Fengxiang Cheng

Fengxiang Cheng, Haoxuan Li, Fenrong Liu et al. · pku

Large language models (LLMs) have achieved remarkable successes on various tasks. However, recent studies have found that there are still significant challenges to the logical reasoning abilities of LLMs, which can be categorized into the following two aspects: (1) Logical question answering: LLMs often fail to generate the correct answer within a complex logical problem which requires sophisticated deductive, inductive or abductive reasoning given a collection of premises. (2) Logical consistency: LLMs are prone to producing responses contradicting themselves across different questions. For example, a state-of-the-art question-answering LLM Macaw, answers Yes to both questions Is a magpie a bird? and Does a bird have wings? but answers No to Does a magpie have wings?. To facilitate this research direction, we comprehensively investigate the most cutting-edge methods and propose a detailed taxonomy. Specifically, to accurately answer complex logic questions, previous methods can be categorized based on reliance on external solvers, prompts, and fine-tuning. To avoid logical contradictions, we discuss concepts and solutions of various logical consistencies, including implication, negation, transitivity, factuality consistencies, and their composites. In addition, we review commonly used benchmark datasets and evaluation metrics, and discuss promising research directions, such as extending to modal logic to account for uncertainty and developing efficient algorithms that simultaneously satisfy multiple logical consistencies.

Jinkun Chen, Fengxiang Cheng, Sijia Han et al.

Reasoning failures in large language models (LLMs) are typically measured only at the end of a generation, yet many failures manifest as a process-level breakdown: the model "loses the thread" mid-reasoning. We study whether such breakdowns are detectable from inference-time observables available in standard APIs (token log probabilities), without any training or fine-tuning. We define a simple instability signal that combines consecutive-step distributional shift (JSD) and uncertainty (entropy), summarize each trace by its peak instability strength, and show that this signal reliably predicts failure. Across GSM8K and HotpotQA, instability strength predicts wrong answers with above-chance AUC and yields monotonic bucket-level accuracy decline at scale across model sizes. Crucially, we show that instability is not uniformly harmful: early instability can reflect subsequent stabilization and a correct final answer (\emph{corrective instability}), whereas late instability is more often followed by failure (\emph{destructive instability}), even at comparable peak magnitudes, indicating that recoverability depends not only on how strongly the distribution changes but also on when such changes occur relative to the remaining decoding horizon. The method is model-agnostic, training-free, and reproducible, and is presented as a diagnostic lens rather than a corrective or control mechanism.

Xiangyu Wang, Haocheng Yang, Fengxiang Cheng et al.

Large Language Models (LLMs) still struggle with complex logical reasoning. While previous works achieve remarkable improvements, their performance is highly dependent on the correctness of translating natural language (NL) problems into a symbolic language (SL). Though numerous works focusing on improving this translation accuracy, they only consider the similarity between the meaning of SL and NL, overlooking another crucial influencing factor, the selection of the target SL type itself. For example, first-order logic language specializes in logical reasoning with categorical syllogisms and complex quantifiers, while Boolean satisfiability formalism excels at representing constraint satisfaction like partial problems. To our knowledge, this is the first paper to claim and verify that different NL logical reasoning problem corresponds to different optimal SL formalization for translation. Based on this, we propose a methods to improve the logical reasoning performance of LLMs by adaptively selecting the most suitable SL for each problem prior to translation. Specifically, we leverage LLMs to select the target SL among first-order logic, logic programming and Boolean satisfiability and then translate the problem in NL to target SL expressions as well as employ the corresponding logical solver to derive the final answer. Experimental results on benchmarks show that our adaptive selection method significantly outperforms translating all into single SL and randomly selecting the SL. On a mixed dataset of these benchmarks, our approach achieves 96% accuracy, which improving performance by 25% compared to the second highest accuracy from the first-order logic translation.