PaperlayerView exact revision on arXiv

Source-linked AI summary

Chain-of-Thought Hub: A Continuous Effort to Measure Large Language Models' Reasoning Performance

Yao Fu, Litu Ou, Mingyu Chen, Yuhao Wan, Hao Peng, Tushar Khot

arXiv:2305.17306v1cs.CLcs.AIcs.LG

TL;DR

Evaluating LLMs is challenging, while complex reasoning is presented as a way to distinguish model capabilities and as relevant to their prospective computational-platform role. CoT Hub compiles reasoning benchmarks into a continuously maintained open-source evaluation suite. Its results show scale-related performance differences and identify model-family comparisons and development directions for open models.

  • Problem

    Evaluating LLMs is challenging, and complex reasoning is presented as a potential differentiator between weaker and stronger models and a capability relevant to their envisioned computational-platform role.

    Source: §1 · p. 1

  • Method

    CoT Hub is a continuous open-source evaluation effort that uses a curated suite of reasoning benchmarks to track LLM capability.

    Source: §1 · p. 1, §2 · p. 2

  • Results

    Reasoning performance correlates with model scale; as of May 2023, PaLM and Claude 2 were comparable to but slightly worse than the GPT family, and LLaMA 65B performed closely to code-davinci-002.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The authors identify building better base models and exploring RLHF as directions for open-source efforts, and say better SFT and RLHF could potentially bring LLaMA 65B to ChatGPT-3.5 performance.

    Source: §4 · p. 4

    Some non-public models remain untested on some datasets, so the reported evaluation is not comprehensive.

    Source: §3 · p. 3

Abstract

from arXiv · show

As large language models (LLMs) are continuously being developed, their evaluation becomes increasingly important yet challenging. This work proposes Chain-of-Thought Hub, an open-source evaluation suite on the multi-step reasoning capabilities of large language models. We are interested in this setting for two reasons: (1) from the behavior of GPT and PaLM model family, we observe that complex reasoning is likely to be a key differentiator between weaker and stronger LLMs; (2) we envisage large language models to become the next-generation computational platform and foster an ecosystem of LLM-based new applications, this naturally requires the foundation models to perform complex tasks that often involve the composition of linguistic and logical operations. Our approach is to compile a suite of challenging reasoning benchmarks to track the progress of LLMs. Our current results show that: (1) model scale clearly correlates with reasoning capabilities; (2) As of May 2023, Claude-v1.3 and PaLM-2 are the only two models that are comparable with GPT-4, while open-sourced models still lag behind; (3) LLaMA-65B performs closely to code-davinci-002, indicating that with successful further development such as reinforcement learning from human feedback (RLHF), it has great potential to be close to GPT-3.5-Turbo. Our results also suggest that for the open-source efforts to catch up, the community may focus more on building better base models and exploring RLHF.

1. Introduction

The introduction argues that complex reasoning can distinguish stronger from weaker language models and matters for their envisioned role as a computational platform. It introduces CoT Hub as a continuously updated open-source benchmark suite and reports broad model-performance patterns.

1. Introduction

  • The introduction identifies complex reasoning as a potential differentiator between stronger and weaker LLMs, drawing on observations about GPT and PaLM models.It points to the distinction between GPT-3.5 and GPT-4 becoming clearer as task complexity rises, and to stronger chain-of-thought reasoning in larger PaLM models.Source: §1 · p. 1
  • The authors motivate measuring reasoning by the prospect of LLMs becoming a computational platform for applications that require complex tasks.They cite LangChain as an example of applications built on backend LLM engines.Source: §1 · p. 1
  • CoT Hub is a continuous open-source effort that tracks LLM reasoning with a curated suite, comparing 19 major models across six benchmarks and more than 100 subtasks.The suite includes bilingual reasoning in Chinese and is being expanded with new models and datasets.Source: §1 · pp. 1–2
  • Reasoning performance correlates with model scale; as of May 2023, PaLM and Claude 2 were comparable to, but slightly worse than, the GPT family, while LLaMA 65B performed closely to code-davinci-002.The authors suggest that successful SFT and RLHF could potentially bring LLaMA 65B to ChatGPT-3.5 performance.Source: §1 · p. 2

2. Method

The method compiles reasoning datasets intended to reflect LLM use and distinguish stronger from weaker models, then evaluates them with few-shot chain-of-thought prompting. It uses final-answer accuracy as a proxy for reasoning and contrasts its focus with other evaluation suites.

