The ability of language models in RAG systems to selectively refuse to answer based on flawed context is critical for safety, yet remains a significant failure point. Our large-scale study reveals that even frontier models struggle in this setting, with refusal accuracy dropping below 50

Effectively evaluating deep research agents that autonomously search the web, analyze information, and generate reports remains a major challenge, particularly when it comes to assessing long reports and giving detailed feedback on their intermediate steps. To address these gaps, we introduce Deep Research Comparator, a platform that offers a holistic framework for deep research agent hosting, side-by-side comparison, fine-grained human feedback collection, and ranking calculation. Given a user query, our platform displays the final reports from two different agents along with their intermediate steps during generation. Annotators can evaluate the overall quality of final reports based on side-by-side comparison, and also provide detailed feedback separately by assessing intermediate steps or specific text spans within the final report. Furthermore, we develop Simple Deepresearch, an end-to-end agent scaffold. This scaffold serves as a baseline that facilitates the easy integration of various large language models to transform them into deep research agents for evaluation. To demonstrate the platform's utility for deep research agent development, we have collected real user preference data from 17 annotators on three deep research agents. A demo video of our platform can be found at https://www.youtube.com/watch?v=g4d2dnbdseg.

Search-augmented LLM agents can produce deep research reports (DRRs), but verifying claim-level factuality remains challenging. Existing fact-checkers usually target general-domain atomic claims, and there is no benchmark to test whether such verifiers transfer to DRRs.Yet building such a benchmark for DRR fact-checkers is itself difficult because it requires expert judgments over cognitively demanding, domain-specific claims.In a controlled study with PhD-level specialists, unassisted experts achieve only 60.8% accuracy on hidden known-answer claims. We therefore propose evolving benchmarking via **Audit-then-Score** (**AtS**), in which labels and rationales remain revisable: when a verifier disagrees with the current benchmark, it submits evidence; an auditor adjudicates the dispute; and accepted revisions update the benchmark before scoring. After three additional **AtS** rounds, expert accuracy rises to 90.9%, showing that experts are better auditors than one-shot labelers.We instantiate **AtS** as **DeepFactBench**, a versioned DRR factuality benchmark with auditable rationales, and introduce **DeepFactEval**, a claim-level verifier.On the frozen **DeepFactBench** release, **DeepFactEval** achieves 83.4% accuracy, outperforming the best prior deep-research and traditional fact-checkers by 14.3 and 24.9 points, respectively, and transferring well to external factuality datasets.

Skills have become the de facto way to enable LLM agents to perform complex real-world tasks with customized instructions, workflows, and tools, but how to learn them automatically and effectively remains unclear. We introduce SkillLearnBench, the first benchmark for evaluating continual skill learning methods, comprising 20 verified, skill-dependent tasks across 15 sub-domains derived from a real-world skill taxonomy , evaluated at three levels: skill quality, execution trajectory, and task outcome. Using this benchmark, we evaluate recent continual learning techniques, those leveraging one-shot, self/teacher feedback, and skill creator to generate skills from agent experiences. We find that all continual learning methods improve over the no-skill baseline, yet consistent gains remain elusive: no method leads across all tasks and LLMs, and scaling to stronger LLMs does not reliably help. Continual learning improves tasks with clear, reusable workflows but struggles on open-ended tasks, and using stronger LLM backbones does not consistently produce better skills. Our analysis also revealed that multiple iterations in continual learning facilitate genuine improvement via external feedback, whereas self-feedback alone induces recursive drift. Our data and code are open-source at https://github.com/cxcscmu/SkillLearnBench to enable further studies of automatic skill generation and continual learning techniques.

Evaluating Retrieval-Augmented Generation (RAG) in large language models (LLMs) is challenging because benchmarks can quickly become stale. Questions initially requiring retrieval may become answerable from pretraining knowledge as newer models incorporate more recent information during pretraining, making it difficult to distinguish evidence-based reasoning from recall. We introduce NeoQA (News Events for Out-of-training Question Answering), a benchmark designed to address this issue. To construct NeoQA, we generated timelines and knowledge bases of fictional news events and entities along with news articles and Q&A pairs to prevent LLMs from leveraging pretraining knowledge, ensuring that no prior evidence exists in their training data. We propose our dataset as a new platform for evaluating evidence-based question answering, as it requires LLMs to generate responses exclusively from retrieved evidence and only when sufficient evidence is available. NeoQA enables controlled evaluation across various evidence scenarios, including cases with missing or misleading details. Our findings indicate that LLMs struggle to distinguish subtle mismatches between questions and evidence, and suffer from short-cut reasoning when key information required to answer a question is missing from the evidence, underscoring key limitations in evidence-based reasoning.

