Long sequences are critical for applications like RAG, long document summarization, multi-modality, etc., and modern LLMs, like Llama 4 Scout, support max sequence length of up to 10 million tokens. However, outside of enterprise labs, long sequence training is challenging for the AI community with limited system support in the open-source space. Out-of-box, even on a modern NVIDIA H100 80GB GPU cluster, training Llama 8B model with sequence over 32K runs out of memory on a basic Hugging Face (HF) model due to two reasons: i) LLM training workloads are not optimized to fully leverage a single GPU memory, ii) existing solutions for leveraging multiple GPU memory are not easily available to HF models, making long sequence training inaccessible. We address this with Arctic Long Sequence Training (ALST). It offers a combination of attention-agnostic single GPU and multi-GPU memory optimizations, that enables it to support out-of-box training of multi-million sequence length for a wide variety of HF models. ALST supports training Meta's Llama 8B model with 500K sequence length on a single H100 GPU, 3.7M on a single 8xH100 GPU node, and over 15M on a 4 node cluster, an increase of over 400x compared to the 32K baseline for the latter. ALST is fully compatible with HF models and open-sourced via Deepspeed https://www.deepspeed.ai/tutorials/ulysses-alst-sequence-pallellism/ and Arctic Training https://github.com/snowflakedb/ArcticTraining/blob/main/projects/sequence-parallelism/README.md.
Recent large language models (LLMs) achieve impressive performance in source-conditioned text generation but often fail to correctly provide fine-grained attributions for their outputs, undermining verifiability and trust. Moreover, existing attribution methods do not explain how and why models leverage the provided source documents to generate their final responses, limiting interpretability. To overcome these challenges, we introduce a modular generation framework, GenerationPrograms, inspired by recent advancements in executable "code agent" architectures. Unlike conventional generation methods that simultaneously generate outputs and attributions or rely on post-hoc attribution, GenerationPrograms decomposes the process into two distinct stages: first, creating an executable program plan composed of modular text operations (such as paraphrasing, compression, and fusion) explicitly tailored to the query, and second, executing these operations following the program's specified instructions to produce the final response. Empirical evaluations demonstrate that GenerationPrograms significantly improves attribution quality at both the document level and sentence level across two long-form question-answering tasks and a multi-document summarization task. We further demonstrate that GenerationPrograms can effectively function as a post-hoc attribution method, outperforming traditional techniques in recovering accurate attributions. In addition, the interpretable programs generated by GenerationPrograms enable localized refinement through modular-level improvements that further enhance overall attribution quality.
Fairness in multi-document summarization (MDS) is crucial for providing comprehensive views across documents with diverse social attribute values, which can significantly impact decision-making. For example, a summarization system that tends to overrepresent negative reviews of products can mislead customers into disregarding good products. Previous works measure fairness in MDS at two levels: summary-level and corpus-level. While summary-level fairness focuses on individual summaries, corpus-level fairness focuses on a corpus of summaries. Recent methods primarily focus on summary-level fairness. We propose FairPO, a preference tuning method that focuses on both summary-level and corpus-level fairness in MDS. To improve summary-level fairness, we propose to generate preference pairs by perturbing document sets. To improve corpus-level fairness, we propose fairness-aware preference tuning by dynamically adjusting the weights of preference pairs. Our experiments show that FairPO outperforms strong baselines while maintaining the critical qualities of summaries. The code is available at https://github.com/leehaoyuan/coverage_fairnes.
Integrating structured information has long improved the quality of abstractive summarization, particularly in retaining salient content. In this work, we focus on a specific form of structure: argument roles, which are crucial for summarizing documents in high-stakes domains such as law. We investigate whether instruction-tuned large language models (LLMs) adequately preserve this information. To this end, we introduce Argument Representation Coverage (ARC), a framework for measuring how well LLM-generated summaries capture salient arguments. Using ARC, we analyze summaries produced by three open-weight LLMs in two domains where argument roles are central: long legal opinions and scientific articles. Our results show that while LLMs cover salient argument roles to some extent, critical information is often omitted in generated summaries, particularly when arguments are sparsely distributed throughout the input. Further, we use ARC to uncover behavioral patterns -- specifically, how the positional bias of LLM context windows and role-specific preferences impact the coverage of key arguments in generated summaries, emphasizing the need for more argument-aware summarization strategies.
