LLMs have revolutionized many areas (e.g. natural language processing, software engineering, etc.) by achieving state-of-the-art performance on extensive downstream tasks. Aiming to achieve robust and general artificial intelligence, there has been a surge of interest in investigating the reasoning ability of the LLMs. Whereas the textual and numerical reasoning benchmarks adopted by previous works are rather shallow and simple, it is hard to conclude that the LLMs possess strong reasoning ability by merely achieving positive results on these benchmarks. Recent efforts have demonstrated that the LLMs are poor at solving sequential decision-making problems that require common-sense planning by evaluating their performance on the reinforcement learning benchmarks. In this work, we conduct an in-depth assessment of several state-of-the-art LLMs' reasoning ability based on the inductive logic programming (ILP) benchmark, which is broadly recognized as a representative and challenging measurement for evaluating logic program induction/synthesis systems as it requires inducing strict cause-effect logic to achieve robust deduction on independent and identically distributed (IID) and out-of-distribution (OOD) test samples. Our evaluations illustrate that compared with the neural program induction systems which are much smaller in model size, the state-of-the-art LLMs are much poorer in terms of reasoning ability by achieving much lower performance and generalization using either natural language prompting or truth-value matrix prompting.
This paper evaluates the reasoning abilities of LLMs with a focus on relational reasoning using the Inductive Logic Programming benchmark.
The performance of state-of-the-art LLMs is compared to the Differentiable Logic Machines (DLM), which specializes in mimicking logical rule operations.
Findings show that LLMs, like GPT-4, perform less effectively in complex relational reasoning tasks compared to DLM, but have potential in logic synthesis.
The study tests various prompting strategies but finds that even advanced techniques like chain-of-thought do not consistently enhance LLMs' performance in relational reasoning.
The results suggest that the current capabilities of LLMs in higher-order reasoning are limited, prompting further research in model training for relational reasoning.
The march towards true artificial intelligence has been marked by the development of extensive models capable of processing and understanding human language. LLMs, recognized for their remarkable achievements across various tasks, have significantly impacted the field. However, the fact that LLMs perform well on tasks that use simple reasoning benchmarks might not be enough to conclude they have strong reasoning capabilities. This paper takes a deeper dive into assessing the reasoning abilities of LLMs, particularly focusing on relational reasoning—a critical element of logical and problem-solving skills.
The study makes use of the Inductive Logic Programming (ILP) benchmark, which challenges systems to deduce strict cause-effect logics based on given information. For instance, if provided data on family relations, the task is to infer complex relationships like grandparent or uncle based solely on baseline knowledge such as parent or sibling relations. Performing well in this space indicates a system's relational reasoning prowess, which can be a stepping stone towards developing general artificial intelligence.
To evaluate the LLMs, a suite of state-of-the-art models were put to the test against a neural program induction system known as the Differentiable Logic Machines (DLM), which uses neural networks to mimic logical rule operations. The models were compared in their effectiveness at reasoning about relationships between objects and processing provided specifications as input-output examples.
The findings reveal that while the largest and most advanced LLMs—GPT-4 and GPT-4 Turbo—demonstrate strong performance in some cases, their abilities are notably diminished when facing tasks that require more complex reasoning. This is starkly contrasted by the DLM, which, although smaller in model size, outperformed LLMs across the board.
One component of the study scrutinized the effectiveness of different prompting strategies, including standard natural language prompting, truth-value matrix prompting, and a state-of-the-art technique known as "chain-of-thought" prompting. The study determined that the chain-of-thought approach, although previously successful in enhancing performance on other reasoning tasks, did not consistently improve outcomes in relational reasoning challenges. Additionally, LLMs showed more promise with tasks related to general graph reasoning when presented as truth-value matrices, suggesting they have potential in logic synthesis.
These findings raise important questions about the current capabilities of LLMs when it comes to higher-order reasoning. While they have proven themselves adept at handling a variety of tasks that demand surface-level textual understanding, the leap to robust reasoning remains significant. This exploration also opens up avenues for future research, particularly in the form of model training that better aligns with the requirements of relational reasoning tasks.
Consequently, the study advocates for a more focused approach to evaluating and enhancing the reasoning abilities of LLMs. By pinpointing where these models excel and where they struggle, researchers can direct their efforts toward developing models that not only parse language but also understand and apply the underlying logic more comprehensively.