By leveraging effective signals to refine their outputs, large language models (LLMs) can achieve superior performance compared to single-pass outputs. However, internal signals often suffer from accumulated hallucinations and a lack of confidence, while external signals are typically difficult to obtain and apply, hindering the development of the self-refinement paradigm. In this work, we propose a novel method named TopoRefine, which integrates the reasoning topology within the model’s outputs as reliable high-level feedback. Specifically, TopoRefine encourages LLMs to explore reasoning paths without assuming a predefined direction for improvement, while using a consistency mechanism to prevent performance degradation. Reasoning topology can provide higher-level information as feedback, offering more precise semantics and facilitating self-analysis. Experiments on mathematical datasets and with various LLMs demonstrate significant improvements using our method. We also provide a detailed analysis of TopoRefine’s efficiency.

Large language models (LLMs) still grapple with complex tasks like mathematical reasoning. Despite significant efforts invested in improving prefix prompts or reasoning process, the crucial role of problem context might have been neglected. Accurate recognition of inputs is fundamental for solving mathematical tasks, as ill-formed problems could potentially mislead LLM’s reasoning. In this study, we propose a new approach named Problem Elaboration Prompting (PEP) to enhance the mathematical capacities of LLMs. Specifically, PEP decomposes and elucidates the problem context before reasoning, therefore enhancing the context modeling and parsing efficiency. Experiments across datasets and models demonstrate promising performances: (1) PEP demonstrates an overall enhancement in various situation. (2) PEP can be easily implemented and integrated with other prompting methods. (3) PEP shows particular strength in handling distraction problems.

Large language models (LLMs) can achieve superior results through iterative refinement based on internal or external signals, compared to the unstable outputs from a single pass. However, the reliability of existing internal signals is questionable due to their susceptibility to intrinsic hallucinations, while external signals are only useful in limited scenarios. In this paper, we introduce a novel framework called Progressive Back-Translation refinement (PBT). Specifically, PBT prompts LLMs to extract and reconstruct the question from the answer, avoiding unreliable inferences, critiques, or judgments on intermediate results. We then derive and provide accurate, fine-grained feedback by identifying discrepancies between the back-translated and original questions. Experiments across various large language models and challenging mathematical datasets demonstrate consistent improvements. We also provide a detailed analysis to confirm the effectiveness of the proposed method.

Haoran Liao, Shaohua Hu, Zhihao Zhu, Hao He, Yaohui Jin. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Game theory, as an analytical tool, is frequently utilized to analyze human behavior in social science research. With the high alignment between the behavior of Large Language Models (LLMs) and humans, a promising research direction is to employ LLMs as substitutes for humans in game experiments, enabling social science research. However, despite numerous empirical researches on the combination of LLMs and game theory, the capability boundaries of LLMs in game theory remain unclear. In this research, we endeavor to systematically analyze LLMs in the context of game theory. Specifically, rationality, as the fundamental principle of game theory, serves as the metric for evaluating players' behavior --- building a clear desire, refining belief about uncertainty, and taking optimal actions. Accordingly, we select three classical games (dictator game, Rock-Paper-Scissors, and ring-network game) to analyze to what extent LLMs can achieve rationality in these three aspects. The experimental results indicate that even the current state-of-the-art LLM (GPT-4) exhibits substantial disparities compared to humans in game theory. For instance, LLMs struggle to build desires based on uncommon preferences, fail to refine belief from many simple patterns, and may overlook or modify refined belief when taking actions. Therefore, we consider that introducing LLMs into game experiments in the field of social science should be approached with greater caution.

In-Context Learning (ICL) is an important paradigm for adapting Large Language Models (LLMs) to downstream tasks through a few demonstrations. Despite the great success of ICL, the limitation of the demonstration number may lead to demonstration bias, i.e. the input-label mapping induced by LLMs misunderstands the task’s essence. Inspired by human experience, we attempt to mitigate such bias through the perspective of the inter-demonstration relationship. Specifically, we construct Comparable Demonstrations (CDs) by minimally editing the texts to flip the corresponding labels, in order to highlight the task’s essence and eliminate potential spurious correlations through the inter-demonstration comparison. Through a series of experiments on CDs, we find that (1) demonstration bias does exist in LLMs, and CDs can significantly reduce such bias; (2) CDs exhibit good performance in ICL, especially in out-of-distribution scenarios. In summary, this study explores the ICL mechanisms from a novel perspective, providing a deeper insight into the demonstration selection strategy for ICL.

