Paper deep dive
From Single- to Cross-Document: Benchmarking Multi-Granularity Event Analysis of Large Language Models
Tao Wen, Shuai Shao, Pei Ke, Xu Han, Jie Zou, Guannan Li, Tao Tian, Jinjie Qiu, Lan Wang, Ke Qin
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 8/1/2026, 1:56:51 AM
Summary
The paper introduces MiGUE-Bench, a systematic benchmark for evaluating Large Language Models (LLMs) on multi-granularity event analysis, ranging from single-document detection to cross-document reasoning. It proposes MiGUE-Pipeline, an LLM-driven self-correcting annotation framework for scalable data construction, and defines four core tasks: event detection, relation reasoning, structure induction, and future prediction.
Entities (10)
Relation Signals (8)
MiGUE-Bench β includestask β Event Detection
confidence 95% Β· we design four core tasks in our benchmark, i.e., event detection, relation reasoning, structure induction, and future prediction
MiGUE-Bench β includestask β Relation Reasoning
confidence 95% Β· we design four core tasks in our benchmark, i.e., event detection, relation reasoning, structure induction, and future prediction
MiGUE-Bench β includestask β Structure Induction
confidence 95% Β· we design four core tasks in our benchmark, i.e., event detection, relation reasoning, structure induction, and future prediction
MiGUE-Bench β includestask β Future Prediction
confidence 95% Β· we design four core tasks in our benchmark, i.e., event detection, relation reasoning, structure induction, and future prediction
MiGUE-Pipeline β supports β MiGUE-Bench
confidence 93% Β· To support large-scale evaluation, we first develop an LLM-driven self-correcting annotation framework called MiGUE-Pipeline... Then, we design four core tasks in our benchmark
Large Language Models β evaluatedby β MiGUE-Bench
confidence 90% Β· MiGUE-Bench, a systematic benchmark for assessing the performance of LLMs in multi-granularity event analysis
MiGUE-Pipeline β usesmodel β GPT-4o
confidence 85% Β· A document will be retained only if it satisfies all the aforementioned criteria based on LLM-as-a-Judge with GPT-4o
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Event analysis is an essential and fundamental direction of information extraction, involving various event-centric tasks at different granularity of documents. While large language models (LLMs) have preliminarily achieved promising performance in part of these tasks individually, their capability in event analysis still lacks comprehensive understanding due to restricted document granularity, task designs, and data source of existing benchmarks. To address these limitations, we introduce MiGUE-Bench, a systematic benchmark for assessing the performance of LLMs in multi-granularity event analysis. To support large-scale evaluation, we first develop an LLM-driven self-correcting annotation framework called MiGUE-Pipeline, enabling scalable acquisition of high-quality source data of events with automatic labels. Then, we design four core tasks in our benchmark, i.e., event detection, relation reasoning, structure induction, and future prediction, to probe model competence at different levels, from atomic event details to complex cross-document narratives. Extensive experiments on state-of-the-art LLMs and retrieval-augmented generation (RAG) methods delineate the current capability boundary and identify critical deficiencies, providing insights into the future improvement of LLMs in challenging event analysis tasks.
Tags
Links
- Source: https://arxiv.org/abs/2607.27654v1
- Canonical: https://arxiv.org/abs/2607.27654v1
Trouble viewing inline? Open PDF directly β
Full Text
52,994 characters extracted from source content.
