Source-linked AI summary
Evaluating the Logical Reasoning Ability of ChatGPT and GPT-4
Hanmeng Liu, Ruoxi Ning, Zhiyang Teng, Jian Liu, Qiji Zhou, Yue Zhang
TL;DR
Logical reasoning remains difficult for natural language understanding systems, motivating evaluation of ChatGPT and GPT-4 across reasoning-heavy benchmarks. The paper tests both models on multi-choice reading comprehension and natural language inference, and releases LogiEval as a prompt-based benchmark suite. Both models perform relatively well on well-known benchmarks but struggle with newly released and out-of-distribution data, especially logical reasoning NLI tasks.
Problem
Achieving human-like logical reasoning capabilities in natural language understanding systems remains challenging despite progress in probing and downstream tasks.
Method
The paper evaluates ChatGPT and GPT-4 on multi-choice reading comprehension and natural language inference benchmarks using instruction prompts, and introduces LogiEval.
Results
Both ChatGPT and GPT-4 outperform traditional fine-tuning on most logical reasoning benchmarks, while facing challenges on new, out-of-distribution, and logical reasoning NLI datasets.
Takeaways & Limitations
Logical reasoning remains challenging for ChatGPT and GPT-4, particularly on out-of-distribution and natural language inference datasets.
Takeaways & Limitations
A limited sample of GPT-4 errors does not establish that the model is incompetent at reasoning questions, although occasional wrong answers indicate room for improvement.
Abstract
from arXiv · showhide
Harnessing logical reasoning ability is a comprehensive natural language understanding endeavor. With the release of Generative Pretrained Transformer 4 (GPT-4), highlighted as "advanced" at reasoning tasks, we are eager to learn the GPT-4 performance on various logical reasoning tasks. This report analyses multiple logical reasoning datasets, with popular benchmarks like LogiQA and ReClor, and newly-released datasets like AR-LSAT. We test the multi-choice reading comprehension and natural language inference tasks with benchmarks requiring logical reasoning. We further construct a logical reasoning out-of-distribution dataset to investigate the robustness of ChatGPT and GPT-4. We also make a performance comparison between ChatGPT and GPT-4. Experiment results show that ChatGPT performs significantly better than the RoBERTa fine-tuning method on most logical reasoning benchmarks. With early access to the GPT-4 API we are able to conduct intense experiments on the GPT-4 model. The results show GPT-4 yields even higher performance on most logical reasoning datasets. Among benchmarks, ChatGPT and GPT-4 do relatively well on well-known datasets like LogiQA and ReClor. However, the performance drops significantly when handling newly released and out-of-distribution datasets. Logical reasoning remains challenging for ChatGPT and GPT-4, especially on out-of-distribution and natural language inference datasets. We release the prompt-style logical reasoning datasets as a benchmark suite and name it LogiEval.
1 Introduction
The paper evaluates ChatGPT and GPT-4 on logical reasoning benchmarks spanning multi-choice reading comprehension and natural language inference, introducing LogiEval as a prompt-based test suite. It finds strong performance on established benchmarks but persistent difficulty with newly released, out-of-distribution, and reasoning-intensive NLI datasets.
- The evaluation draws on established logical reasoning benchmarks such as LogiQA and ReClor, which are based on challenging reasoning examinations.
- Human-like logical reasoning remains challenging for natural language understanding systems despite progress in probing and downstream tasks.
- The study evaluates ChatGPT and GPT-4 on multi-choice reading comprehension and natural language inference benchmarks requiring logical reasoning.
- The authors introduce LogiEval, a test suite for testing prompt-based large language models.
- Both models perform well on well-known logical reasoning reading comprehension benchmarks but struggle with out-of-distribution datasets and logical reasoning NLI tasks.
2 Evaluation Settings
The evaluation tests ChatGPT and GPT-4 on logical reasoning through multi-choice reading comprehension and natural language inference benchmarks, including established, newly released, and out-of-distribution datasets. It uses RoBERTa-base as a fine-tuning baseline and instruction-prompted model evaluations.
- Task settings: The evaluation covers multi-choice reading comprehension and natural language inference tasks requiring logical reasoning.The reading-comprehension setting asks systems to select the most appropriate answer from candidate options, while NLI classifies the relationship between a premise and hypothesis.
- Multi-choice reading comprehension: The reading-comprehension benchmarks include LogiQA, ReClor, and AR-LSAT, spanning civil-service and LSAT-derived logical reasoning questions.LogiQA uses Chinese Civil Service Examination questions; ReClor uses LSAT examples; AR-LSAT contains ordering, grouping, and assignment games.
- Natural language inference: The NLI evaluation includes ConTRoL, MED, HELP, ConjNLI, TaxiNLI, and MNLI, covering contextual, monotonicity, conjunction, and other logical categories.ConTRoL contains 36.2% premise-hypothesis pairs categorized as logical reasoning, while MED and HELP focus on monotonicity reasoning.
- Out-of-distribution data: The study constructs a LogiQA 2.0 out-of-distribution dataset from Chinese Civil Servant Exams released from 2022 onward.The test set consists of logical reasoning tests designed by experts and is released with the authors’ data.
3 Results
ChatGPT and GPT-4 perform relatively well on established logical-reasoning benchmarks, but their performance declines on out-of-distribution data and remains limited on logical-reasoning NLI tasks. GPT-4 generally exceeds ChatGPT on reading-comprehension benchmarks, while NLI results are mixed and instruction following remains challenging.
