abstractive summarization

To address severe information loss in hundred-document-scale multi-document event summarization, this work systematically compares compression-based (multi-stage pipeline) and full-text-based (direct long-context modeling) approaches. Leveraging long-context Transformers—including Llama-3.1, Command-R, and Jamba-1.5-Mini—augmented with retrieval enhancement, hierarchical compression, and incremental summarization, we find that full-text modeling combined with retrieval achieves the best overall performance; in contrast, compression methods better preserve local information at intermediate stages but suffer from global context loss and consequent information decay. Building on these insights, we propose a hybrid paradigm that synergistically integrates compression and full-text modeling at critical stages. Empirical evaluation demonstrates that this architecture significantly improves summary coherence, coverage, and factual consistency. The approach offers a scalable, principled solution for large-scale event summarization, advancing the state of the art in handling ultra-long document collections.

To address factual inconsistency, information loss, and poor traceability in long-document summarization, this paper proposes a sentence-level highlighting-guided self-planning generation framework. First, it identifies salient sentences via importance modeling and generates a traceable content plan; subsequently, summary generation is conditioned on this plan, effectively decoupling content selection from surface realization. This novel paradigm significantly enhances summary faithfulness and fine-grained detail retention. On the GovReport benchmark, our approach achieves a +4.1-point improvement in ROUGE-L and a 35% gain in SummaC score. Qualitative analysis confirms more complete preservation of critical details, as well as improved cross-domain accuracy and analytical depth in generated summaries.

This work investigates whether plan-guided summarization improves faithfulness of small language models (SLMs) on long narrative texts. Addressing the susceptibility of existing fine-grained plans to hallucination, we propose a high-level planning method grounded in narrative structure. Through automated evaluation and rigorous human assessment—specifically targeting faithfulness and hallucination—we find that neither fine-grained nor our novel high-level plan guidance significantly outperforms the plan-free baseline. The root cause is high hallucination rates inherent in the plans themselves, which undermine guidance efficacy and even propagate factual errors. To our knowledge, this is the first systematic study exposing the limitations of plan-guided summarization for complex narratives. Our findings caution against uncritical adoption of planning in long-text and low-resource settings, where plan hallucinations critically compromise reliability. The study provides key empirical evidence for developing trustworthy abstractive summarization systems, highlighting the necessity of hallucination-robust planning mechanisms.

To address hallucination accumulation and factual inconsistency arising from hierarchical merging in abstractive summarization of ultra-long documents (>100K tokens), this paper proposes a context-aware hierarchical merging paradigm. The method systematically introduces three novel context-enhancement mechanisms: (i) key-segment replacement, (ii) prompt-guided summary refinement, and (iii) implicit provenance via citation alignment. Empirical analysis identifies the combination of refinement and extractive key-segment selection as optimal. Evaluated on legal and narrative domain datasets using Llama 3.1-based models, the approach significantly outperforms zero-shot and standard hierarchical merging baselines. It achieves simultaneous improvements in factual consistency and coherence, empirically validating the critical role of source-document context at every summarization layer.

Large language models (LLMs) struggle with long-document summarization due to event information dispersion and disrupted narrative logic, leading to inaccurate comprehension. To address this, we propose HERA—a novel framework introducing *context packing and reordering*. First, the input document is segmented based on semantic structure; second, cross-segment retrieval identifies fragments pertaining to the same event, which are then reordered into a logically coherent context; finally, the reordered context is fed into a zero-shot LLM for summary generation. HERA requires no fine-tuning, external training, or additional computational resources. Crucially, it enhances the model’s ability to capture event-level causal and temporal dependencies. Evaluated on two major long-document summarization benchmarks, HERA consistently outperforms strong baselines across ROUGE, BERTScore, and faithfulness metrics—demonstrating significant and robust improvements in both informativeness and factual consistency.