Expand or collapse full text
From Single- to Cross-Document: Benchmarking Multi-Granularity Event Analysis of Large Language Models Tao Wen Laboratory of Intelligent Collaborative Computing University of Electronic Science and Technology of China Chengdu, China enril.wentao@std.uestc.edu.cn Shuai Shao Laboratory of Intelligent Collaborative Computing University of Electronic Science and Technology of China Chengdu, China 202221080331@std.uestc.edu.cn Pei Ke β Laboratory of Intelligent Collaborative Computing University of Electronic Science and Technology of China Chengdu, China kepei@uestc.edu.cn Xu Han Department of Computer Science and Technology Tsinghua University Beijing, China han-xu@mail.tsinghua.edu.cn Jie Zou School of Computer Science and Engineering University of Electronic Science and Technology of China Chengdu, China jie.zou@uestc.edu.cn Guannan Li Laboratory of Intelligent Collaborative Computing University of Electronic Science and Technology of China Chengdu, China lgn4sci@std.uestc.edu.cn Tao Tian Jinjie Qiu Laboratory of Intelligent Collaborative Computing University of Electronic Science and Technology of China Chengdu, China taotian@std.uestc.edu.cn 202522900127@std.uestc.edu.cn Lan Wang School of Information and Software Engineering University of Electronic Science and Technology of China Chengdu, China 202521090303@std.uestc.edu.cn Ke Qin Laboratory of Intelligent Collaborative Computing University of Electronic Science and Technology of China Chengdu, China qinke@uestc.edu.cn Abstract Event analysis is an essential and fundamental direction of informa- tion extraction, involving various event-centric tasks at different granularity of documents. While large language models (LLMs) have preliminarily achieved promising performance in part of these tasks individually, their capability in event analysis still lacks com- prehensive understanding due to restricted document granularity, task designs, and data source of existing benchmarks. To address these limitations, we introduce MiGUE-Bench, a systematic bench- mark for assessing the performance of LLMs in multi-granularity event analysis. To support large-scale evaluation, we first develop an LLM-driven self-correcting annotation framework called MiGUE- Pipeline, enabling scalable acquisition of high-quality source data of events with automatic labels. Then, we design four core tasks in our benchmark, i.e., event detection, relation reasoning, structure induction, and future prediction, to probe model competence at dif- ferent levels, from atomic event details to complex cross-document β Corresponding author. This work is licensed under a Creative Commons Attribution 4.0 International License. SIGIR β26, Melbourne, VIC, Australia. Β© 2026 Copyright held by the owner/author(s). ACM ISBN 979-8-4007-2599-9/2026/07 https://doi.org/10.1145/3805712.3808607 narratives. Extensive experiments on state-of-the-art LLMs and retrieval-augmented generation (RAG) methods delineate the cur- rent capability boundary and identify critical deficiencies, providing insights into the future improvement of LLMs in challenging event analysis tasks. CCS Concepts β’ Computing methodologiesβ Information extraction. Keywords Large Language Model, Event Analysis, Automatic Evaluation ACM Reference Format: Tao Wen, Shuai Shao, Pei Ke, Xu Han, Jie Zou, Guannan Li, Tao Tian, Jinjie Qiu, Lan Wang, and Ke Qin. 2026. From Single- to Cross-Document: Benchmarking Multi-Granularity Event Analysis of Large Language Models. In Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR β26), July 20β24, 2026, Melbourne, VIC, Australia. ACM, New York, NY, USA, 9 pages. https://doi.org/10.1145/ 3805712.3808607 1 Introduction Event analysis is a foundational pillar of information extraction, inherently requiring to extract the structure of events from raw tex- tual data at different granularity levels [6]. This essential direction arXiv:2607.27654v1 [cs.CL] 30 Jul 2026 SIGIR β26, July 20β24, 2026, Melbourne, VIC, Australia.Tao Wen et al. encompasses a hierarchical task taxonomy from the fine-grained event detection of individual occurrences to the induction of global event structures across different documents [12,13,24,26,29,32, 39,46]. Mastering the capabilities of analyzing events from intra- document event mentions to inter-document event relations is indispensable for the application of current information extraction systems in real-world scenarios, moving beyond local semantic matching towards global relation reasoning. Early works on event analysis have extensively explored task- specific deep learning paradigms, ranging from sequence labeling architectures [8,35,40] to graph-based reasoning frameworks [24, 36]. Despite their advanced performance on specific datasets, the generalization ability of these models is rather limited. Thus, recent works resort to large language models (LLMs) and preliminarily show their promising performance in part of event analysis tasks, such as event extraction [23], event relation prediction [5,19], and event reasoning [34,41]. Equipped with strong abilities