We present NewsQs (news-cues), a dataset that provides question-answer pairs for multiple news documents. To create NewsQs, we augment a traditional multi-document summarization dataset with questions automatically generated by a T5-Large model fine-tuned on FAQ-style news articles from the News On the Web corpus. We show that fine-tuning a model with control codes produces questions that are judged acceptable more often than the same model without them as measured through human evaluation. We use a QNLI model with high correlation with human annotations to filter our data. We release our final dataset of high-quality questions, answers, and document clusters as a resource for future work in query-based multi-document summarization.

Training a supervised news summarization model requires large amounts of high-quality training data consisting of news articles paired with reference summaries. However, obtaining such data is costly, and existing datasets contain considerable amount of noise. We present a new large-scale and high-quality dataset for supervised abstractive news summarization containing 1.3 million training samples, which we call CCSUM.(1) In creating this dataset, we take advantage of the journalistic inverted-pyramid style in news writing: In some articles, the first sentence can be considered a summary of the reported story. Accordingly, among 35 million CommonCrawl News articles, we identify pairs of articles about the same news story and use one article's first sentence as the summary for the other article. To ensure high quality, we apply strict filters whose parameters we optimize using Bayesian optimization. We show that the resulting dataset is more factual and informative than established summarization datasets; less than 1% of the summaries have major factual inconsistencies with the corresponding news articles, compared to 5.5% to 15.4% in existing datasets, according to our human evaluation. Summarization models trained on our dataset are more favored compared to those trained on CNN/Daily Mail. The proposed dataset can open new opportunities for future research in abstractive summarization.

Multimodal Large Language Models (MLLMs) excel at synthesizing key information from diverse sources. However, generating accurate and faithful multimodal summaries is challenging, primarily due to the lack of appropriate multimodal datasets for fine-tuning that meaningfully integrate textual and visual modalities. To address this gap, we present a new dataset specifically designed for image-text multimodal summarization, harnessing the capabilities of state-of-the-art MLLMs. We generate summaries from Wikipedia sections and corresponding images and evaluate them across text-based, visual and multimodal dimensions, employing reference-free metrics. To refine the dataset, we: (1) filter the MLLM-generated summaries by training a critic model on human annotations and using its predictions to remove low-quality summaries; (2) fine-tune the MLLM with the filtered high-quality summaries; (3) use the fine-tuned model in turn to regenerate the summaries. This self-refinement process notably improves summary quality, as measured by human judgments and automatic multimodal metrics, resulting in a valuable dataset for multimodal summarization research.(1)

In this work, We present Unified Embeddings for Multimodal Retrieval (UNIMUR), a simple but effective approach that embeds multimodal inputs and retrieves visual and textual outputs via frozen Large Language Models (LLMs). Specifically, UNIMUR jointly retrieves multimodal outputs via unified multimodal embedding and applies dual alignment training to account for both visual and textual semantics. Thus, unlike previous approaches, UNIMUR significantly reduces LLM's modality bias towards generating text-only outputs. Meanwhile, the proposed unified multimodal embedding mitigates the inconsistency between visual and textual outputs and provides coherent multimodal outputs. Empirically, UNIMUR also achieves strong image/text retrieval ability outperforming existing approaches on zero-shot multimodal response retrieval on MMDialog, improving the overall R@1 by 6.5% while boosting the image retrieval rate and having better cross-modal consistency on multimodal outputs. UNIMUR also achieves 2.4% and 3.9% improvement on context-based image retrieval tasks on MMDialog and VisDial respectively when compared to previous approaches, validating its generalization ability across multiple tasks.

Neural models for abstractive summarization tend to generate output that is fluent and well-formed but lacks semantic faithfulness, or factuality, with respect to the input documents. In this paper, we analyze the tradeoff between abstractiveness and factuality of generated summaries across multiple datasets and models, using extensive human evaluations of factuality. In our analysis, we visualize the rates of change in factuality as we gradually increase abstractiveness using a decoding constraint, and we observe that, while increased abstractiveness generally leads to a drop in factuality, the rate of factuality decay depends on factors such as the data that the system was trained on. We introduce two datasets with human factuality judgements; one containing 10.2k generated summaries with systematically varied degrees of abstractiveness; the other containing 4.2k summaries from five different summarization models. We propose new factuality metrics that adjust for the degree of abstractiveness, and we use them to compare the abstractiveness-adjusted factuality of previous summarization works, providing baselines for future work.

