Xiaotian Liu, Ali Pesaranghader, Hanze Li, Punyaphat Sukcharoenchaikul, Jaehong Kim, Tanmana Sadhu, Hyejeong Jeon, Scott Sanner. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Large language models (LLM) are being increasingly applied to tasks requiring commonsense reasoning. Despite their outstanding potential, the reasoning process of LLMs is prone to errors and hallucinations that hinder their applicability, especially in high-stakes scenarios. Several works have attempted to enhance commonsense reasoning performance of LLMs by (i) using prompting styles that elicit more accurate reasoning, (ii) utilizing the LLM as a semantic parser for a symbolic reasoner, or (iii) enforcing the LLM to simulate a logical inference rule. However, all these solutions have critical limitations: they are unable to leverage the internal commonsense knowledge of the LLM in tandem with an axiomatic knowledge base, they lack a mechanism to reliably repair erroneous inference steps, and their application is restricted to small knowledge bases that fit the context limit of the LLM. In this work, we present LLM-based Typed Hyperresolution (LLM-TH), a logical commonsense reasoning framework that leverages "theory resolution", a concept from classical logical inference which enables integrating LLMs into the "resolution" inference rule, thus mitigating reasoning errors and hallucinations and enabling verification of the reasoning procedure. LLM-TH is also equipped with a mechanism for repairing erroneous inference steps supported by theoretical guarantees. Using "Hyperresolution" and "Typed inference" schemes, we show that LLM-TH can efficiently reason over large knowledge bases consisting of tens of thousands of rules with arbitrary predicate arities. Our experiments on three diverse language-based reasoning tasks—preference reasoning, multi-domain deductive reasoning, and geographical question answering—showcase that LLM-TH, using merely a BART 406M parameter NLI entailment model, significantly reduces reasoning errors compared to baselines using Llama3-70B, Gemini1.5-Flash, GPT-3.5-Turbo, and Mixtral-46.7B.

The ability to actively acquire information is essential for open-world planning under partial observability and incomplete knowledge. However, most existing embodied AI systems either assume a known object category or rely on passive perception strategies that exhaustively gather object and relational information from the environment. Such a strategy becomes insufficient in visually complex open-world settings. For instance, a typical household may contain thousands of novel and uniquely configured objects, most of which are irrelevant to the agent’s current task. Consequently, open-world agents must be capable of actively identifying and prioritizing task-relevant objects to enable efficient and goal-directed knowledge acquisition. In this work, we introduce ActiveVOO, a novel zero-shot framework for open-world embodied planning that emphasizes object-centric active knowledge acquisition. ActiveVOO employs lifted regression to generate compact, first-order subgoal descriptions that identify task-relevant objects, and provides a principled mechanism to quantify the utility of sensing actions based on commonsense priors derived from LLMs and VLMs. We evaluate ActiveVOO on the visual ALFWorld benchmark, where it achieves substantial improvements over existing LLM- and VLM-based planning approaches, notably outperforming VLMs fine-tuned on ALFWorld data. This work establishes a principled foundation for developing embodied agents capable of actively and efficiently acquiring knowledge to plan and act in open-world environments.

Recent advances in Large Language Models (LLM) have led to substantial interest in their application to commonsense reasoning tasks.Despite their potential, LLMs are susceptible to reasoning errors and hallucinations that may be harmful in use cases where accurate reasoning is critical.This challenge underscores the need for verifiable, debuggable, and repairable LLM reasoning.Recent works have made progress toward verifiable reasoning with LLMs by using them as either (i) a reasoner over an axiomatic knowledge base, or (ii) a semantic parser for use in existing logical inference systems.However, both settings are unable to extract commonsense axioms from the LLM that are not already formalized in the knowledge base, and also lack a reliable method to repair missed commonsense inferences.In this work, we present LLM-TRes, a logical reasoning framework based on the notion of "theory resolution" that allows for seamless integration of the commonsense knowledge from LLMs with a verifiable logical reasoning framework that mitigates hallucinations and facilitates debugging of the reasoning procedure as well as repair.We crucially prove that repaired axioms are theoretically guaranteed to be given precedence over flawed ones in our theory resolution inference process.We conclude by evaluating on three diverse language-based reasoning tasks -preference reasoning, deductive reasoning, and causal commonsense reasoning -and demonstrate the superior performance of LLM-TRes vs. state-of-the-art LLM-based reasoning methods in terms of both accuracy and reasoning correctness.

Harmful and offensive communication or content is detrimental to social bonding and the mental state of users on social media platforms. Text detoxification is a crucial task in natural language processing (NLP), where the goal is removing profanity and toxicity from text while preserving its content. Supervised and unsupervised learning are common approaches for designing text detoxification solutions. However, these methods necessitate fine-tuning, leading to computational overhead. In this paper, we propose GPT-DETOX as a framework for prompt-based in-context learning for text detoxification using GPT-3.5 Turbo. We utilize zero-shot and few-shot prompting techniques for detoxifying input sentences. To generate few-shot prompts, we propose two methods: word-matching example selection (WMES) and context-matching example selection (CMES). We additionally take into account ensemble in-context learning (EICL) where the ensemble is shaped by base prompts from zero-shot and all few-shot settings. We use ParaDetox and APPDIA as benchmark detoxification datasets. Our experimental results show that the zero-shot solution achieves promising performance, while our best few-shot setting outperforms the state-of-the-art models on ParaDetox and shows comparable results on APPDIA. Our EICL solutions obtain the greatest performance, adding at least 10% improvement.. against both datasets.