of context understanding and knowledge utilization, LLMs have shown great potential in dealing with complex event analysis tasks in a zero-shot manner, gradually becoming a research focus in this field. However, we argue that there still lacks a comprehensive bench- mark for assessing LLMsβ capabilities in event analysis. Existing benchmarks mostly suffer from restricted granularity, lacking data source, and homogeneous task design, hindering a systematic un- derstanding of LLMsβ deficiencies at different event-centric tasks: β’Restricted Granularity: Most of the current benchmarks only fall into the assessment at constrained granularity of docu- ments. While some of the datasets aim to measure the model performance in understanding events within a document [41, 44], others focus on the event relations across multiple documents [3,18,47], both of which fail to provide an en- tire perspective of LLMsβ capabilities at different granularity levels. β’Lacking Data Source: Most existing benchmarks are retrofit- ted from legacy datasets [14], which restrict the scalability to diverse and contemporary corpora. The lack of data sources also causes missing cross-document dependencies, making it improper to measure LLMsβ capabilities to analyze events across multiple documents. β’ Homogeneous Task Design: Existing benchmarks are mostly targeting at isolated evaluation dimensions and objects, ana- lyzing the ability of LLMs to deal with specific event relations [47] or tasks [5,19,34,41]. Such homogeneous and narrow task design may be unable to reveal the capability bound- ary of general LLMs, exaggerating their performance in the limited task scope. To address these limitations, we introduce MiGUE-Bench, a comprehensive benchmark designed for Multi-GranUlar Event analysis. MiGUE-Bench aims to cover the evolving process of event analysis from single-document event detection and relation rea- soning to cross-document event structure induction and future prediction (as shown in Table 1), enabling a holistic assessment of LLMsβ capabilities in different granularity levels of documents. Firstly, to deal with the limitation of document granularity and data source, we develop an LLM-driven automatic pipeline named MiGUE-Pipeline that includes document filtering, event / relation Table 1: Comparison of different event analysis benchmarks. Benchmark Single-Doc Analysis Cross-Doc Analysis Multiple Relations Event Structure ACE05-EN [11]ββ Causal-TimeBank [33]ββββ MAVEN [45]ββ MAVEN-ERE [44]ββββ TCELongBench [47]βββ EventRelBench [14]ββββ MiGUE-Bench (Ours)β annotation, and document clustering, enabling scalable data con- struction from raw corpora to support various event-centric tasks. Secondly, to solve the challenge of homogeneous task design, we devise four task types to cover different stages of event analysis, Specifically, MiGUE-Detection assesses the performance in recog- nizing and extracting triggers from event mentions, while MiGUE- Reasoning aims to evaluate the capability to reason within and across fragmented document contexts to extract implicit event rela- tions. To further obtain a global understanding of event dynamics across different documents, MiGUE-Induction is targeted at evaluat- ing the ability to acquire the global topological structure of events. Finally, MiGUE-Prediction tests the capability to infer the future de- velopment of events based on the understanding of event dynamics across different documents. Our benchmark is expected to reveal the deficiencies of LLMs via comprehensive and challenging task design. In summary, our key contributions are three-fold 1 : β’ We develop the first open-source automatic pipeline named MiGUE-Pipeline for multi-granularity event data genera- tion, alleviating the dependence on manual annotation and flexibly supporting various event-centric tasks. β’We introduce MiGUE-Bench, a comprehensive benchmark that provides systematic assessment for LLMsβ capabilities in the entire spectrum of event analysis. β’ We conduct extensive experiments on state-of-the-art LLMs and retrieval-augmented generation frameworks, uncover- ing critical deficiencies of LLMs on event analysis and pro- viding insights for future improvement. 2 Related Work LLMs for Event Analysis. Recent advances in large language models (LLMs) have enabled a prompt-based paradigm for event- centric understanding and reasoning in a zero-/few-shot setting. Existing studies show that LLMs can support core event analysis capabilities, spanning event extraction [9,20,23,25] and event relation prediction [4,5,19,41]. Beyond event understanding in a single document, LLMs have been utilized for temporal/causal event reasoning across documents, commonly with retrieval modules to support more coherent information extraction [34,41,46]. To summarize, existing studies suggest that LLMs hold strong potential for bridging multiple granularity of event analysis, ranging from single-document understanding to cross-document reasoning. 