2. Method

  • The dataset collection is intended to reflect actual LLM use and clearly differentiate stronger from weaker models.The paper first describes test-data collection, then how model performance is obtained.Source: §2 · p. 2
  • GSM8k tests arithmetic reasoning and composing math steps in language, while MATH covers advanced math and science reasoning across seven categories.GSM8k contains 8,000 problems; MATH contains 12,000 competition problems.Source: §2 · p. 2
  • MMLU tests high-school and college-level knowledge and reasoning, while C-Eval spans 52 disciplines and four difficulty levels in Chinese.MMLU covers 57 subjects; C-Eval comprises 13,000 multiple-choice questions.Source: §2 · p. 2
  • BigBench Hard contains language and symbolic reasoning tasks suited to chain-of-thought prompting, and HumanEval tests coding with handwritten Python problems.BigBench Hard has 6,500 problems across 23 subsets; HumanEval has 164 problems.Source: §2 · p. 2
  • The evaluation uses few-shot chain-of-thought prompting, rather than answer-only prompting, and favors few-shot over zero-shot to accommodate pretrained as well as instruction-tuned checkpoints.The paper notes that zero-shot prompting may underestimate pretrained checkpoints.Source: §2 · p. 2
  • Unlike HeLM, Chatbot Arena, and Open LLM Leaderboard, CoT Hub focuses on reasoning, uses chain-of-thought prompting for most results, and considers both open and closed models.The paper characterizes HeLM as broader in task coverage, Chatbot Arena as measuring dialog preference, and Open LLM Leaderboard as focused on open models.Source: §2 · p. 2
  • Final-answer accuracy serves as the reasoning proxy because intermediate-step correctness is reported to correlate strongly with final accuracy.The evaluation does not score intermediate steps directly.Source: §2 · p. 2

3. Experiments

The suite compares popular model families on reasoning benchmarks, with results showing an open/closed-source gap and performance correlated with scale. It also reports RLHF among most leading models and highlights LLaMA 65B’s performance relative to code-davinci-002.

3. Experiments

  • The evaluated model families include GPT, Claude, PaLM, LLaMA, and T5.Table 1 reports overall model performance on the Chain-of-Thought Hub suite.Source: §3 · p. 2, Table 1 · p. 3
  • The results show a gap between leading GPT, Claude, and PaLM models and open-source LLaMA and FlanT5 models.Among open-source models, LLaMA 65B performs close to code-davinci-002.Source: Table 1 · p. 3, §3 · p. 4
  • Results are taken from prior papers when available and otherwise tested by the authors, but some non-public models remain untested on some datasets.The authors describe the evaluation as non-comprehensive and the hub as a continuous effort.Source: §3 · p. 3
  • Most of the top six leaderboard models have undergone reinforcement learning from human feedback, with PaLM-2 the exception.The authors identify RLHF as an area for further community study.Source: §3 · p. 4
  • Model performance generally correlates with scale, approximately following a log-linear trend, while models not disclosing scale generally perform better.The figure plots performance against the log of model scale measured in billions of parameters.Source: §3 · p. 4, Figure 1 · p. 4
  • On MMLU, LLaMA 65B is close to code-davinci-002; on GSM8k, it is worse but close and substantially better than other open-source models.The authors suggest that correctly applying RLHF to LLaMA 65B may bring it close to ChatGPT.Source: §3 · p. 4

4. Conclusion and Future Work

CoT Hub reports differences in reasoning performance between smaller and larger models and between closed-source and open-source models. The conclusion identifies directions for improving open-source models and plans to expand the evaluation platform.

  • CoT Hub shows reasoning-performance differences between smaller and larger models and between closed-source and open-source models.Source: §4 · p. 4
  • The authors identify building better base models and exploring RLHF as directions for improving open-source models.They also suggest that LLaMA 65B could perform on par with ChatGPT-3.5 if aligned with better SFT and RLHF.Source: §4 · p. 4
  • Future work plans to expand CoT Hub’s datasets and model coverage, explore API-assisted MATH solving, and use the platform to guide open-source LLM development.Planned datasets include commonsense reasoning, math theorem proving, and API calling; proposed additions include more models.Source: §4 · p. 4

Related papers

Loading 2305.17306v1…