Benchmark Self-Evolving: A Multi-Agent Framework for Dynamic LLM Evaluation

1 Introduction

• Static benchmarks may be inadequate amid evolving LLM abilities and contamination, while annotating new benchmarks from scratch is costly.The paper therefore updates existing benchmark datasets rather than constructing entirely new ones.Source: §1 · p. 1
• The framework reframes benchmark instances through scalable, robust, and fine-grained evaluation, respectively testing diverse queries, context perturbations, and problem-solving sub-abilities.Robust evaluation includes paraphrasing, noise, and polarity reversal.Source: §1 · p. 2
• A four-component, GPT-4-powered system filters original instances, creates variants, verifies them, and formulates incorrect options for an additional consistency check.Only instances passing verification are used for dynamic evaluation.Source: §1 · p. 2
• Across four tasks, scalable and robust evaluation produces a general performance decline for all models, while fine-grained probing and widened gaps sharpen capability assessment and model selection.The results also reveal limited generalizability and robustness to diverse and complex queries.Source: §1 · p. 2

2.1 Evolving Instance Taxonomy

• For scalable evaluation, alternative questions test different context facets, while more complex questions require additional reasoning without changing the context.Question generation preserves the original context to maintain instance accuracy.Source: §2.1 · p. 2
• Robust evaluation paraphrases or adds irrelevant or adversarial sentences to contexts, while polarity reversal changes key details and requires a corresponding answer change.Paraphrasing and noising retain the original answer label.Source: §2.1 · p. 3
• Fine-grained evaluation probes task planning, implicit knowledge identification, and relevant context retrieval as problem-solving sub-abilities.These questions examine reasoning steps, underlying facts or rules, and pertinent information in the context.Source: §2.1 · p. 3

2.2 Multi-Agent Evolving Instance Generator

• The instance pre-filter selects manageable original examples that GPT-4 can answer correctly, establishing a foundation for subsequent generation.It compares GPT-4’s prediction with the reference answer.Source: §2.2 · p. 3
• The instance creator applies six predefined reframing operations, retaining the original answer when required or inferring a new answer after reformulation.The operations modify contexts or questions while holding the other component fixed.Source: §2.2 · p. 4
• A verifier checks generated instances, while an option formulator supplies an incorrect answer for double-checking and binary-choice evaluation.Only instances judged correct and paired with an incorrect alternative pass the double-verification process.Source: §2.2 · p. 4

3.1 Setup

• The setup extends GSM8K, CLUTRR, StrategyQA, and BoolQ using 100 instances per dataset, with task-specific reframing and two evaluation formats.Scalable and robust evaluations use two-shot CoT; fine-grained evaluations use zero-shot binary-choice questions.Source: §3.1 · p. 5

3.2 Overall Comparison

• Most models show reduced performance on scalable and robust evaluations across the first three tasks; the comparisons include evolving and original results in Table 3.The authors say original results may overestimate proficiency; evolving evaluation also widens performance gaps between models and across tasks.Source: §3.2 · pp. 5–6, Table 3 · p. 6

3.3 Analysis of Varied Reframing Operations

• Question complicating disrupts models most, followed by polarity reversing and question alternating; paraphrasing and noising have limited impact.The operation comparison reports accuracy across datasets against original results in Figure 3.Source: §3.3 · p. 6, Figure 3 · p. 6

3.4 Further Analysis on Sub-Ability

• Fine-grained results reverse the original ChatGLM–ChatGPT ranking, and option-A preference in some models may impair decision-making.GPT-4 performs best and Mistral worst across the three sub-abilities; planning is the weakest skill for all models.Source: §3.4 · p. 7, Figure 4 · p. 7

3.5 Quality of Evolving Instances

• Human annotators judged 110 of 115 sampled generated instances accurate (95.7%), while the system filters nearly 9% of originals and about 24% after initial GPT-4 processing.The reported filtering supports the authors’ double-verification strategy for instance quality.Source: §3.5 · p. 7

3.6 Impact on Data Contamination

• Under original evaluation, both contaminated models improve over vanilla; under dynamic evaluation, their performance gap from vanilla decreases, especially in scalable and fine-grained assessments.Fine-grained results show in-domain contamination outperforming the original model, while direct contamination performs worse.Source: §3.6 · pp. 7–8, Figure 5 · p. 8

LLMs Evaluation

• LLM evaluation includes automatic metrics, human preference, and LLM feedback; automatic evaluation is described as cost-effective for extensive assessment.Examples include automatic metrics (Liang et al., 2022), human preference, and LLM feedback (Liu et al., 2023b; Zhang et al., 2023).Source: §4 · p. 8
• Static benchmarks may be inadequate as LLMs evolve, while training-data overlap with benchmark data risks biased evaluation.The related work frames contamination as raising doubts about whether performance reflects generalization or memorization (Biderman et al., 2023).Source: §4 · p. 8