The rise of interactive recommendation assistants has led to a novel domain of natural language (NL) recommendation that would benefit from improved multi-aspect reasoning to retrieve relevant items based on NL statements of preference. Such preference statements often involve multiple aspects, e.g., "I would like meat lasagna but I'm watching my weight". Unfortunately, progress in this domain is slowed by the lack of annotated data. To address this gap, we curate a novel dataset which captures logical reasoning over multi-aspect, NL preference-based queries and a set of multiple-choice, multi-aspect item descriptions. We focus on the recipe domain in which multi-aspect preferences are often encountered due to the complexity of the human diet. The goal of publishing our dataset is to provide a benchmark for joint progress in three key areas: 1) structured, multi-aspect NL reasoning with a variety of properties (e.g., level of specificity, presence of negation, and the need for commonsense, analogical, and/or temporal inference), 2) the ability of recommender systems to respond to NL preference utterances, and 3) explainable NL recommendation facilitated by aspect extraction and reasoning. We perform experiments using a variety of methods (sparse and dense retrieval, zero- and few-shot reasoning with large language models) in two settings: a monolithic setting which uses the full query and an aspect-based setting which isolates individual query aspects and aggregates the results. GPT-3 results in much stronger performance than other methods with 73% zero-shot accuracy and 83% few-shot accuracy in the monolithic setting. Aspect-based GPT-3, which facilitates structured explanations, also shows promise with 68% zero-shot accuracy. These results establish baselines for future research into explainable recommendations via multi-aspect preference-based NL reasoning.

OBJECTIVE: In biomedicine, there is a wealth of information hidden in unstructured narratives such as research articles and clinical reports. To exploit these data properly, a word sense disambiguation (WSD) algorithm prevents downstream difficulties in the natural language processing applications pipeline. Supervised WSD algorithms largely outperform un- or semisupervised and knowledge-based methods; however, they train 1 separate classifier for each ambiguous term, necessitating a large number of expert-labeled training data, an unattainable goal in medical informatics. To alleviate this need, a single model that shares statistical strength across all instances and scales well with the vocabulary size is desirable. MATERIALS AND METHODS: Built on recent advances in deep learning, our deepBioWSD model leverages 1 single bidirectional long short-term memory network that makes sense prediction for any ambiguous term. In the model, first, the Unified Medical Language System sense embeddings will be computed using their text definitions; and then, after initializing the network with these embeddings, it will be trained on all (available) training data collectively. This method also considers a novel technique for automatic collection of training data from PubMed to (pre)train the network in an unsupervised manner. RESULTS: We use the MSH WSD dataset to compare WSD algorithms, with macro and micro accuracies employed as evaluation metrics. deepBioWSD outperforms existing models in biomedical text WSD by achieving the state-of-the-art performance of 96.82% for macro accuracy. CONCLUSIONS: Apart from the disambiguation improvement and unsupervised training, deepBioWSD depends on considerably less number of expert-labeled data as it learns the target and the context terms jointly. These merit deepBioWSD to be conveniently deployable in real-time biomedical applications.

Increasingly, Internet of Things (IoT) domains, such as sensor networks, smart cities, and social networks, generate vast amounts of data. Such data are not only unbounded and rapidly evolving. Rather, the content thereof dynamically evolves over time, often in unforeseen ways. These variations are due to so-called concept drifts, caused by changes in the underlying data generation mechanisms. In a classification setting, concept drift causes the previously learned models to become inaccurate, unsafe and even unusable. Accordingly, concept drifts need to be detected, and handled, as soon as possible. In medical applications and emergency response settings, for example, change in behaviours should be detected in near real-time, to avoid potential loss of life. To this end, we introduce the McDiarmid Drift Detection Method (MDDM), which utilizes McDiarmid's inequality [1] in order to detect concept drift. The MDDM approach proceeds by sliding a window over prediction results, and associate window entries with weights. Higher weights are assigned to the most recent entries, in order to emphasize their importance. As instances are processed, the detection algorithm compares a weighted mean of elements inside the sliding window with the maximum weighted mean observed so far. A significant difference between the two weighted means, upper-bounded by the McDiarmid inequality, implies a concept drift. Our extensive experimentation against synthetic and real-world data streams show that our novel method outperforms the state-of-the-art. Specifically, MDDM yields shorter detection delays as well as lower false negative rates, while maintaining high classification accuracies.

As the Internet grows rapidly, finding desirable information becomes a tedious and time consuming task. Topic-specific web crawlers, as utopian solutions, tackle this issue through traversing the Web and collecting information related to the topic of interest. In this regard, various methods are proposed. Nevertheless, they hardly consider desired sense of the given topic which would certainly play an important role to find relevant web pages. In this paper, we attempt to improve topic-specific web crawling by disambiguating the sense of the topic. This would avoid crawling irrelevant links interlaced with other senses of the topic. For this purpose, by considering links hypertext semantic, we employ Lin semantic similarity measure in our crawler, named LinCrawler, to distinguish topic sense-related links from the others. Moreover, we compare LinCrawler against TFCrawler which only considers frequency of terms in hypertexts. Experimental results show LinCrawler outperforms TFCrawler to collect more relevant web pages.