This work proposes GloSA-sum, a novel text summarization approach that addresses the limitations of existing methods—namely, their frequent neglect of global document structure, which undermines logical coherence, and the high computational cost of large language model–based alternatives. GloSA-sum is the first to integrate topological data analysis (TDA) into summarization by constructing a semantic-weighted graph and leveraging persistent homology to identify core semantics and logical backbones. It employs a preservation pooling mechanism to retain critical structural elements and combines a hierarchical strategy with a lightweight iterative process to efficiently handle long documents. Evaluated across multiple datasets, GloSA-sum significantly reduces redundancy while maintaining semantic fidelity and logical integrity, and further enhances the performance of downstream large language model tasks.

This work addresses the “length collapse” phenomenon in large language models, wherein coherence and controllability significantly degrade when generating open-ended texts exceeding 2,000 words. To mitigate this, the authors propose the Interleaved Structural Chain-of-Thought (IS-CoT) framework, which introduces, for the first time, an internal mechanism that dynamically interleaves structured reasoning with text generation. By embedding a Plan-Write-Reflect loop, IS-CoT enables continuous self-adjustment of writing strategies without external intervention while maintaining global alignment. Leveraging a high-quality dataset of interleaved reasoning trajectories constructed via multi-teacher distillation, the authors train IS-Writer-8B, which achieves state-of-the-art performance on benchmarks such as LongBench-Write—surpassing DeepSeek-V3.2 by 3.08 points—and demonstrates length compliance and coherence comparable to much larger closed-source models.

This study investigates the effectiveness and limitations of reasoning capabilities in large language models (LLMs) for abstractive summarization. Addressing the underexamined hypothesis that “reasoning universally improves summary quality,” we systematically evaluate eight explicit and implicit reasoning strategies—including chain-of-thought, self-consistency, and stepwise reasoning—across three prominent reasoning-oriented LLMs and eight diverse benchmark datasets. Results reveal a critical trade-off: explicit reasoning enhances linguistic fluency but degrades factual consistency, whereas implicit reasoning exhibits the opposite pattern; moreover, increasing internal reasoning steps does not consistently improve performance. Our core contribution is the empirical identification of an inherent tension between summary quality and factual fidelity in abstractive summarization, leading to the principle that “faithful compression outweighs excessive reasoning.” This work provides both empirical evidence and methodological guidance for the principled design and deployment of reasoning mechanisms in LLM-based summarization systems.

Traditional multi-document summarization struggles with integrating multi-perspective narrative texts (e.g., legal testimonies, historical accounts) due to its overemphasis on concision, compromising temporal coherence and factual completeness. To address this, we formally introduce *Narrative Consolidation*—a novel task requiring precise chronological ordering, comprehensive content coverage, and seamless integration of complementary details. We propose the Temporal Alignment Event Graph (TAEG), a graph-based model that explicitly unifies event alignment with temporal structure modeling. TAEG incorporates graph centrality measures (e.g., PageRank) to automatically select authoritative narrative versions. Evaluated on the Four Gospels dataset, our method achieves perfect temporal consistency (Kendall’s Tau = 1.000) and improves ROUGE-L F1 by 357.2% over baselines, demonstrating that explicit temporal backbone modeling is essential for effective narrative consolidation.

Large language models (LLMs) struggle to efficiently process raw web content due to its unstructured formatting and contextual constraints. Method: This paper proposes a chain-of-summarization approach grounded in the dialectical “thesis–antithesis–synthesis” framework, iteratively refining summaries through multi-turn questioning and abstraction to produce concise, information-dense plain-text summaries—without task-specific fine-tuning. The method jointly addresses both explicit and implicit information needs, enhancing interpretability and downstream utility. Results: Experiments on TriviaQA, TruthfulQA, and SQuAD demonstrate that our approach outperforms zero-shot LLM baselines by 66% and surpasses dedicated abstractive summarizers (e.g., BRIO, PEGASUS) by 27%, while achieving higher question-answering accuracy using fewer tokens.