- Multi-choice reading comprehension: 38.44% accuracy was achieved by ChatGPT on LogiQA 2.0 out-of-distribution data, below RoBERTa-base and its performance on established benchmarks.ChatGPT also scored only 20.42% on AR-LSAT.
- Multi-choice reading comprehension: 33.48% accuracy was achieved by GPT-4 on AR-LSAT, showing a substantial decline on this newly released dataset.GPT-4 scored 58.49% on LogiQA 2.0 out-of-distribution data, below its LogiQA 2.0 test-set performance.
- Natural language inference: 58.45% accuracy was achieved by ChatGPT on ConTRoL, outperforming RoBERTa-base by nearly 10 percentage points on a logical-reasoning NLI dataset.ChatGPT also exceeded RoBERTa on ConjNLI, HELP, MED, and TaxiNLI, but the reported margins were small overall.
- Natural language inference: 89.42% accuracy was achieved by GPT-4 on MED, while its NLI performance varied across datasets and was lower than ChatGPT on ConTRoL.GPT-4 scored 72.71% on ConjNLI, 46.01% on HELP, and 60.08% on TaxiNLI.
- Natural language inference: GPT-4 performed less strongly on logical-reasoning NLI than on multi-choice reading comprehension and did not output NLI labels steadily despite natural-language instructions.The authors infer that GPT-4 follows instructions better for multi-choice reading comprehension than for NLI.
4 Analysis
The analysis finds that ChatGPT and GPT-4 outperform RoBERTa on most logical reasoning benchmarks, but GPT-4 remains vulnerable to unfamiliar data and reasoning errors. Additional tests show benefits from conversational context and zero-shot Chain-of-Thought prompting.
- Benchmark performance: ChatGPT and GPT-4 surpass RoBERTa on most logical reasoning benchmarks, including LogiQA, ReClor, and AR-LSAT.The models nevertheless show noticeable performance drops on out-of-distribution datasets.
- Answer and Reason: GPT-4’s incorrect reasoning can result from overlooking relevant dimensions of a problem, such as fossil location rather than chronology.In the cited example, failure to attend to fossil location leads to the wrong answer.
- Answer and Reason: Four of GPT-4’s first 10 LogiQA errors were logical errors, three were scope errors, and three involved unresolved semantic ambiguity.The authors caution that this limited sample does not establish general incompetence at reasoning questions.
- In-Context Learning: GPT-4 tends to produce more correct answers within the same conversation after approximately eight rounds, even without feedback.The result is presented as evidence of improved in-context performance after seeing more examples.
- In-Context Learning: On 20-instance tests, GPT-4 answered 9 versus 5 LogiQA 2.0 OOD questions and 13 versus 7 ConTRoL questions correctly with versus without context.The comparison tests examples in one conversation window against separate new conversation windows.
5 Conclusion
The conclusion evaluates ChatGPT and GPT-4 across logical reasoning datasets covering multi-choice reading comprehension and natural language inference. Both models outperform traditional fine-tuning on most benchmarks but remain challenged by new, out-of-distribution, and reasoning-intensive NLI data.
- 5 Conclusion: ChatGPT and GPT-4 were evaluated on multiple logical reasoning datasets spanning multi-choice reading comprehension and natural language inference.The evaluation included popular, less-known, and newly released datasets.
- 5 Conclusion: ChatGPT and GPT-4 outperform the traditional fine-tuning method on most logical reasoning benchmarks.The conclusion reports this pattern across the evaluated benchmarks rather than for a single dataset.
- 5 Conclusion: Both models face challenges with new and out-of-distribution data and do not perform astonishingly well on natural language inference requiring logical reasoning.These difficulties remain despite stronger performance on many logical reasoning benchmarks.
A Instruction and Prompt for the multi-choice reading comprehension task
The multi-choice reading comprehension prompt presents a passage and question, then requires selecting exactly one of four answer options. The model must output A, B, C, or D according to the selected option.
- A Instruction and Prompt for the multi-choice reading comprehension task: The task requires choosing the only correct answer from four options after reading a passage and question.The output is encoded as A, B, C, or D according to the option’s position.
- A Instruction and Prompt for the multi-choice reading comprehension task: The instructions require reading the question and options thoroughly before selecting the answer.
B Instruction and Prompt for the natural language inference task
The natural language inference prompt asks whether a hypothesis follows from a premise. It maps the judgment to entailment, contradiction, or insufficient information labels.
- B Instruction and Prompt for the natural language inference task: The task asks whether a hypothesis is entailed by the information contained in a premise.The model must decide among logical entailment, falsity, and indeterminacy.
- B Instruction and Prompt for the natural language inference task: The response labels are e for entailment, c for logical falsity, and n when the premise cannot determine the hypothesis.
C GPT-4 Answers with/without Context
Without context, GPT-4 selected option B, while with context it selected option A; both options were presented as analogous to the original argument’s logical flaw.
- Logical comparison: The original question asks which option exhibits the most similar logical loophole.The target pattern concerns treating a risk as unimportant because the adverse outcome does not occur in most cases.
- With context: With context, GPT-4 selected A, applying the same loophole to obesity and heart-disease risk in China.Both arguments move from higher conditional risk to dismissing the need to emphasize or distinguish the associated danger.
- Without context: Without context, the answer was B, which parallels the original argument by downplaying increased risk because only a minority experience the negative outcome.The reasoning compares gun-related homicide risk with promiscuous-sex AIDS risk.