1 The code of MiGUE-Pipeline is released at https://github.com/EnrilAmirite/ MiGUE-Pipeline, while the MiGUE-Bench dataset is available at https://github.com/ EnrilAmirite/MiGUE-Bench. From Single- to Cross-Document: Benchmarking Multi-Granularity Event Analysis of Large Language ModelsSIGIR β26, July 20β24, 2026, Melbourne, VIC, Australia. Despite this advance, current work on LLM-based event analysis typically focuses on isolated task settings, making it hard to delin- eate the strengths and weaknesses of LLMs along the whole process of event analysis, moving from local semantics to global structures. Therefore, a more systematic benchmark is necessary to compre- hensively reflect the LLMsβ performance at multiple granularity levels of event analysis. Benchmarks for LLMs in Event Analysis. Motivated by the recent progress of LLMs on event-centric tasks, several benchmarks have been proposed to probe LLM capabilities along different facets of event analysis. Existing works typically formulate evaluation as instruction-following tasks with structured outputs, aiming to measure whether LLMs can (i) recover event-centric descriptions from long and temporally dense narratives [41,44], (i) reason about temporal/causal dependencies among events [3,18,47], and (i) infer event relations under diverse types and conduct event reasoning with different formats [14, 47]. Though existing benchmarks offer a preliminary view of LLMsβ event analysis capabilities, they mostly focus on specific event- centric tasks with restricted granularity, data sources, and task design. For comparison, our work aims to provide an complete picture of LLM performance involving different stages in the whole process of event analysis, with multiple granularity levels of docu- ments, diverse data sources, and broad task scopes. 3 MiGUE-Pipeline To support scalable data construction of MiGUE-Bench, we present MiGUE-Pipeline, a four-stage framework comprising document filtering, event annotation, relation annotation, and cluster generation, as shown in Figure 1. Transforming raw corpora into high-quality multi-granularity event data resources can serve as the foundation for subsequent evaluation. 3.1 Document Filtering We firstly employ an LLM-driven filtering mechanism that pre- liminarily evaluates each candidate document across three critical dimensions: factual informativeness (prioritizing objective reporting over subjective commentary or fragmentary discourse), text qual- ity (ensuring the documentsβ grammatical fluency and structural coherence), event density (excluding redundant or non-informative contents). A document will be retained only if it satisfies all the aforementioned criteria based on LLM-as-a-Judge with GPT-4o [27]. The multi-dimensional filtering ensures that the resulting docu- ments are both grammatically fluent and rich in event dynamics. 3.2 Event Annotation Since identifying events in open documents is challenging, we con- duct a pilot study on part of documents with GPT-5.2 [37] followed by manual check, finding six typical error types, i.e., No-Occurrence, Negated-Claim, Assumption, Abstraction, Named-Entity, and Nar- rative. Detailed descriptions of all these error types are provided in Table 2. Based on these findings, we propose a two-stage pro- tocol for event annotation. Firstly, we make an effective LLM (i.e., DeepSeek-V3.2 [10]) generate a candidate set of event triggers. Then, we devise a retrieval-augmented reflection method to select semantically similar examples in the data pool of each error type Table 2: Description of Event Annotation Error Types. Error TypeDescription No-OccurrenceNo-Occurrence errors arise when the model labels the expressions describing future plans, intentions, or predictions rather than events that have actually occurred as triggers. Although such expressions are event-like in form, they are not realized on the current timeline and should not be extracted as events. Negated-ClaimNegated-Claim errors arise when the model fails to distinguish between the statements about events that did not actually occur and genuinely occurring events that express negation (e.g., denying). Negation op- erators themselves (e.g., not, no, never) do not introduce new events, whereas verbs such as deny, reject, and refute denote individual events and should be identified as triggers. AssumptionAssumption errors arise when the model identifies as triggers those verbs or action expressions that carry event semantics but occur only in conditional, hypothetical, intentional, advisory, or exhortative contexts, rather than being asserted as factual events on the real timeline. AbstractionAbstraction errors arise when the model labels the expressions that are event-like in form but function rhetorically or metaphorically rather than denote a specific real-world event as triggers. Such expressions lack clear temporal anchoring, identifiable participants, and concrete boundaries. Thus, they are not extractable factual events. Named-EntityNamed-Entity errors arise when the model incorrectly labels nominal elements of an event (e.g., participants, carriers, results, named entities, event-denoting nouns, or state descriptions) as triggers. NarrativeNarrative errors arise when the model labels the expressions that narrate, explain, modify, or evaluate an event rather than denote the core action as event triggers. Such expressions (e.g., stance and manner) do not introduce new events but only supplement or interpret existing ones. Thereby, they should not be annotated as events. from our pilot study, respectively. This method requires the LLM to self-correct the results by reasoning against similar historical failures. Manual validation on a subset of 200 data samples shows that our automatic protocol yields a precision of 0.87 during event annotation. 