable instructions, rather than decomposing one instruction into multiple constraints.
6 Conclusion
The paper concludes that DRFR and INFOBENCH enable more nuanced instruction-following evaluation, while GPT-4 provides an accurate, efficient, and cost-effective annotation alternative. Evaluations still identify challenges with complex instructions.
- DRFR offers clearer performance insights than traditional evaluation methods, especially for complex instruction-following scenarios.
- GPT-4 emerged as a cost-effective, accurate, and efficient annotator for large-scale evaluation.
- The evaluated LLMs show substantial progress, particularly closed-source models, but still need improvement on complex instructions, numerical reasoning, and linguistic comprehension.
Limitations
The study’s evidence is limited by restricted human annotation coverage, a modest manually constructed benchmark, and omission of truthfulness and harmlessness from the evaluation.
- Human annotations covered only 10% of the instruction set, so protocol reliability and annotation comparisons rely on 50 instructions.
- INFOBENCH contains 500 instructions and 2,250 decomposed questions, and manual instruction writing limits substantial scaling.
- The evaluation focuses on explicit instruction intentions and excludes truthfulness and harmlessness.
A.1 Instruction Dataset Statistics
The Hard Set contains longer instructions and more decomposed questions per instruction than the Easy Set, indicating greater constraint complexity.
- Hard Set instructions average around 50 words, whereas Easy Set instructions predominantly range from 10–20 words.
- Hard Set instructions typically contain 5–7 questions, compared with approximately 2–3 questions in the Easy Set.
- Table 6 categorizes questions by constraint type across the Easy and Hard Sets, documenting their distribution and complexity gradation.
A.2 Prompt Design for GPT-4 Automatic Evaluation
The automatic-evaluation prompt asks GPT-4 to judge decomposed requirements sequentially with binary answers, while few-shot examples did not significantly improve results.
- GPT-4 evaluates each decomposed question with a YES or NO response using the input, generated text, and question.
- The evaluation prompt omits the original instruction because the decomposed questions are hypothesized to capture its requirements.
- Adding few-shot examples produced no significant improvement, possibly because instructions varied and the answer format was simple.
A.3 Human Annotation Guidance
The study uses structured materials and annotation interfaces to guide human evaluation of generated responses across systems and constraint types.
- The Amazon Mechanical Turk interface and annotation guidelines support the crowd-sourced review process.
- The dataset and evaluation materials distinguish instruction content, generated responses, decomposed questions, and constraint-type labels.
- Table 7 provides an example of the multi-turn GPT-4 evaluation prompt, while Table 8 reports automated DRFR results for six selected LLMs.
- Annotators review an input, generated text, and questions answered with YES, NO, or UNKNOWN under stated guidelines.
A.4 GPT-4-1106 Evaluation Results
Evaluation with gpt-4-1106-preview reproduces earlier performance trends while exposing recurring difficulties in numerical, linguistic, spatial, and prerequisite-dependent requirements.
- gpt-4-1106-preview largely reproduces earlier evaluation trends, although its scores are marginally lower than gpt-4-0314.
- Around 30%–40% of questions show medium to high human disagreement, and these questions are excluded when assessing GPT-4 annotation errors.
- GPT-4 annotations can be misled by content that repeats instruction wording, while human disagreement and annotation errors remain recognized sources of discrepancy.
- GPT-4 can misjudge requirements involving exact counting, unique categories, spatial dimensions, or omitted prerequisite details.
- None of the evaluated models fully satisfy the fourth and fifth Number constraints, highlighting difficulty with numerical reasoning.
- None of the models completely satisfy the third and fourth Linguistic constraints, although GPT-4-1106 attempts the complex word-linking rule.