This paper introduces PeopleJoin, a benchmark for evaluating LM-mediated collaborative problem solving. Given a user request, PeopleJoin agents must identify teammates who might be able to assist, converse with these teammates to gather information, and finally compile a useful answer or summary for the original user. PeopleJoin comprises two evaluation domains: PeopleJoin-QA, focused on questions about tabular data, and PeopleJoin-DocCreation, focused on document creation tasks. The two domains are adapted from existing NLP benchmarks for database question answering and multi-document summarization; here, however, the information needed to complete these tasks is distributed across synthetic ``organizations'' of 2--20 users, simulating natural multi-user collaboration scenarios. We implemented several popular LM agent architectures, evaluating their accuracy and efficiency at completing tasks, and highlight new research questions that can be studied using PeopleJoin.
The exponential growth of scientific publications has made it increasingly difficult for researchers to stay updated and synthesize knowledge effectively. This paper presents XSum, a modular pipeline for multi-document summarization (MDS) in the scientific domain using Retrieval-Augmented Generation (RAG). The pipeline includes two core components: a question-generation module and an editor module. The question-generation module dynamically generates questions adapted to the input papers, ensuring the retrieval of relevant and accurate information. The editor module synthesizes the retrieved content into coherent and well-structured summaries that adhere to academic standards for proper citation. Evaluated on the SurveySum dataset, XSum demonstrates strong performance, achieving considerable improvements in metrics such as CheckEval, G-Eval and Ref-F1 compared to existing approaches. This work provides a transparent, adaptable framework for scientific summarization with potential applications in a wide range of domains. Code available at https://github.com/webis-de/scolia25-xsum
Large Language Models (LLMs) have significantly advanced text generation capabilities, including tasks like summarization, often producing coherent and fluent outputs. However, faithfulness to source material remains a significant challenge due to the generation of hallucinations. While extensive research focuses on detecting and reducing these inaccuracies, less attention has been paid to the positional distribution of hallucination within generated text, particularly in long outputs. In this work, we investigate where hallucinations occur in LLM-based long response generation, using long document summarization as a key case study. Focusing on the challenging setting of long context-aware long response generation, we find a consistent and concerning phenomenon: hallucinations tend to concentrate disproportionately in the latter parts of the generated long response. To understand this bias, we explore potential contributing factors related to the dynamics of attention and decoding over long sequences. Furthermore, we investigate methods to mitigate this positional hallucination, aiming to improve faithfulness specifically in the concluding segments of long outputs.
As Large Language Models (LLMs) are increasingly applied to document-based tasks - such as document summarization, question answering, and information extraction - where user requirements focus on retrieving information from provided documents rather than relying on the model's parametric knowledge, ensuring the trustworthiness and interpretability of these systems has become a critical concern. A central approach to addressing this challenge is attribution, which involves tracing the generated outputs back to their source documents. However, since LLMs can produce inaccurate or imprecise responses, it is crucial to assess the reliability of these citations. To tackle this, our work proposes two techniques. (1) A zero-shot approach that frames attribution as a straightforward textual entailment task. Our method using flan-ul2 demonstrates an improvement of 0.27% and 2.4% over the best baseline of ID and OOD sets of AttributionBench, respectively. (2) We also explore the role of the attention mechanism in enhancing the attribution process. Using a smaller LLM, flan-t5-small, the F1 scores outperform the baseline across almost all layers except layer 4 and layers 8 through 11.