Data Contamination

• Prior contamination work detects contaminated benchmark data and evaluates on untainted data, while perplexity on newly sampled data may not fully reflect model capabilities.The paper presents its self-evolving framework as a dynamic alternative that can mitigate evaluation bias from contamination.Source: §4 · p. 8

5 Conclusion

• The multi-agent framework enhances existing benchmarks for scalable, robust, and fine-grained evaluation of LLMs.The authors say it provides a more accurate and comprehensive evaluation.Source: §5 · p. 8
• LLMs show a general performance decline and significant performance discrepancies across models and tasks under the framework.The conclusion says these discrepancies help distinguish models’ capabilities.Source: §5 · p. 8
• The authors say the evaluation supports selecting suitable LLMs for specific applications and guiding their further development.Source: §5 · p. 8

Limitations

• Computational limits restricted evaluation to four datasets and 100 selected instances per dataset, yielding nearly 1,600 evolving instances.The study describes this as a limitation on benchmark coverage.Source: Limitations · p. 9
• The evaluation covers three closed-source and two open-source LLMs, and the authors acknowledge that this limits the model scope.They state that further experiments on more LLMs will follow.Source: Limitations · p. 9
• Despite pre-filtering and double verification, the GPT-4-powered system may produce a small number of inaccurate instances, potentially reducing assessment accuracy.Human verification provided evidence of such inaccuracies.Source: Limitations · p. 9

Risks

• For factual datasets such as BoolQ, polarity reversal may introduce counterfactual information into altered contexts.If models inadvertently use these inaccuracies as learning material, their performance and reliability could be negatively affected.Source: Risks · p. 9
• Extensive use of OpenAI APIs to generate evolving instances may increase environmental impact, especially for large benchmarks.The authors note that energy consumption and carbon footprint could be considerable at substantial benchmark size.Source: Risks · p. 9

Ethics Statement

• The authors use publicly available datasets, release reviewed evolving instances, and restrict the framework to ethically guided instance generation while denouncing harmful or inaccurate instances.They say this setting supports transparency and reproducibility.Source: Ethics Statement · p. 9

A.1 Algorithm Design

• The appendix points to Algorithm 1 for pseudocode of the multi-agent evolving-instance system.Source: §A.1 · p. 11

A.2 Prompts of Multi-Agent Evolving Instance Setter

• Tables 4–12 provide prompts for the instance generator, verifier, and candidate option formulator.Source: §A.2 · p. 11

B.1 Dataset Descriptions

• The appendix describes GSM8K, CLUTRR, StrategyQA, and BoolQ as the four datasets used in the evaluation.The descriptions characterize their tasks as mathematical reasoning, kinship reasoning, commonsense yes/no reasoning, and reading comprehension, respectively.Source: §B.1 · p. 11

B.2 Instance Filtering Statistics

• Table 13 reports the percentage of instances filtered by double verification.The supplied passage does not provide the table’s percentages.Source: §B.2 · p. 11

B.3 Selection Bias Analysis

• Selection bias is a preference for particular option IDs; the paper estimates prior probabilities across permutations and uses a sampling approximation because closed-source prediction distributions are inaccessible.The permutation-based debiasing method is described as theoretically eliminating selection bias.Source: §B.3 · p. 11
• The analysis defines xI as default-ordered option IDs paired with option contents permuted by I, and assumes the debiased prediction is invariant to option ordering when estimating the prior.It describes summing logarithms across permutations to compute the prior probability.Source: §B.3 · p. 11
• The analysis identifies Pprior (idi | C) as the prior probability of selecting idi and Pdebiased as the debiased probability of the option content indexed by the permutation.Source: §B.3 · p. 11

• Perplexity quantifies dataset complexity and predictability.Source: §B.4 · p. 14
• Reframed GSM8K datasets show increased perplexity across reframing operations, indicating greater complexity and lower predictability than original instances.Source: Figure 6 · p. 15, §B.4 · p. 15
• The analysis compares the perplexity of newly evolving GSM8K datasets with their original counterparts.Source: §B.4 · p. 15

• In three GSM8K cases, ChatGPT answers the originals but fails evolved versions involving added reasoning, altered data points, or a question about the first reasoning step.Source: §B.5 · p. 15
• The evolving instances expose model vulnerability to perturbations and diminished generalization on more complex problems.Source: §B.5 · p. 15
• They also expose inadequacies in addressing questions that target specific sub-abilities.Source: §B.5 · p. 15

• One instruction-tuning dataset combines 4,000 general Alpaca-GPT4 instances with 4,000 task-training instances, 1,000 each from GSM8K, CLUTRR, StrategyQA, and BoolQ.This dataset simulates in-domain contamination.Source: §B.6 · p. 15
• A separate model is fine-tuned with 400 original benchmark instances to simulate direct contamination.Source: §B.6 · p. 15
• Tables 14–16 show before-and-after instances for Question Complicating, Polarity Reversing, and Planning operations.Source: Table 14 · p. 16, Table 15 · p. 16, Table 16 · p. 16