3.3 Relation Annotation To further acquire the relation between the events within a single document or among different ones, we devise different strategies to ensure the quality of automatic relation annotation. Single-Document Relation Annotation. Inspired by existing works [5], we design a constraint-aware multi-step reasoning strat- egy to capture event dependencies within a document. For each event pair, LLMs iteratively select the most confident relation type. Upon selecting a candidate label at each iteration, LLMs are prompted to validate its decision against a set of explicit logical constraints (e.g., a coreference label inherently precludes causal or sub-event relations). Based on self-evaluation, the LLM should select one of the following operations, i.e., confirming the current label to proceed to the next most confident relation type, revising the current label to select an alternative one, and restarting the entire decision process due to the conflict during reasoning. Once a label is confirmed, we dynamically prune the search space by removing logically incompatible relation types. To further enhance reliability, we employ majority voting with three LLMs (i.e., GPT-4o, DeepSeek-V3.2 [10], and Qwen3-Max [38]). If all the models yield identical predictions, the result is accepted; otherwise, the instance is escalated to a high-capacity LLM (e.g., Claude-4.5-Opus) for final results. This approach achieves an accuracy rate of 0.82 on 100 data samples during manual check, which shows its effectiveness. SIGIR β26, July 20β24, 2026, Melbourne, VIC, Australia.Tao Wen et al. MiGUE-Pipeline Relation Annotation LLM1LLM2LLM3 Coreference Cause Cause ν 1 ν 2 ν 3 Document Filtering Raw Corpus Filtered Documents Cluster Generation ... Documents with Coreferential Relations Document Affinity Graph Generated Document Cluster Has Coreference Event Pair(s) MiGUE-Benchmark Event Annotation Candidate Triggers announced suspend safeguard hit ... Retrieval Failure Types No-OccurrenceAbstractionNamed-Entity AssumptionNarrativeNegated-Claim Iterative Self-correction Criteria Factual Informativeness Text Quality Event Density LLM Satisfy all the three criteria? suspend Corrected Triggers announced hit ... safeguard ... Pilot Study Constraint-Aware Multi-step Reasoning (CAMR) Single-Document Event Pair 1. Type Iteration 2. Constrain-Aware Self-Validation ConfirmRestartRevise 3. Majority-Consensus Propagate-via-Coreference (PvC) Cross-Document CASR Constrained search space Coreference: β MiGUE-Detection Task Description ν·νν ννν‘ Options A. announced D. hit B. suspend C. safeguard According to foreign media reports, Ukraineβs energy minister <announced> that the country will <suspend> electricity exports starting on the 11th after Russian missile strikes <hit> its energy facilities, in order to <safeguard> the stability of its domestic energy system. Event Mention (Sentence ν) Please select all the event triggers you believe are correct ... Question (Multiple-Choice) MiGUE-Reasoning CoreferenceSubeventCausalTemporal Options A. after D. containsB. before C. is contained ? Cross-Document ? Single-Document Task Description ν·νν ννν Relational Dimension ν Please select the relationship you believe is correct from the options below based on the given event mentions ... Question (Single-Choice) Document(s) ν·,(ν· β² ) & Event Pairs ν,νβ² MiGUE-Induction Interval-based Temporal OrderingCausal Graph Construction Event Trigger Set β° Time ννννν ν ν ν‘ννν‘ ν ννν ...ν ν ν‘ννν‘ ν ννν ν ν ν ν ν ν ν β Generative Result ν ννν , ν ννν , ν ννν , ν ννν , ν ννν ν ν ν‘ννν‘ , ν ν ν‘ννν‘ , ν ν ν‘ννν‘ , ν ν ν‘ννν‘ , ν ν ν‘ννν‘ Options A. ν, ν, ν, ν, β B. ν, ν, ν, ν, β Correct and complete causal chains Distractors: Irrelevant Links Incomplete Path Incorrect Order Task Description ν·νν ννννν Task Description ν·νν ννν’ Document Set ν Please return only the event start time sequence you derive after reasoning ... Question (Generative) Please return only the event end time sequence you derive after reasoning ... ... select all causal chains that you believe are complete and correct ... Question (Multiple-Choice) Event Trigger Set β° With Causal Relationships C. ν, ν, ν, ν D. ν, ν, ν, β E. ν, ν, ν, ν, ν, β MiGUE-Prediction LLM Adversarial Strategies Counterfactual Over-generalization Temporal TrapIrrelevant Distraction Disturbing Options Extracted A. Temporary closure measures will be implemented for road traffic on the Crimean Bridge. Task