Agentic search has emerged as a promising paradigm for adaptive retrieval systems powered by large language models (LLMs). However, existing benchmarks primarily focus on quality, overlooking efficiency factors that are critical for real-world deployment. Moreover, real-world user queries often contain underspecified preferences, a challenge that remains largely underexplored in current agentic search evaluation. As a result, many agentic search systems remain impractical despite their impressive performance. In this work, we introduce HotelQuEST, a benchmark comprising 214 hotel search queries that range from simple factual requests to complex queries, enabling evaluation across the full spectrum of query difficulty. We further address the challenge of evaluating underspecified user preferences by collecting clarifications that make annotators' implicit preferences explicit for evaluation. We find that LLM-based agents achieve higher accuracy than traditional retrievers, but at substantially higher costs due to redundant tool calls and suboptimal routing that fails to match query complexity to model capability. Our analysis exposes inefficiencies in current agentic search systems and demonstrates substantial potential for cost-aware optimization.

Small language models (SLMs) have emerged as efficient alternatives to large language models for task-specific applications. However, they are often employed in high-volume, low-latency settings, where efficiency is crucial. We propose TASC, Task-Adaptive Sequence Compression, a framework for SLM acceleration comprising two use-cases: When performing SLM fine-tuning, we propose TASC-ft, which iteratively enriches the tokenizer vocabulary with high-frequency output n-grams and then fine-tunes the model to utilize the expanded vocabulary. Next, we propose an inference-time method, termed TASC-spec. TASC-spec is a lightweight, training-free speculative decoding method that constructs an n-gram draft model from the task's output corpus, mixing task and context n-gram information.TASC-spec avoids any additional training, while bypassing draft-target vocabulary alignment constraints. We demonstrate the effectiveness of both methods across multiple low output-variability generation tasks. Our methods show consistent improvements in inference efficiency while maintaining task performance.

Grounded text generation, encompassing tasks such as long-form question-answering and summarization, necessitates both content selection and content consolidation. Current end-to-end methods are difficult to control and interpret due to their opaqueness. Accordingly, recent works have proposed a modular approach, with separate components for each step. Specifically, we focus on the second subtask, of generating coherent text given pre-selected content in a multi-document setting. Concretely, we formalize Fusion-in-Context (FiC) as a standalone task, whose input consists of source texts with highlighted spans of targeted content. A model then needs to generate a coherent passage that includes all and only the target information. Our work includes the development of a curated dataset of 1000 instances in the reviews domain, alongside a novel evaluation framework for assessing the faithfulness and coverage of highlights, which strongly correlate to human judgment. Several baseline models exhibit promising outcomes and provide insightful analyses. This study lays the groundwork for further exploration of modular text generation in the multi-document setting, offering potential improvements in the quality and reliability of generated content.(1)

Multi-document summarization (MDS) is a challenging task, often decomposed to subtasks of salience and redundancy detection, followed by text generation. In this context, alignment of corresponding sentences between a reference summary and its source documents has been leveraged to generate training data for some of the component tasks. Yet, this enabling alignment step has usually been applied heuristically on the sentence level on a limited number of subtasks. In this paper, we propose extending the summary-source alignment framework by (1) applying it at the more fine-grained proposition span level, (2) annotating alignment manually in a multi-document setup, and (3) revealing the great potential of summary-source alignments to yield several datasets for at least six different tasks. Specifically, for each of the tasks, we release a manually annotated test set that was derived automatically from the alignment annotation. We also release development and train sets in the same way, but from automatically derived alignments. Using the datasets, each task is demonstrated with baseline models and corresponding evaluation metrics to spur future research on this broad challenge.