Prompts have been shown to be an effective method to adapt a frozen Pretrained Language Model (PLM) to perform well on downstream tasks. Prompts can be represented by a human-engineered word sequence or by a learned continuous embedding. In this work, we investigate conditional and compositional differentiable prompting. We propose a new model, Prompt Production System (ProPS), which learns to transform task instructions or input metadata, into continuous prompts that elicit task-specific outputs from the PLM. Our model uses a modular network structure based on our neural formulation of Production Systems, which allows the model to learn discrete rules — neural functions that learn to specialize in transforming particular prompt input patterns, making it suitable for compositional transfer learning and few-shot learning. We present extensive empirical and theoretical analysis and show that ProPS consistently surpasses other PLM adaptation techniques, and often improves upon fully fine-tuned models, on compositional generalization tasks, controllable summarization and multilingual translation, while needing fewer trainable parameters.

Despite significant progress in understanding and improving faithfulness in abstractive summarization, the question of how decoding strategies affect faithfulness is less studied. We present a systematic study of the effect of generation techniques such as beam search and nucleus sampling on faithfulness in abstractive summarization. We find a consistent trend where beam search with large beam sizes produces the most faithful summaries while nucleus sampling generates the least faithful ones. We propose two faithfulness-aware generation methods to further improve faithfulness over current generation techniques: (1) ranking candidates generated by beam search using automatic faithfulness metrics and (2) incorporating lookahead heuristics that produce a faithfulness score on the future summary. We show that both generation methods significantly improve faithfulness across two datasets as evaluated by four automatic faithfulness metrics and human evaluation. To reduce computational cost, we demonstrate a simple distillation approach that allows the model to generate faithful summaries with just greedy decoding. Our code is publicly available at https://github.com/amazon-science/faithful-summarization-generation

Readability refers to how easily a reader can understand a written text. Several factors affect the readability level, such as the complexity of the text, its subject matter, and the reader's background knowledge. Generating summaries based on different readability levels is critical for enabling knowledge consumption by diverse audiences. However, current text generation approaches lack refined control, resulting in texts that are not customized to readers' proficiency levels. In this work, we bridge this gap and study techniques to generate summaries at specified readability levels. Unlike previous methods that focus on a specific readability level (e.g., lay summarization), we generate summaries with fine-grained control over their readability. We develop three text generation techniques for controlling readability: (1) instruction-based readability control, (2) reinforcement learning to minimize the gap between requested and observed readability and (3) a decoding approach that uses lookahead to estimate the readability of upcoming decoding steps. We show that our generation methods significantly improve readability control on news summarization (CNN/DM dataset), as measured by various readability metrics and human judgement, establishing strong baselines for controllable readability in summarization.

Information extraction (IE) and summarization are closely related, both tasked with presenting a subset of the information contained in a natural language text. However, while IE extracts structural representations, summarization aims to abstract the most salient information into a generated text summary – thus potentially encountering the technical limitations of current text generation methods (e.g., hallucination). To mitigate this risk, this work uses structured IE graphs to enhance the abstractive summarization task. Specifically, we focus on improving Multi-Document Summarization (MDS) performance by using cross-document IE output, incorporating two novel components: (1) the use of auxiliary entity and event recognition systems to focus the summary generation model; (2) incorporating an alignment loss between IE nodes and their text spans to reduce inconsistencies between the IE graphs and text representations. Operationally, both the IE nodes and corresponding text spans are projected into the same embedding space and pairwise distance is minimized. Experimental results on multiple MDS benchmarks show that summaries generated by our model are more factually consistent with the source documents than baseline models while maintaining the same level of abstractiveness.

Generating concise summaries of news events is a challenging natural language processing task. While journalists often curate timelines to highlight key sub-events, newcomers to a news event face challenges in catching up on its historical context. In this paper, we address this need by introducing the task of background news summarization, which complements each timeline update with a background summary of relevant preceding events. We construct a dataset by merging existing timeline datasets and asking human annotators to write a background summary for each timestep of each news event. We establish strong baseline performance using state-of-the-art summarization systems and propose a query-focused variant to generate background summaries. To evaluate background summary quality, we present a question-answering-based evaluation metric, Background Utility Score (BUS), which measures the percentage of questions about a current event timestep that a background summary answers. Our experiments show the effectiveness of instruction fine-tuned systems such as Flan-T5, in addition to strong zero-shot performance using GPT-3.5.