Description ν·νν νννν Document Sequence ν with timestamps ν― Time Generation Options B. The Crimean Bridge will be permanently closed due to security risks. C. The Crimean Bridge will quickly resume full traffic. D. Russia and Ukraine will reach a new consensus on the Black Sea Grain Initiative. Based on the information provided in the documents, which of the following future predictions is most likely to be accurate? Question (Single-Choice) Random Sampling Figure 1: Overview of MiGUE-Pipeline and MiGUE-Bench. Cross-Document Relation Annotation. For comparison, cross- document relation annotation faces the challenges of relational spar- sity and computational intractability. To address these challenges, we propose a propagate-via-coreference strategy. For preparation, we constrain the search space using three proximity heuristics, including temporal proximity (where documents must fall within a one-month timestamp window), semantic affinity (where event mentions should exceed a semantic similarity threshold), and entity overlap (where events should share common participants). Poten- tially coreferential pairs are then validated using the same protocol as in the single-document part. In the propagate-via-coreference strategy, we densify the relation network by treating coreferential events as bridges for propagation. By applying formal transitiv- ity rules (e.g., ifν 1 ννννν ======= ν 2 ,ν 1 ννν’ν ν ββ ν 3 , thenν 2 ννν’ν ν ββ ν 3 ), we recover implicit dependencies across different documents. This sys- tematic approach circumvents exhaustive pairing while improving the global structural integrity of the event relation network. 3.4 Cluster Generation Since documents that share identical real-world events are poten- tially linked by their underlying narrative contents, we utilize event coreference as a primary signal for document cluster generation, supporting the data construction of downstream tasks that require sufficient cross-document relations. Specifically, we construct a document affinity graph whose vertices represent individual docu- ments and edges denote the presence of coreferential event pairs. To partition this graph into dense subgraphs, we employ the Leiden community detection algorithm [43], which isolates clusters with maximum narrative coherence. By operating on these localized subgraphs whose relational density is highest, we streamline the discovery of closely-connected documents for more efficient data construction of cross-document event-centric tasks. 4 MiGUE-Benchmark In this section, we systematically delineate the construction of MiGUE-Bench, including four core tasks. For each task , we provide a rigorous definition involving task inputs and outputs, followed by its data construction strategy with MiGUE-Pipeline. 4.1 MiGUE-Detection Since the ability of precisely detecting events is the foundational skill of event analysis, we devise the MiGUE-Detection task to mea- sure LLMsβ performance on distinguishing factual event triggers. Task Definition. Given a task descriptionν·νν ννν‘ , a sentenceν, and a candidate set of event triggersC= ν 1 ,ν 2 , ...,ν ν | 3β©½ νβ©½ 6 whereνindicates the size ofC, the model is required to provide all the valid triggers that appear in ν within the candidate setC. Data Construction. The data instances for this task are directly acquired from the event annotation stage (Section 3.2). To further improve the difficulty of this fundamental task, we incorporate ad- versarial distractors into the options. Specifically, we inject typical error types identified during event annotation of MiGUE-Pipeline into candidate options for each instance. By integrating these fail- ures modes as distractors, we aim to assess the fine-grained seman- tic understanding ability of LLMs. From Single- to Cross-Document: Benchmarking Multi-Granularity Event Analysis of Large Language ModelsSIGIR β26, July 20β24, 2026, Melbourne, VIC, Australia. Table 3: Statistics of MiGUE-Bench including the number of data instances (#Instance) and the average number of input tokens (#Token) / options (#Option). TaskSubtask#Instance #Token #Option Detection-33055.184.56 Reasoning Intra Temporal300796.844.00 Causal300937.314.00 Coreference300890.002.00 Subevent300770.583.00 Cross Temporal3001209.374.00 Causal3001202.184.00 Coreference3001202.242.00 Subevent3001141.263.00 Induction Temporal Order150671.77- Causal Graph1101431.274.48 Prediction-3002112.034.50 Overall-3,2901061.243.85 4.2 MiGUE-Reasoning Inferring relationships between events within / across documents is essential for event analysis. MiGUE-Reasoning focuses on evalu- ating LLMsβ capacity of reasoning across different events. Task Definition. Given the task descriptionν·νν ννν , the event pair (ν,ν β² )which comes from the documentν·/ν· β² , respectively, the tar- get relational dimensionR β νννννννν,νΆνν’ν νν,νν’ννν£ννν‘,νΆν β ννννννννν, and a candidate set of relation labelsC=ν 1 ,ν 2 , ...,ν ν | 2β©½ νβ©½4, the model is required to select the a label fromCthat precisely characterizes the relationship within the corresponding dimension. According to whetherν·andν· β² are the same docu- ment, this task can be further categorized into two subtasks, i.e., Reasoning-Intra and Reasoning-Cross. Data Construction. To ensure that our benchmark probes deep reasoning over events rather than surface-level pattern matching, we explicitly exclude pairs linked by overt linguistic cues, such as direct temporal markers (afterward) or causal connectives (because). We retain only those relations that necessitate contextual under- standing and multi-hop reasoning. We also prioritize event pairs that the effective LLMs (such as GPT-4o) initially misclassify to further enhance the difficulty. 