In the field of multi-document summarization (MDS), transformer-based models have demonstrated remarkable success, yet they suffer an input length limitation. Current methods apply truncation after the retrieval process to fit the context length; however, they heavily depend on manually well-crafted queries, which are impractical to create for each document set for MDS. Additionally, these methods retrieve information at a coarse granularity, leading to the inclusion of irrelevant content. To address these issues, we propose a novel retrieval-based framework that integrates query selection and document ranking and shortening into a unified process. Our approach identifies the most salient elementary discourse units (EDUs) from input documents and utilizes them as latent queries. These queries guide the document ranking by calculating relevance scores. Instead of traditional truncation, our approach filters out irrelevant EDUs to fit the context length, ensuring that only critical information is preserved for summarization. We evaluate our framework on multiple MDS datasets, demonstrating consistent improvements in ROUGE metrics while confirming its scalability and flexibility across diverse model architectures. Additionally, we validate its effectiveness through an in-depth analysis, emphasizing its ability to dynamically select appropriate queries and accurately rank documents based on their relevance scores. These results demonstrate that our framework effectively addresses context-length constraints, establishing it as a robust and reliable solution for MDS.
Recent advances in long-context reasoning abilities of language models led to interesting applications in large-scale multi-document summarization. However, prior work has shown that these long-context models are not effective at their claimed context windows. To this end, retrieval-augmented systems provide an efficient and effective alternative. However, their performance can be highly sensitive to the choice of retrieval context length. In this work, we present a hybrid method that combines retrieval-augmented systems with long-context windows supported by recent language models. Our method first estimates the optimal retrieval length as a function of the retriever, summarizer, and dataset. On a randomly sampled subset of the dataset, we use a panel of LLMs to generate a pool of silver references. We use these silver references to estimate the optimal context length for a given RAG system configuration. Our results on the multi-document summarization task showcase the effectiveness of our method across model classes and sizes. We compare against length estimates from strong long-context benchmarks such as RULER and HELMET. Our analysis also highlights the effectiveness of our estimation method for very long-context LMs and its generalization to new classes of LMs.
Generating long, coherent text remains a challenge for large language models (LLMs), as they lack hierarchical planning and structured organization in discourse generation. We introduce Structural Alignment, a novel method that aligns LLMs with human-like discourse structures to enhance long-form text generation. By integrating linguistically grounded discourse frameworks into reinforcement learning, our approach guides models to produce coherent and well-organized outputs. We employ a dense reward scheme within a Proximal Policy Optimization framework, assigning fine-grained, token-level rewards based on the discourse distinctiveness relative to human writing. Two complementary reward models are evaluated: the first improves readability by scoring surface-level textual features to provide explicit structuring, while the second reinforces deeper coherence and rhetorical sophistication by analyzing global discourse patterns through hierarchical discourse motifs, outperforming both standard and RLHF-enhanced models in tasks such as essay generation and long-document summarization. All training data and code will be publicly shared at https://github.com/minnesotanlp/struct_align.
We investigate a critical yet under-explored question in Large Vision-Language Models (LVLMs): Do LVLMs genuinely comprehend interleaved image-text in the document? Existing document understanding benchmarks often assess LVLMs using question-answer formats, which are information-sparse and difficult to guarantee the coverage of long-range dependencies. To address this issue, we introduce a novel and challenging Multimodal Document Summarization Benchmark (M-DocSum-Bench), which comprises 500 high-quality arXiv papers, along with interleaved multimodal summaries aligned with human preferences. M-DocSum-Bench is a reference-based generation task and necessitates the generation of interleaved image-text summaries using provided reference images, thereby simultaneously evaluating capabilities in understanding, reasoning, localization, and summarization within complex multimodal document scenarios. To facilitate this benchmark, we develop an automated framework to construct summaries and propose a fine-grained evaluation method called M-DocEval. Moreover, we further develop a robust summarization baseline, i.e., M-DocSum-7B, by progressive two-stage training with diverse instruction and preference data. The extensive results on our M-DocSum-Bench reveal that the leading LVLMs struggle to maintain coherence and accurately integrate information within long and interleaved contexts, often exhibiting confusion between similar images and a lack of robustness. Notably, M-DocSum-7B achieves state-of-the-art performance compared to larger and closed-source models (including GPT-4o, Gemini Pro, Claude-3.5-Sonnet and Qwen2.5-VL-72B, etc.), demonstrating the potential of LVLMs for improved interleaved image-text understanding. The code, data, and models are available at https://github.com/stepfun-ai/M-DocSum-Bench.