The common practice for assessing automatic evaluation metrics is to measure the correlation between their induced system rankings and those obtained by reliable human evaluation, where a higher correlation indicates a better metric. Yet, an intricate setting arises when an NLP task is evaluated by multiple Quality Criteria (QCs), like for text summarization where prominent criteria including relevance, consistency, fluency and coherence. In this paper, we challenge the soundness of this methodology when multiple QCs are involved, concretely for the summarization case. First, we show that the allegedly best metrics for certain QCs actually do not perform well, failing to detect even drastic summary corruptions with respect to the considered QC. To explain this, we show that some of the high correlations obtained in the multi-QC setup are spurious. Finally, we propose a procedure that may help detecting this effect. Overall, our findings highlight the need for further investigating metric evaluation methodologies for the multiple-QC case.

In e-commerce sites, customer questions on the product details-page express the customers' information needs about the product. The answers to these questions often provide the necessary information. In this work, we present and address the novel task of generating product insights from community questions and answers (Q&A). These insights can be presented to customers to assist them in their shopping journey. Our method first generates concise, self-contained sentences based on the information in the Q&A. Then insights are selected based on the prominence of their associated questions. Empirical evaluation attests to the effectiveness of our approach in generating well-formed, objective, and helpful insights that are often not available in the product description or in summaries of customer reviews.

Phrase similarity is a key component of many NLP applications.Current phrase similarity methods focus on embedding the phrase itself and use the phrase context only during training of the pretrained model.To better leverage the information in the context, we propose McPhraSy (Multi-context Phrase Similarity), a novel algorithm for estimating the similarity of phrases based on multiple contexts.At inference time, McPhraSy represents each phrase by considering multiple contexts in which it appears and computes the similarity of two phrases by aggregating the pairwise similarities between the contexts of the phrases.Incorporating context during inference enables McPhraSy to outperform current state-of-theart models on two phrase similarity datasets by up to 13.3%.Finally, we also present a new downstream task that relies on phrase similarity -keyphrase clustering -and create a new benchmark for it in the product reviews domain.We show that McPhraSy surpasses all other baselines for this task.

The product reviews summarization task aims to automatically produce a short summary for a set of reviews of a given product. Such summaries are expected to aggregate a range of different opinions in a concise, coherent and informative manner. This challenging task gives rise to two shortcomings in existing work. First, summarizers tend to favor generic content that appears in reviews for many different products, resulting in template-like, less informative summaries. Second, as reviewers often disagree on the pros and cons of a given product, summarizers sometimes yield inconsistent, self-contradicting summaries. We propose the PASS system (Perturb-and-Select Summarizer) that employs a large pre-trained Transformer-based model (T5 in our case), which follows a few-shot fine-tuning scheme. A key component of the PASS system relies on applying systematic perturbations to the model’s input during inference, which allows it to generate multiple different summaries per product. We develop a method for ranking these summaries according to desired criteria, coherence in our case, enabling our system to almost entirely avoid the problem of self-contradiction. We compare our system against strong baselines on publicly available datasets, and show that it produces summaries which are more informative, diverse and coherent.

In recent years online shopping has gained momentum and became an important venue for customers wishing to save time and simplify their shopping process. A key advantage of shopping online is the ability to read what other customers are saying about products of interest. In this work, we aim to maintain this advantage in situations where extreme brevity is needed, for example, when shopping by voice. We suggest a novel task of extracting a single representative helpful sentence from a set of reviews for a given product. The selected sentence should meet two conditions: first, it should be helpful for a purchase decision and second, the opinion it expresses should be supported by multiple reviewers. This task is closely related to the task of Multi Document Summarization in the product reviews domain but differs in its objective and its level of conciseness. We collect a dataset in English of sentence helpfulness scores via crowd-sourcing and demonstrate its reliability despite the inherent subjectivity involved. Next, we describe a complete model that extracts representative helpful sentences with positive and negative sentiment towards the product and demonstrate that it outperforms several baselines.

