Have Large Language Models Learned to Reason? A Characterization via 3-SAT Phase Transition

Large Language Models (LLMs) have been touted as AI models possessing advanced reasoning abilities. In theory, autoregressive LLMs with Chain-of-Thought (CoT) can perform more serial computations to solve complex reasoning tasks. However, recent studies suggest that, despite this capacity, LLMs do not truly learn to reason but instead fit on statistical features. To study the reasoning capabilities in a principled fashion, we adopt a computational theory perspective and propose an experimental protocol centered on 3-SAT -- the prototypical NP-complete problem lying at the core of logical reasoning and constraint satisfaction tasks. Specifically, we examine the phase transitions in random 3-SAT and characterize the reasoning abilities of state-of-the-art LLMs by varying the inherent hardness of the problem instances. By comparing DeepSeek R1 with other LLMs, our findings reveal two key insights (1) LLM accuracy drops significantly on harder instances, suggesting all current models struggle when statistical shortcuts are unavailable (2) Unlike other LLMs, R1 shows signs of having learned the underlying reasoning. Following a principled experimental protocol, our study moves beyond the benchmark-driven evidence often found in LLM reasoning research. Our findings highlight important gaps and suggest clear directions for future research.

Valid Text-to-SQL Generation with Unification-based DeepStochLog

Large language models have been used to translate natural language questions to SQL queries. Without hard constraints on syntax and database schema, they occasionally produce invalid queries that are not executable. These failures limit the usage of these systems in real-life scenarios. We propose a neurosymbolic framework that imposes SQL syntax and schema constraints with unification-based definite clause grammars and thus guarantees the generation of valid queries. Our framework also builds a bi-directional interface to language models to leverage their natural language understanding abilities. The evaluation results on a subset of SQL grammars show that all our output queries are valid. This work is the first step towards extending language models with unification-based grammars. We demonstrate this extension enhances the validity, execution accuracy, and ground truth alignment of the underlying language model by a large margin. Our code is available at https://github.com/ML-KULeuven/deepstochlog-lm.

Neurosymbolic Decision Trees

Neurosymbolic (NeSy) AI studies the integration of neural networks (NNs) and symbolic reasoning based on logic. Usually, NeSy techniques focus on learning the neural, probabilistic and/or fuzzy parameters of NeSy models. Learning the symbolic or logical structure of such models has, so far, received less attention. We introduce neurosymbolic decision trees (NDTs), as an extension of decision trees together with a novel NeSy structure learning algorithm, which we dub NeuID3. NeuID3 adapts the standard top-down induction of decision tree algorithms and combines it with a neural probabilistic logic representation, inherited from the DeepProbLog family of models. The key advantage of learning NDTs with NeuID3 is the support of both symbolic and subsymbolic data (such as images), and that they can exploit background knowledge during the induction of the tree structure, In our experimental evaluation we demonstrate the benefits of NeSys structure learning over more traditonal approaches such as purely data-driven learning with neural networks.

The Gradient of Algebraic Model Counting

Algebraic model counting unifies many inference tasks on logic formulas by exploiting semirings. Rather than focusing on inference, we consider learning, especially in statistical-relational and neurosymbolic AI, which combine logical, probabilistic and neural representations. Concretely, we show that the very same semiring perspective of algebraic model counting also applies to learning. This allows us to unify various learning algorithms by generalizing gradients and backpropagation to different semirings. Furthermore, we show how cancellation and ordering properties of a semiring can be exploited for more memory-efficient backpropagation. This allows us to obtain some interesting variations of state-of-the-art gradient-based optimisation methods for probabilistic logical models. We also discuss why algebraic model counting on tractable circuits does not lead to more efficient second-order optimization. Empirically, our algebraic backpropagation exhibits considerable speed-ups as compared to existing approaches.

Relational Neurosymbolic Markov Models

Sequential problems are ubiquitous in AI, such as in reinforcement learning or natural language processing. State-of-the-art deep sequential models, like transformers, excel in these settings but fail to guarantee the satisfaction of constraints necessary for trustworthy deployment. In contrast, neurosymbolic AI (NeSy) provides a sound formalism to enforce constraints in deep probabilistic models but scales exponentially on sequential problems. To overcome these limitations, we introduce relational neurosymbolic Markov models (NeSy-MMs), a new class of end-to-end differentiable sequential models that integrate and provably satisfy relational logical constraints. We propose a strategy for inference and learning that scales on sequential settings, and that combines approximate Bayesian inference, automated reasoning, and gradient estimation. Our experiments show that NeSy-MMs can solve problems beyond the current state-of-the-art in neurosymbolic AI and still provide strong guarantees with respect to desired properties. Moreover, we show that our models are more interpretable and that constraints can be adapted at test time to out-of-distribution scenarios.