Abstractive multi-document summarization (MDS) is the task of automatically summarizing information in multiple documents, from news articles to conversations with multiple speakers. The training approaches for current MDS models can be grouped into four approaches: end-to-end with special pre-training ("direct"), chunk-then-summarize, extract-then-summarize, and inference with GPT-style models. In this work, we evaluate MDS models across training approaches, domains, and dimensions (reference similarity, quality, and factuality), to analyze how and why models trained on one domain can fail to summarize documents from another (News, Science, and Conversation) in the zero-shot domain transfer setting. We define domain-transfer "failure" as a decrease in factuality, higher deviation from the target, and a general decrease in summary quality. In addition to exploring domain transfer for MDS models, we examine potential issues with applying popular summarization metrics out-of-the-box.
Document structure analysis, aka document layout analysis, is crucial for understanding both the physical layout and logical structure of documents, serving information retrieval, document summarization, knowledge extraction, etc. Hierarchical Document Structure Analysis (HDSA) specifically aims to restore the hierarchical structure of documents created using authoring software with hierarchical schemas. Previous research has primarily followed two approaches: one focuses on tackling specific subtasks of HDSA in isolation, such as table detection or reading order prediction, while the other adopts a unified framework that uses multiple branches or modules, each designed to address a distinct task. In this work, we propose a unified relation prediction approach for HDSA, called UniHDSA, which treats various HDSA sub-tasks as relation prediction problems and consolidates relation prediction labels into a unified label space. This allows a single relation prediction module to handle multiple tasks simultaneously, whether at a page-level or document-level structure analysis. To validate the effectiveness of UniHDSA, we develop a multimodal end-to-end system based on Transformer architectures. Extensive experimental results demonstrate that our approach achieves state-of-the-art performance on a hierarchical document structure analysis benchmark, Comp-HRDoc, and competitive results on a large-scale document layout analysis dataset, DocLayNet, effectively illustrating the superiority of our method across all sub-tasks. The Comp-HRDoc benchmark and UniHDSA's configurations are publicly available at https://github.com/microsoft/CompHRDoc.
The applications of Large Language Models (LLMs) in political science are rapidly expanding. This paper demonstrates how LLMs, when augmented with predefined functions and specialized tools, can serve as dynamic agents capable of streamlining tasks such as data collection, preprocessing, and analysis. Central to this approach is agentic retrieval-augmented generation (Agentic RAG), which equips LLMs with action-calling capabilities for interaction with external knowledge bases. Beyond information retrieval, LLM agents may incorporate modular tools for tasks like document summarization, transcript coding, qualitative variable classification, and statistical modeling. To demonstrate the potential of this approach, we introduce CongressRA, an LLM agent designed to support scholars studying the U.S. Congress. Through this example, we highlight how LLM agents can reduce the costs of replicating, testing, and extending empirical research using the domain-specific data that drives the study of political institutions.
Automatic patent summarization approaches that help in the patent analysis and comprehension procedure are in high demand due to the colossal growth of innovations. The development of natural language processing (NLP), text mining, and deep learning has notably amplified the efficacy of text summarization models for abundant types of documents. Summarizing patent text remains a pertinent challenge due to the labyrinthine writing style of these documents, which includes technical and legal intricacies. Additionally, these patent document contents are considerably lengthier than archetypal documents, which complicates the process of extracting pertinent information for summarization. Embodying extractive and abstractive text summarization methodologies into a hybrid framework, this study proposes a system for efficiently creating abstractive summaries of patent records. The procedure involves leveraging the LexRank graph-based algorithm to retrieve the important sentences from input parent texts, then utilizing a Bidirectional Auto-Regressive Transformer (BART) model that has been fine-tuned using Low-Ranking Adaptation (LoRA) for producing text summaries. This is accompanied by methodical testing and evaluation strategies. Furthermore, the author employed certain meta-learning techniques to achieve Domain Generalization (DG) of the abstractive component across multiple patent fields.