4.3 MiGUE-Induction Global comprehension of event development necessitates induc- ing the topological structure of event across documents. MiGUE- Induction aims to measure the capacities of LLMs for structure induction through two sub-tasks: Interval-based Temporal Ordering and Causal Graph Construction. 4.3.1 Interval-based Temporal Ordering. Traditional temporal or- dering often simplifies events into instantaneous points, relying on binary before/after comparisons [47]. This task design fails to capture the temporal duration and interval logic (e.g., overlap and containment) inherently in real-world narratives. To address this limitation, we decouple each event into its start and end temporal anchors for fine-grained ordering. Task Definition. Given the task descriptionν·νν ννννν , a set of event triggersE= ν 1 , . . .,ν ν | ν β₯3, and a set of documents D= ν· 1 , . . .,ν· ν | ν β₯2where these events occur, the model should output two distinct permutations ofE, representing the chronological order of event starts and event ends, respectively. Data Construction. The data instances are derived from cross- document temporal relations (Section 3.3) and document clusters (Section 3.4) generated by MiGUE-Pipeline. We specifically leverage the results of relation propagation, which yields dense temporal networks across different documents. we prioritize clusters with high relation density, ensuring that the selected events form an interconnected network rather than isolated chains. Furthermore, to prevent the models from relying on superficial pattern matching, we remove absolute temporal markers (e.g., April 14, 2022) in the data instances. This forces the model to induct the relative temporal order from the narrative contents and implicit cues within the provided documents. 4.3.2 Causal Graph Construction. Causality among events of dif- ferent documents commonly manifests as a directed acyclic graph (DAG), where events may have multiple preconditions or cause several subsequent consequences. Since causality is essential for structure induction of events, we design a task to assess the capa- bility of constructing causal graphs. Task Definition. Given the task descriptionν·νν ννν’ , a set of event triggersE=ν 1 , ...,ν ν | ν β₯3, a set of documentsD=ν· 1 , ...,ν· ν | ν β₯2where these events occur, and a set of candidate options C ν ν = ν 1 ,ν 2 , . . .,ν ν | 4β©½ νβ©½6, where each option corresponds to a causal pathβ¨ν 1 β ν 2 β Β· β ν ν β©, the model is required to identify all the longest valid causal chains within the causal DAG. Data Construction. The data instances are curated from causal networks that are built based on causal relations (Section 3.3) and cluster generation (Section 3.4). To acquire challenging evaluation instances, we treat each maximal path in causal DAGs as a correct option and design three types of distractors as disturbing options, including irrelevant links (i.e., chains containing event pairs with no logical dependency), incomplete path (i.e., sub-paths that are causally valid but fail to include all possible intermediate or terminal events) , and incorrect order (i.e., chains where events are causally related but presented in an incorrect logical order). 4.4 MiGUE-Prediction The purpose of event analysis is not merely the retrospective un- derstanding of what has occurred, but also the prediction of what will follow. Thus, we devise the MiGUE-Prediction task as follows. Task Definition. Given the the task descriptionν·νν νννν , a tempo- rally ordered document sequenceD= ν· 1 , ...,ν· νβ1 |3β©½ νβ©½6 with their timestamps, and a candidate set of eventsC=ν 1 , . . .,ν ν | 4β©½ νβ©½5, the model is required to select the event that occurs in the correct terminal document ν· ν fromC. Data Construction. Inspired by existing works [21,24,30], we extract the document sequences from the clusters (Section 3.4) that exhibit cross-document causal chains (Section 3.3). We further in- crease difficulty by appending redundant contextβchronologically consistent but irrelevant documents from the same clusterβto dis- courage shallow pattern matching. For option design, we devise the correct option as a paraphrased version of the event occurring in ν· ν with specific details removed to prevent trivial lexical shortcuts. SIGIR β26, July 20β24, 2026, Melbourne, VIC, Australia.Tao Wen et al. Table 4: Accuracy (Acc.) and Micro-F1 of different LLMs and RAG methods on