Despite recent improvements in abstractive summarization, most current approaches generate summaries that are not factually consistent with the source document, severely restricting their trust and usage in real-world applications. Recent works have shown promising improvements in factuality error identification using text or dependency arc entailments; however, they do not consider the entire semantic graph simultaneously. To this end, we propose FactGraph, a method that decomposes the document and the summary into structured meaning representations (MR), which are more suitable for factuality evaluation. MRs describe core semantic concepts and their relations, aggregating the main content in both document and summary in a canonical form, and reducing data sparsity. FactGraph encodes such graphs using a graph encoder augmented with structure-aware adapters to capture interactions among the concepts based on the graph connectivity, along with text representations using an adapter-based text encoder. Experiments on different benchmarks for evaluating factuality show that FactGraph outperforms previous approaches by up to 15%. Furthermore, FactGraph improves performance on identifying content verifiability errors and better captures subsentence-level factual inconsistencies.

Abstractive summarization models are typically pre-trained on large amounts of generic texts, then fine-tuned on tens or hundreds of thousands of annotated samples. However, in opinion summarization, large annotated datasets of reviews paired with reference summaries are not available and would be expensive to create. This calls for fine-tuning methods robust to overfitting on small datasets. In addition, generically pre-trained models are often not accustomed to the specifics of customer reviews and, after fine-tuning, yield summaries with disfluencies and semantic mistakes. To address these problems, we utilize an efficient few-shot method based on adapters which, as we show, can easily store in-domain knowledge. Instead of fine-tuning the entire model, we add adapters and pre-train them in a task-specific way on a large corpus of unannotated customer reviews, using held-out reviews as pseudo summaries. Then, fine-tune the adapters on the small available human-annotated dataset. We show that this self-supervised adapter pre-training improves summary quality over standard fine-tuning by 2.0 and 1.3 ROUGE-L points on the Amazon and Yelp datasets, respectively. Finally, for summary personalization, we condition on aspect keyword queries, automatically created from generic datasets. In the same vein, we pre-train the adapters in a query-based manner on customer reviews and then fine-tune them on annotated datasets. This results in better-organized summary content reflected in improved coherence and fewer redundancies.

This paper presents an efficient graph-enhanced approach to multi-document summarization (MDS) with an encoder-decoder Transformer model. This model is based on recent advances in pre-training both encoder and decoder on very large text data (Lewis et al., 2019), and it incorporates an efficient encoding mechanism (Beltagy et al., 2020) that avoids the quadratic memory growth typical for traditional Transformers. We show that this powerful combination not only scales to large input documents commonly found when summarizing news clusters; it also enables us to process additional input in the form of auxiliary graph representations, which we derive from the multi-document clusters. We present a mechanism to incorporate such graph information into the encoder-decoder model that was pre-trained on text only. Our approach leads to significant improvements on the Multi-News dataset, overall leading to an average 1.8 ROUGE score improvement over previous work (Li et al., 2020). We also show improvements in a transfer-only setup on the DUC-2004 dataset. The graph encodings lead to summaries that are more abstractive. Human evaluation shows that they are also more informative and factually more consistent with their input documents.

We analyze the tradeoff between factuality and abstractiveness of summaries. We introduce abstractiveness constraints to control the degree of abstractiveness at decoding time, and we apply this technique to characterize the abstractiveness-factuality tradeoff across multiple widely-studied datasets, using extensive human evaluations. We train a neural summarization model on each dataset and visualize the rates of change in factuality as we gradually increase abstractiveness using our abstractiveness constraints. We observe that, while factuality generally drops with increased abstractiveness, different datasets lead to different rates of factuality decay. We propose new measures to quantify the tradeoff between factuality and abstractiveness, incl. µ QAGS, which balances factuality with abstractiveness. We also quantify this tradeoff in previous works, aiming to establish baselines for the abstractiveness-factuality tradeoff that future publications can compare against.

Pre-trained and fine-tuned news summarizers are expected to generalize to news articles unseen in the fine-tuning (training) phase. However, these articles often contain specifics, such as new events and people, a summarizer could not learn about in training. This applies to scenarios such as a news publisher training a summarizer on dated news and summarizing incoming recent news. In this work, we explore the first application of transductive learning to summarization where we further fine-tune models on test set inputs. Specifically, we construct pseudo summaries from salient article sentences and input randomly masked articles. Moreover, this approach is also beneficial in the fine-tuning phase, where we jointly predict extractive pseudo references and abstractive gold summaries in the training set. We show that our approach yields state-of-the-art results on CNN/DM and NYT datasets, improving ROUGE-L by 1.05 and 0.74, respectively. Importantly, our approach does not require any changes of the original architecture. Moreover, we show the benefits of transduction from dated to more recent CNN news. Finally, through human and automatic evaluation, we demonstrate improvements in summary abstractiveness and coherence.