Conditional text generation is to generate text sequences conditioning on linguistic or non-linguistic data. The main line of existing work proposed deterministic models to improve the fidelity of the generated text but often ignored the diversity. Another line relied on conditional variational auto-encoders (CVAEs), which increased the diversity over their deterministic backbones. However, CVAEs regard diversity as an implicit objective and may not be optimal. In this paper, we raise two questions: i) Can diversity be further improved with an explicit objective? ii) Since fidelity and diversity are two conflicting objectives, how can we obtain different multi-objective optimal solutions according to user preferences? To answer question i), we propose a multi-objective reinforcement learning (MORL) method which explicitly takes CIDEr and Self-CIDEr scores as the fidelity-oriented and diversity-oriented rewards respectively. To answer question ii), we propose a preference-controlled MORL method, which can obtain infinite multi-objective optimal solutions by tuning the preference variable. We conduct extensive experiments on paraphrasing and image captioning tasks, which show that in the fidelity-diversity trade-off space, our model outperforms both deterministic and CVAE-based baselines.

Unsupervised text-graph alignment (UTGA) is a fundamental task that bidirectionally generates texts and graphs without parallel data. Most available models of UTGA suffer from information asymmetry, a common phenomenon that texts and graphs include additional information invisible to each other. On the one hand, these models fail to supplement asymmetric information effectively due to the lack of ground truths. On the other hand, it is challenging to indicate asymmetric information with explicit indicators because it cannot be decoupled from the data directly. To address the challenge posed by information asymmetry, we propose the assumption that asymmetric information is encoded in unobservable latent variables and only affects the one-way generation processes. These latent variables corresponding to asymmetric information should obey prior distributions recovered approximately from original data. Therefore, we first propose a taxonomy of the latent variable that classifies the latent variable into transferrable (TV) and non-transferable (NTV) variables and further distinguish NTV as the dependent variable (DV) and the independent variable (IV). Next, we propose three latent VAE-based regularizations on TV, DV, and IV to constrain their distributions to well-designed prior distributions to introduce asymmetric information into models and enhance the preservation of shared contents. Finally, we impose the three proposed constraints on a cycle-consistent learning framework, back-translation (BT), named ConstrainedBT. Experimental results on three UTGA tasks demonstrate the effectiveness of ConstrainedBT on the information-asymmetric challenge.

Large language models (LLMs) have shown incredible performance on many tasks such as dialogue generation, commonsense reasoning and question answering.In-context learning (ICL) is an important paradigm for adapting LLMs to the downstream tasks by prompting few demonstrations.However, the distribution of demonstrations can severely affect the performance, especially for challenging classification tasks.In this paper, we propose the concept of task-level thinking steps that can eliminate bias introduced by demonstrations.Further, to help LLMs distinguish confusing classes, we design a progressive revision framework, which can improve the thinking steps by correcting hard demonstrations.Experimental results prove the superiority of our proposed method, achieving best performance on three kinds of challenging classification tasks in the zero-shot and few-shot settings.Besides, with task-level thinking steps, automatically generated chain-of-thoughts (CoTs) bring more competitive performance.

Chain-of-Thought (CoT) is a technique that guides Large Language Models (LLMs) to decompose complex tasks into multi-step reasoning through intermediate steps in natural language form.Briefly, CoT enables LLMs to think step by step.However, although many Natural Language Understanding (NLU) tasks also require thinking step by step, LLMs perform less well than small-scale Masked Language Models (MLMs).To migrate CoT from LLMs to MLMs, we propose Chain-of-Thought Tuning (CoTT), a two-step reasoning framework based on prompt tuning, to implement step-by-step thinking for MLMs on NLU tasks.From the perspective of CoT, CoTT's two-step framework enables MLMs to implement task decomposition; CoTT's prompt tuning allows intermediate steps to be used in natural language form.Thereby, the success of CoT can be extended to NLU tasks through MLMs.To verify the effectiveness of CoTT, we conduct experiments on two NLU tasks: hierarchical classification and relation extraction, and the results show that CoTT outperforms baselines and achieves state-of-the-art performance.

Counterfactually-Augmented Data (CAD) has the potential to improve language models’ Out-Of-Distribution (OOD) generalization capability, as CAD induces language models to exploit causal features and exclude spurious correlations. However, the empirical results of OOD generalization on CAD are not as efficient as expected. In this paper, we attribute the inefficiency to Myopia Phenomenon caused by CAD: language models only focus on causal features that are edited in the augmentation and exclude other non-edited causal features. As a result, the potential of CAD is not fully exploited. Based on the structural properties of CAD, we design two additional constraints to help language models extract more complete causal features contained in CAD, thus improving the OOD generalization capability. We evaluate our method on two tasks: Sentiment Analysis and Natural Language Inference, and the experimental results demonstrate that our method could unlock CAD’s potential and improve language models’ OOD generalization capability.