Artificial intelligence (AI) is defined as the ability of machines to perform tasks that are usually associated with intelligent beings. Argument and debate are fundamental capabilities of human intelligence, essential for a wide range of human activities, and common to all human societies. The development of computational argumentation technologies is therefore an important emerging discipline in AI research 1 . Here we present Project Debater, an autonomous debating system that can engage in a competitive debate with humans. We provide a complete description of the system’s architecture, a thorough and systematic evaluation of its operation across a wide range of debate topics, and a detailed account of the system’s performance in its public debut against three expert human debaters. We also highlight the fundamental differences between debating with humans as opposed to challenging humans in game competitions, the latter being the focus of classical ‘grand challenges’ pursued by the AI research community over the past few decades. We suggest that such challenges lie in the ‘comfort zone’ of AI, whereas debating with humans lies in a different territory, in which humans still prevail, and for which novel paradigms are required to make substantial progress.

Product reviews summarization is a type of Multi-Document Summarization (MDS) task in which the summarized document sets are often far larger than in traditional MDS (up to tens of thousands of reviews). We highlight this difference and coin the term "Massive Multi-Document Summarization" (MMDS) to denote an MDS task that involves hundreds of documents or more. Prior work on product reviews summarization considered small samples of the reviews, mainly due to the difficulty of handling massive document sets. We show that summarizing small samples can result in loss of important information and provide misleading evaluation results. We propose a schema for summarizing a massive set of reviews on top of a standard summarization algorithm. Since writing large volumes of reference summaries needed for advanced neural network models is impractical, our solution relies on weak supervision. Finally, we propose an evaluation scheme that is based on multiple crowdsourced reference summaries and aims to capture the massive review collection. We show that an initial implementation of our schema significantly improves over several baselines in ROUGE scores, and exhibits strong coherence in a manual linguistic quality assessment.

Wikipedia is a very popular source of encyclopedic knowledge which provides highly reliable articles in a variety of domains. This richness and popularity created a strong motivation among NLP researchers to develop relatedness measures between Wikipedia concepts. In this paper, we introduce WORD (Wikipedia Oriented Relatedness Dataset), a new type of concept relatedness dataset, composed of 19,276 pairs of Wikipedia concepts. This is the first human annotated dataset of Wikipedia concepts, whose purpose is twofold. On the one hand, it can serve as a benchmark for evaluating concept-relatedness methods. On the other hand, it can be used as supervised data for developing new models for concept relatedness prediction. Among the advantages of this dataset compared to its term-relatedness counterparts, are its built-in disambiguation solution, and its richness with meaningful multiword terms. Based on this benchmark we develop a new tool, named WORT (Wikipedia Oriented Relatedness Tool), for measuring the level of relatedness between pairs of concepts. We show that the relatedness predictions of WORT outperform state of the art methods.

Searching for sentences containing claims in a large text corpus is a key component in developing an argumentative content search engine. Previous works focused on detecting claims in a small set of documents or within documents enriched with argumentative content. However, pinpointing relevant claims in massive unstructured corpora, received little attention. A step in this direction was taken in (Levy et al. 2017), where the authors suggested using a weak signal to develop a relatively strict query for claim–sentence detection. Here, we leverage this work to define weak signals for training DNNs to obtain significantly greater performance. This approach allows to relax the query and increase the potential coverage. Our results clearly indicate that the system is able to successfully generalize from the weak signal, outperforming previously reported results in terms of both precision and coverage. Finally, we adapt our system to solve a recent argument mining task of identifying argumentative sentences in Web texts retrieved from heterogeneous sources, and obtain F1 scores comparable to the supervised baseline.

Automatic claim detection is a fundamental argument mining task that aims to automatically mine claims regarding a topic of consideration. Previous works on mining argumentative content have assumed that a set of relevant documents is given in advance. Here, we present a first corpus– wide claim detection framework, that can be directly applied to massive corpora. Using simple and intuitive empirical observations, we derive a claim sentence query by which we are able to directly retrieve sentences in which the prior probability to include topic-relevant claims is greatly enhanced. Next, we employ simple heuristics to rank the sentences, leading to an unsupervised corpus–wide claim detection system, with precision that outperforms previously reported results on the task of claim detection given relevant documents and labeled data.