NeSyA: Neurosymbolic Automata

Neurosymbolic Artificial Intelligence (NeSy) has emerged as a promising direction to integrate low level perception with high level reasoning. Unfortunately, little attention has been given to developing NeSy systems tailored to temporal/sequential problems. This entails reasoning symbolically over sequences of subsymbolic observations towards a target prediction. We show that using a probabilistic semantics symbolic automata, which combine the power of automata for temporal structure specification with that of propositional logic, can be used to reason efficiently and differentiably over subsymbolic sequences. The proposed system, which we call NeSyA (Neuro Symbolic Automata), is shown to either scale or perform better than existing NeSy approaches when applied to problems with a temporal component.

EXPLAIN, AGREE, LEARN: Scaling Learning for Neural Probabilistic Logic

Neural probabilistic logic systems follow the neuro-symbolic (NeSy) paradigm by combining the perceptive and learning capabilities of neural networks with the robustness of probabilistic logic. Learning corresponds to likelihood optimization of the neural networks. However, to obtain the likelihood exactly, expensive probabilistic logic inference is required. To scale learning to more complex systems, we therefore propose to instead optimize a sampling based objective. We prove that the objective has a bounded error with respect to the likelihood, which vanishes when increasing the sample count. Furthermore, the error vanishes faster by exploiting a new concept of sample diversity. We then develop the EXPLAIN, AGREE, LEARN (EXAL) method that uses this objective. EXPLAIN samples explanations for the data. AGREE reweighs each explanation in concordance with the neural component. LEARN uses the reweighed explanations as a signal for learning. In contrast to previous NeSy methods, EXAL can scale to larger problem sizes while retaining theoretical guarantees on the error. Experimentally, our theoretical claims are verified and EXAL outperforms recent NeSy methods when scaling up the MNIST addition and Warcraft pathfinding problems.

Can Large Language Models Reason? A Characterization via 3-SAT

Large Language Models (LLMs) are said to possess advanced reasoning abilities. However, some skepticism exists as recent works show how LLMs often bypass true reasoning using shortcuts. Current methods for assessing the reasoning abilities of LLMs typically rely on open-source benchmarks that may be overrepresented in LLM training data, potentially skewing performance. We instead provide a computational theory perspective of reasoning, using 3-SAT -- the prototypical NP-complete problem that lies at the core of logical reasoning and constraint satisfaction tasks. By examining the phase transitions in 3-SAT, we empirically characterize the reasoning abilities of LLMs and show how they vary with the inherent hardness of the problems. Our experimental evidence shows that LLMs cannot perform true reasoning, as is required for solving 3-SAT problems.

On the Hardness of Probabilistic Neurosymbolic Learning

The limitations of purely neural learning have sparked an interest in probabilistic neurosymbolic models, which combine neural networks with probabilistic logical reasoning. As these neurosymbolic models are trained with gradient descent, we study the complexity of differentiating probabilistic reasoning. We prove that although approximating these gradients is intractable in general, it becomes tractable during training. Furthermore, we introduce WeightME, an unbiased gradient estimator based on model sampling. Under mild assumptions, WeightME approximates the gradient with probabilistic guarantees using a logarithmic number of calls to a SAT solver. Lastly, we evaluate the necessity of these guarantees on the gradient. Our experiments indicate that the existing biased approximations indeed struggle to optimize even when exact solving is still feasible.

CLEVR-POC: Reasoning-Intensive Visual Question Answering in Partially Observable Environments

The integration of learning and reasoning is high on the research agenda in AI. Nevertheless, there is only a little attention to use existing background knowledge for reasoning about partially observed scenes to answer questions about the scene. Yet, we as humans use such knowledge frequently to infer plausible answers to visual questions (by eliminating all inconsistent ones). Such knowledge often comes in the form of constraints about objects and it tends to be highly domain or environment-specific. We contribute a novel benchmark called CLEVR-POC for reasoning-intensive visual question answering (VQA) in partially observable environments under constraints. In CLEVR-POC, knowledge in the form of logical constraints needs to be leveraged to generate plausible answers to questions about a hidden object in a given partial scene. For instance, if one has the knowledge that all cups are colored either red, green or blue and that there is only one green cup, it becomes possible to deduce the color of an occluded cup as either red or blue, provided that all other cups, including the green one, are observed. Through experiments, we observe that the low performance of pre-trained vision language models like CLIP (~ 22%) and a large language model (LLM) like GPT-4 (~ 46%) on CLEVR-POC ascertains the necessity for frameworks that can handle reasoning-intensive tasks where environment-specific background knowledge is available and crucial. Furthermore, our demonstration illustrates that a neuro-symbolic model, which integrates an LLM like GPT-4 with a visual perception network and a formal logical reasoner, exhibits exceptional performance on CLEVR-POC.