Despite the recent success of end-to-end deep neural networks, there are growing concerns about their lack of logical reasoning abilities, especially on cognitive tasks with perception and reasoning processes. A solution is the neural symbolic learning (NeSyL) method that can effectively utilize pre-defined logic rules to constrain the neural architecture making it perform better on cognitive tasks. However, it is challenging to apply NeSyL to these cognitive tasks because of the lack of supervision, the non-differentiable manner of the symbolic system, and the difficulty to probabilistically constrain the neural network. In this paper, we propose WS-NeSyL, a weakly supervised neural symbolic learning model for cognitive tasks with logical reasoning. First, WS-NeSyL employs a novel back search algorithm to sample the possible reasoning process through logic rules. This sampled process can supervise the neural network as the pseudo label. Based on this algorithm, we can backpropagate gradients to the neural network of WS-NeSyL in a weakly supervised manner. Second, we introduce a probabilistic logic regularization into WS-NeSyL to help the neural network learn probabilistic logic. To evaluate WS-NeSyL, we have conducted experiments on three cognitive datasets, including temporal reasoning, handwritten formula recognition, and relational reasoning datasets. Experimental results show that WS-NeSyL not only outperforms the end-to-end neural model but also beats the state-of-the-art neural symbolic learning models.

Reasoning and knowledge-related skills are considered as two fundamental skills for natural language understanding (NLU) tasks such as machine reading comprehension (MRC) and natural language inference (NLI). However, it is not clear to what extent an NLU task defined on a dataset correlates to a specific NLU skill. On the one hand, evaluating the correlation requires an understanding of the significance of the NLU skill in a dataset. Significance judges whether a dataset includes sufficient material to help the model master this skill. On the other hand, it is also necessary to evaluate the dependence of the task on the NLU skill. Dependence is a measure of how much the task defined on a dataset depends on the skill. In this paper, we propose a systematic method to diagnose the correlations between an NLU dataset and a specific skill, and then take a fundamental reasoning skill, logical reasoning, as an example for analysis. The method adopts a qualitative indicator to indicate the significance while adopting a quantitative indicator to measure the dependence. We perform diagnosis on 8 MRC datasets (including two types) and 3 NLI datasets and acquire intuitively reasonable results. We then perform the analysis to further understand the results and the proposed indicators. Based on the analysis, although the diagnostic method has some limitations, it is still an effective method to perform a basic diagnosis of the correlation between the dataset and logical reasoning skill, which also can be generalized to other NLU skills.

Wenqing Chen, Jidong Tian, Yitian Li, Hao He, Yaohui Jin. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Recently, language models (LMs) have achieved significant performance on many NLU tasks, which has spurred widespread interest for their possible applications in the scientific and social area.However, LMs have faced much criticism of whether they are truly capable of reasoning in NLU.In this work, we propose a diagnostic method for first-order logic (FOL) reasoning with a new proposed benchmark, LogicNLI.LogicNLI is an NLI-style dataset that effectively disentangles the target FOL reasoning from commonsense inference and can be used to diagnose LMs from four perspectives: accuracy, robustness, generalization, and traceability.Experiments on BERT, RoBERTa, and XLNet, have uncovered the weaknesses of these LMs on FOL reasoning, which motivates future exploration to enhance the reasoning ability.

Liqiang Xiao, Jun Ma, Xin Luna Dong, Pascual Martínez-Gómez, Nasser Zalmout, Wei Chen, Tong Zhao, Hao He, Yaohui Jin. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. 2021.

Recent work for image captioning mainly followed an extract-then-generate paradigm, pre-extracting a sequence of object-based features and then formulating image captioning as a single sequence-to-sequence task. Although promising, we observed two problems in generated captions: 1) content inconsistency where models would generate contradicting facts; 2) not informative enough where models would miss parts of important information. From a causal perspective, the reason is that models have captured spurious statistical correlations between visual features and certain expressions (e.g., visual features of "long hair" and "woman"). In this paper, we propose a dependent multi-task learning framework with the causal intervention (DMTCI). Firstly, we involve an intermediate task, bag-of-categories generation, before the final task, image captioning. The intermediate task would help the model better understand the visual features and thus alleviate the content inconsistency problem. Secondly, we apply Pearl's do-calculus on the model, cutting off the link between the visual features and possible confounders and thus letting models focus on the causal visual features. Specifically, the high-frequency concept set is considered as the proxy confounders where the real confounders are inferred in the continuous space. Finally, we use a multi-agent reinforcement learning (MARL) strategy to enable end-to-end training and reduce the inter-task error accumulations. The extensive experiments show that our model outperforms the baseline models and achieves competitive performance with state-of-the-art models.