In the standard agnostic multiclass model, pairs are sampled independently from some underlying distribution. This distribution induces a conditional probability over the labels given an instance, and our goal in this paper is to learn this conditional distribution. Since even unconditional densities are quite challenging to learn, we give our learner access to pairs. Assuming a base learner oracle in this model, we might seek a boosting algorithm for constructing a strong learner. Unfortunately, without further assumptions, this is provably impossible. However, we give a new boosting algorithm that succeeds in the following sense: given a base learner guaranteed to achieve some average accuracy (i.e., risk), we efficiently construct a learner that achieves the same level of accuracy with arbitrarily high probability. We give generalization guarantees of several different kinds, including distribution-free accuracy and risk bounds. None of our estimates depend on the number of boosting rounds and some of them admit dimension-free formulations.

GRASP (GReedy Augmented Sequential Patterns) is an algorithm for automatically extracting patterns that characterize subtle linguistic phenomena. To that end, GRASP augments each term of input text with multiple layers of linguistic information. These different facets of the text terms are systematically combined to reveal rich patterns. We report highly promising experimental results in several challenging text analysis tasks within the field of Argumentation Mining. We believe that GRASP is general enough to be useful for other domains too. For example, each of the following sentences includes a claim for a [topic]: 1. Opponents often argue that the open primary is unconstitutional. [Open Primaries] 2. Prof. Smith suggested that affirmative action devalues the accomplishments of the chosen. [Affirmative Action] 3. The majority stated that the First Amendment does not guarantee the right to offend others. [Freedom of Speech] These sentences share almost no words in common, however, they are similar at a more abstract level. A human observer may notice the following underlying common structure, or pattern: [someone][argue/suggest/state][that][topic term][sentiment term]. GRASP aims to automatically capture such underlying structures of the given data. For the above examples it finds the pattern [noun][express][that][noun,topic][sentiment], where [express] stands for all its (in)direct hyponyms, and [noun,topic] means a noun which is also related to the topic.

Measuring word relatedness is an important ingredient of many NLP applications. Several datasets have been developed in order to evaluate such measures. The main drawback of existing datasets is the focus on single words, although natural language contains a large proportion of multiword terms. We propose the new TR9856 dataset which focuses on multi-word terms and is significantly larger than existing datasets. The new dataset includes many real world terms such as acronyms and named entities, and further handles term ambiguity by providing topical context for all term pairs. We report baseline results for common relatedness methods over the new data, and exploit its magnitude to demonstrate that a combination of these methods outperforms each individual method.

While discussing a concrete controversial topic, most humans will find it challenging to swiftly raise a diverse set of convincing and relevant claims that should set the basis of their arguments. Here, we demonstrate the initial capabilities of a system that, given a controversial topic, can automatically pinpoint relevant claims in Wikipedia, determine their polarity with respect to the given topic, and articulate them per the user's request.

While discussing a concrete controversial topic, most humans will find it challenging to swiftly raise a diverse set of convincing and relevant claims that should set the basis of their arguments. Here, we formally define the challenging task of automatic claim detection in a given context and discuss its associated unique difficulties. Further, we outline a preliminary solution to this task, and assess its performance over annotated real world data, collected specifically for that purpose over hundreds of Wikipedia articles. We report promising results of a supervised learning approach, which is based on a cascade of classifiers designed to properly handle the skewed data which is inherent to the defined task. These results demonstrate the viability of the introduced task.

We describe a novel and unique argumentative structure dataset.This corpus consists of data extracted fro m hundreds of Wikipedia articles using a meticulously monitored manual annotation process.The result is 2,683 argument elements, collected in the context of 33 controversial topics, organized under a simp le claim-evidence structure.The obtained data are publicly available for academic research.