DARG: Dynamic Evaluation of Large Language Models via Adaptive Reasoning Graph

Summary

The paper introduces a framework, Dynamic Evaluation of Large Language Models via Adaptive Reasoning Graphs (DARG), which dynamically extends benchmarks by generating evaluation data with controlled complexity. This framework addresses key limitations of static benchmarks, such as data contamination and a lack of adaptability to evolving LLM capabilities.

Strengths

1. The method of dynamically generating evaluation data based on adaptive reasoning graphs is promising. It addresses critical limitations of static benchmarks, such as data contamination and the inability to adapt to the evolving capabilities of LLMs. 2. The evaluation encompasses a wide range of LLMs and tasks, offering a thorough analysis of how different models perform across various complexity levels. This provides valuable insights into model robustness and generalization capabilities. 3. The paper effectively highlights how biases in LLMs can be exacerbated under complex testing conditions. This is a crucial aspect for developing fairer models and contributes significantly to the ongoing discourse on ethical AI. 4. The use of reasoning graphs to represent the underlying structures of problem-solving processes and the subsequent perturbation to create novel test samples is methodologically sound. This ensures that the generated data retains linguistic diversity and is representative of real-world scenarios.

Weaknesses

1. According to Figure 2, the ranking order among various models remains consistent as the complexity metrics increase. This observation contradicts the conclusion in the introduction (line 58) regarding the unreliable assessment of LLMs' capabilities using static benchmarks. This inconsistency needs to be addressed and clarified. 2. While the paper demonstrates the applicability of DARG across four reasoning tasks, it is unclear how well this approach generalizes to other types of tasks (e.g., knowledge based QA), particularly those that do not naturally lend themselves to graph-based representations. 3. The graph extraction and data generation process heavily relies on closed-source LLMs, such as GPT-4. Although rule-based constraints and data verification modules are incorporated, the dependence on proprietary models raises questions about the reproducibility of the proposed framework.

Questions

1. Can the authors provide more detailed explanations or additional analyses to reconcile the observed inconsistencies between the model rankings and the stated conclusions? 2. How does the DARG framework perform when applied to other non-reasoning tasks? Can the authors provide preliminary results or case studies on different task domains?

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

The authors have adequately addressed the limitations

Summary

This paper proposed a dynamic evaluation of LLMs --- DARG. The authors first generate a reasoning graph of the problem, then perturb the problem's complexity along various "dimensions", then convert the more complicated graph back to natural language questions. The authors evaluate several LLMs on 4 perturbed datasets and observe consistent performance drop, indicating that LLMs may not reason very well, and previous good static evaluation results may be due to data contamination.

Strengths

This paper tries to tackle an important evaluation issue in LLMs by providing dynamic yet controlled method.

Weaknesses

The particular method only applies to problems that have a clear reasoning graph and it relies on a rule-based system and a non-ambiguous way to increase complexity, which may not be immediate clear how to apply to any dataset. However, it is a good starting point.

Questions

1. For GSM8K, when you increase the numerical complexity by +k, does that mean you sample numbers of which the range is increased by k? or you simply +k to all the numbers? 2. For figure 9, can you also add an overlapping radar plot such that your claims in line 195-199 is more easily visualized? 3. Line 203-208, can you explain how the evaluation is done? Did you ask the model being evaluated to generate a reasoning graph and compare with the ground truth? 4. For BBQ, how is the attributes' polarity determined? Do you have a predetermined set of attributes to sample from? 5. Line 232-234, I am not quite getting the claim here. Figure 4 the lower right subfigure is presenting the avoidance rate, and it seems Mistral 7B has the highest rate rather than GPT4 and Gemini. Also is the Mistral model here aligned? If so, then isn't it against the argument that over-alignment is causing the issue here? 6. Do we know the model performance when the reasoning graph stays the same, but only the question is paraphrased? Misc: - line 200 "Trubo"-> "Turbo"

Rating

6

Confidence

2

Soundness

3

Presentation

3

Contribution

3

Limitations

Limitations have been discussed.

Summary

The paper proposes a new framework, DARG, to extend current reasoning benchmarks with controlled and diversity dynamically. Authors evaluate multiple LLMs on those generated output from DARG and concluded that the proposed method is useful for evaluating LLMs in a dynamic and adaptive way.

Strengths

1. The paper proposes a new method to add control for benchmark augmentation when LLMs are involved, which is useful for evaluating LLMs. 2. The paper conducts experiments on significant amount of LLMs and multiple reasoning categories, indicating the generality of the method. 3. The method is well-motivated and stated.

Weaknesses

1. The paper only involves evaluations for DARG that use ChatGPT as the graph construction and graph-to-text generation engine, which may cause bias on the extended benchmarks. It would be good to see if the selection of the LLM used for those components can affect the final evaluation results. It is even useful to check if replacing those LLM components with humans will lead to a significant change. 2. DARG's rule-based function is not clearly explained in the paper, which seems important to ensure the quality of reasoning graph generation. 3. Due to the uncertainty of the quality of the generated benchmarks, it would be nice to reflect the correct rate evaluated with human eval on the LLM evaluation results with DARG, such as Figure 2. With that, it will be easier to see how confident the results are for the ranking of those LLMs.

Questions

See weakness above

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

See weakness above.

Summary

This paper introduces DARG (Dynamic Evaluation of LLMs via Adaptive Reasoning Graph), a framework for dynamically generating evaluation data for large language models (LLMs) with controlled complexity. The DARG framework constructs reasoning graphs from existing benchmarks, perturbs these graphs to generate more complex samples, and uses LLMs with code verification to filter out incorrect perturbations. The authors apply DARG to generate new test data for 4 reasoning tasks: math, social, spatial, and symbolic reasoning. One of the key experimental results shows that the performance of all LMs generally decreases as complexity increases, potentially indicating that the newly constructed datasets are indeed meaningfully challenging. They also demonstrate that DARG-generated data can be used to improve model performance through fine-tuning.

Strengths

* Novel approach to augment existing static benchmarks, especially the idea of using code generation to filter out bad examples. * Comprehensive evaluation and interesting analysis of performance dropping across benchmarks. * Demonstrates potential of generated data for improving models via fine-tuning

Weaknesses

- Baseline comparisons: Main experimental results show that the generated benchmarks are challenging for llms. However, this isn't the first work introducing the idea of creating harder benchmarks. So, it would be useful to know the gap the proposed work fills. For instance, it would be valuable to see comparisons against other dynamic evaluation methods like DyVal (which the authors do mention in the related work). - The perturbations used, especially for math problems, may be overly simplistic. Simply increasing numerical values or graph complexity doesn't necessarily capture all aspects that make math problems more challenging or interesting.

Questions

1. Have you explored using only open-source models for the graph extraction and data generation steps? How well does DARG work without access to closed-source LLMs? 2. Could you expand on the fine-tuning experiments? Specifically, what happens if we _add_ gsm8k training data to the newly generated samples?

Rating

6

Confidence

5

Soundness

3

Presentation

3

Contribution

2

Limitations

Yes.

Thanks for your response!

Thanks for your response. I'll keep my positive score with an increased confidence in the work.