Works on scene text visual question answering (TextVQA) always emphasize the importance of reasoning questions and image contents. However, we find current TextVQA models lack reasoning ability and tend to answer questions by exploiting dataset bias and language priors. Moreover, our observations indicate that recent accuracy improvement in TextVQA is mainly contributed by stronger OCR engines, better pre-training strategies and more Transformer layers, instead of newly proposed networks. In this work, towards the reasoning ability, we 1) conduct module-wise contribution analysis to quantitatively investigate how existing works improve accuracies in TextVQA; 2) design a gradient-based explainability method to explore why TextVQA models answer what they answer and find evidence for their predictions; 3) perform qualitative experiments to visually analyze models reasoning ability and explore potential reasons behind such a poor ability.

Jointly using the extractive and abstractive summarization methods can combine their complementary advantages, generating both informative and concise summary. Existing methods that adopt an extract-then-abstract strategy have achieved impressive results, yet they suffer from the information loss in the abstraction step because they compress all the selected sentences without distinguish. Especially when the whole sentence is summary-worthy, salient content would be lost by compression. To address this problem, we propose HySum, a hybrid framework for summarization that can flexibly switch between copying sentence and rewriting sentence according to the degree of redundancy. In this way, our approach can effectively combine the advantages of two branches of summarization, juggling informativity and conciseness. Moreover, we based on Hierarchical Reinforcement Learning, propose an end-to-end reinforcing method to bridge together the extraction module and rewriting module, which can enhance the cooperation between them. Automatic evaluation shows that our approach significantly outperforms the state-of-the-arts on the CNN/DailyMail corpus. Human evaluation also demonstrates that our generated summaries are more informative and concise than popular models.

Paraphrase generation aims to generate semantically consistent sentences with different syntactic realizations.Most of the recent studies rely on the typical encoder-decoder framework where the generation process is deterministic.However, in practice, the ability to generate multiple syntactically different paraphrases is important.Recent work proposed to cooperate variational inference on a target-related latent variable to introduce the diversity.But the latent variable may be contaminated by the semantic information of other unrelated sentences, and in turn, change the conveyed meaning of generated paraphrases.In this paper, we propose a semantically consistent and syntactically variational encoder-decoder framework, which uses adversarial learning to ensure the syntactic latent variable be semantic-free.Moreover, we adopt another discriminator to improve the word-level and sentence-level semantic consistency.So the proposed framework can generate multiple semantically consistent and syntactically different paraphrases.The experiments show that our model outperforms the baseline models on the metrics based on both n-gram matching and semantic similarity, and our model can generate multiple different paraphrases by assembling different syntactic variables.

Previous studies have shown that hierarchical multi-task learning (MTL) can utilize task dependencies by stacking encoders and outperform democratic MTL.However, stacking encoders only considers the dependencies of feature representations and ignores the label dependencies in logically dependent tasks.Furthermore, how to properly utilize the labels remains an issue due to the cascading errors between tasks.In this paper, we view logically dependent MTL from the perspective of causal inference and suggest a mediation assumption instead of the confounding assumption in conventional MTL models.We propose a model including two key mechanisms: label transfer (LT) for each task to utilize the labels of all its lower-level tasks, and Gumbel sampling (GS) to deal with cascading errors.In the field of causal inference, GS in our model is essentially a counterfactual reasoning process, trying to estimate the causal effect between tasks and utilize it to improve MTL.We conduct experiments on two English datasets and one Chinese dataset.Experiment results show that our model achieves state-of-the-art on six out of seven subtasks and improves predictions' consistency.

Modeling content importance is an essential yet challenging task for summarization.Previous work is mostly based on statistical methods that estimate word-level salience, which does not consider semantics and larger context when quantifying importance.It is thus hard for these methods to generalize to semantic units of longer text spans.In this work, we apply information theory on top of pretrained language models and define the concept of importance from the perspective of information amount.It considers both the semantics and context when evaluating the importance of each semantic unit.With the help of pre-trained language models, it can easily generalize to different kinds of semantic units (n-grams or sentences).Experiments on CNN/Daily Mail and New York Times datasets demonstrate that our method can better model the importance of content than prior work based on F1 and ROUGE scores.

Knowledge graph embedding, which projects the symbolic relations and entities onto low-dimension continuous spaces, is essential to knowledge graph completion. Recently, translation-based embedding models (e.g. TransE) have aroused increasing attention for their simplicity and effectiveness. These models attempt to translate semantics from head entities to tail entities with the relations and infer richer facts outside the knowledge graph. In this paper, we propose a novel knowledge graph embedding method named TransMS, which translates and transmits multidirectional semantics: i) the semantics of head/tail entities and relations to tail/head entities with nonlinear functions and ii) the semantics from entities to relations with linear bias vectors. Our model has merely one additional parameter α than TransE for each triplet, which results in its better scalability in large-scale knowledge graph. Experiments show that TransMS achieves substantial improvements against state-of-the-art baselines, especially the Hit@10s of head entity prediction for N-1 relations and tail entity prediction for 1-N relations improved by about 27.1% and 24.8% on FB15K database respectively.