This work tackles the problem of overoptimization in reinforcement learning from human feedback (RLHF), a prevalent technique for aligning models with human preferences. RLHF relies on reward or preference models trained on \emph{fixed preference datasets}, and these models are unreliable when evaluated outside the support of this preference data, leading to the common reward or preference hacking phenomenon. We propose novel, pessimistic objectives for RLHF which are provably robust to overoptimization through the use of pessimism in the face of uncertainty, and design practical algorithms, P3O and PRPO, to optimize these objectives. Our approach is derived for the general preference optimization setting, but can be used with reward models as well. We evaluate P3O and PRPO on the tasks of fine-tuning language models for document summarization and creating helpful assistants, demonstrating remarkable resilience to overoptimization.
Recent advances in test-time scaling have shown promising results in improving Large Language Model (LLM) performance through strategic computation allocation during inference. While this approach has demonstrated strong improvements in logical and mathematical reasoning tasks, its application to natural language generation (NLG), particularly summarization, remains unexplored. Multi-Document Summarization (MDS), a fundamental task in NLG, presents unique challenges by requiring models to extract and synthesize essential information across multiple lengthy documents. Unlike reasoning tasks, MDS demands a more nuanced approach to prompt design and ensemble methods, as no single "best" prompt can satisfy diverse summarization requirements. We propose a novel framework leveraging test-time scaling for MDS. Our approach employs prompt ensemble techniques to generate multiple candidate summaries using various prompts, then combines them with an aggregator to produce a refined summary. To evaluate our method effectively, we also introduce two new LLM-based metrics: the Consistency-Aware Preference (CAP) score and LLM Atom-Content-Unit (LLM-ACU) score, which assess summary quality while addressing the positional bias inherent in traditional automatic evaluation. Our extensive experiments demonstrate that this framework significantly enhances summary quality while also revealing the practical scaling boundaries to MDS tasks.
This paper introduces Leaderboard Auto Generation (LAG), a novel and well-organized framework for automatic generation of leaderboards on a given research topic in rapidly evolving fields like Artificial Intelligence (AI). Faced with a large number of AI papers updated daily, it becomes difficult for researchers to track every paper's proposed methods, experimental results, and settings, prompting the need for efficient automatic leaderboard construction. While large language models (LLMs) offer promise in automating this process, challenges such as multi-document summarization, leaderboard generation, and experiment fair comparison still remain under exploration. LAG solves these challenges through a systematic approach that involves the paper collection, experiment results extraction and integration, leaderboard generation, and quality evaluation. Our contributions include a comprehensive solution to the leaderboard construction problem, a reliable evaluation method, and experimental results showing the high quality of leaderboards.
Contextual memory integration remains a high challenge in the development of language models, particularly in tasks that require maintaining coherence over extended sequences. Traditional approaches, such as self-attention mechanisms and memory-augmented architectures, often prioritize short-term dependencies, leading to fragmentation and inconsistency in long-range contextual understanding. Inspired by principles of synaptic plasticity observed in biological neural systems, a novel mechanism, Synaptic Resonance, is introduced to dynamically reinforce relevant memory pathways during training and inference. Unlike static memory representations, this mechanism continuously adjusts synaptic weight matrices based on contextual relevance, allowing for improved information retention without excessive computational overhead. Evaluations conducted on an open-source language model demonstrate reductions in perplexity, enhancements in contextual coherence, and increased robustness against input noise, highlighting the effectiveness of reinforcement-driven memory modulation. Comparative analysis against baseline models further reveals that the proposed approach achieves higher memory retention efficiency while maintaining computational feasibility. The architectural modifications integrate seamlessly into existing transformer-based frameworks, ensuring stable convergence and efficient inference without sacrificing scalability. Applications benefiting from improved long-term contextual consistency, such as dialogue systems and document summarization, stand to gain from this approach. Empirical findings suggest that dynamically reinforced memory pathways offer a promising alternative to conventional memory mechanisms, addressing longstanding limitations in extended sequence modeling.