MiGUE-Benchmark. GranularitySingle-DocumentCross-Document TaskDetectionReasoningInductionPrediction Subtask- Reasoning-IntraReasoning-CrossTemporal Order Causal Graph- Temp. Cau. Coref. Sub.Temp. Cau. Coref. Sub.StartEnd MetricMicro-F1Acc. Acc. Acc. Acc.Acc. Acc. Acc. Acc.Acc.Acc.Acc.Acc. Closed-Source LLMs GPT-5.2-Pro0.83980.66000.6373 0.97350.80330.67860.6408 0.96360.81880.8412 0.83330.43750.4231 Gemini-3-Pro0.71510.68670.62960.95000.78010.79670.55710.91000.80340.66230.37750.21590.5814 Claude-4.5-Opus 0.85520.7100 0.67780.9700 0.82890.7781 0.70480.93330.84210.87420.80790.58140.5544 Claude-4.5-Haiku0.61010.67670.60000.93000.70000.46330.42380.74000.71640.44370.17220.19320.4452 Qwen3-Max0.65430.69670.60370.90500.80220.70670.60000.77000.80990.44370.19870.27270.4585 Open-Source LLMs DeepSeek-V3.20.81680.66000.66300.96250.79560.67320.62860.8567 0.84800.82110.63000.28410.5210 GLM-4.70.81330.69330.61670.96330.77660.70660.53330.86000.83000.80670.56000.27270.4867 Kimi-K20.58940.59000.56300.92000.70440.61670.51430.84000.78070.35760.19870.26140.5150 Qwen3-8B0.61380.40000.45990.86770.51330.49000.33900.73330.61700.13910.07950.13640.3355 Qwen3-30B 0.62980.41000.49260.62110.60220.53720.50950.62000.78950.17220.05300.25000.3522 Qwen3-235B0.61150.56000.54810.87250.61330.55670.52860.76670.83620.35100.12580.17050.4352 Llama-3.1-7B0.49590.23260.21850.32960.22380.21670.16670.18000.13160.05960.01320.03410.3654 RAG Methods based on LLMs UltraRAG (w/ Gemini-3-Pro)0.74570.69390.61480.95250.82630.80000.53330.89000.83040.69540.38410.20450.6246 UltraRAG (w/ Qwen3-Max) 0.69360.70000.60740.91000.79110.70670.61900.79000.79530.47680.21850.29550.4751 UltraRAG (w/ Qwen3-8B)0.62870.36670.43330.86900.51780.50470.33010.69000.54090.14570.07280.11360.3522 LightRAG (w/ Gemini-3-Pro)0.71370.71000.61810.95500.73280.75000.55710.84670.78950.70330.41960.20670.6185 LightRAG (w/ Qwen3-Max) 0.64650.71810.61320.89220.74970.69670.62380.79330.79240.48340.23180.29550.4950 LightRAG (w/ Qwen3-8B)0.58510.38930.44810.86470.51450.48670.40120.64670.53800.11260.03970.12770.3445 Following existing works [16], we generate disturbing options us- ing four adversarial strategies, including counterfactual (i.e., events that are logically opposite to the true outcome), overgeneralization (i.e., plausible but exaggerated conclusions that transcend the evi- dence), temporal trap (i.e., past events from the document sequence), and irrelevant distraction (i.e., fabricated but contextually consistent events unsupported by the document sequence). 4.5 Quality Control To ensure the quality of MiGUE-Bench, we implement a multi-stage validation protocol. Firstly, we automatically filter all the instances to maintain a balanced distribution across document sources, event densities, and difficulties. Then, we thoroughly detect and exclude offensive, biased, or harmful content, ensuring alignment with ethi- cal standards. Finally, we conduct manual review on all the instances to verify the correctness of answers and eliminate ambiguity in all the options. The statistics of the final dataset are shown in Table 3. 5 Experiment 5.1 Setting Source Data. As MiGUE-Pipeline is a scalable event data construc- tion pipeline, our corpus is mainly drawn from two sources: (1) high-quality open-source news document datasets [31], and (2) more than 9,000 news reports collected from Chinese official and mainstream media outlets between 2022 and 2024. During data collection, we performed a preliminary filtering based on news tags to ensure that the corpus covers major events from this period as comprehensively as possible. All the corpora are subsequently processed by our MiGUE-Pipeline to automatically construct event data 2 . Base Model Selection. To comprehensively evaluate the per- formance of LLMs, we select two lines of work, i.e., LLMs and RAG methods. For LLMs, we involve mainstream closed-source (including GPT-5.2-Pro [37], Gemini-3-Pro [15], Claude-4.5-Opus [2], Claude-4.5-Haiku [1], and Qwen3-Max [38]) and open-source LLMs (including DeepSeek-V3.2 [10], GLM-4.7 [42], Kimi-K2 [22], Qwen-3-8B/30B/235B [38], and Llama-3.1-7B [28]) as representa- tives, covering the models with different scales, families, and capac- ities. Furthermore, we also adopt two mainstream RAG frameworks including LightRAG [17] and UltraRAG [7]. The retrieval corpora contain 5,000 documents from the document filtering stage (Section 3.1) of MiGUE-Pipeline 3 . In our experiment, we utilize Gemini-3- Pro, Qwen3-Max, and Qwen3-8B as base models for RAG methods. Evaluation Metric. For MiGUE-Detection, we adopt Micro-F1 to jointly reflect the precision and recall of this multi-choice selection task. As for the other tasks, accuracy is adopted for measurement. 5.2 Main Result The results in Table 4 show that closed-source LLMs mostly out- perform open-source ones, especially on cross-document event analysis tasks, demonstrating the effectiveness of state-of-the-art proprietary models. Among them, GPT-5.2-Pro and Claude-4.5- Opus achieve the best performance across nearly all the tasks, while 2 Since the source data collected is initially Chinese, we use MiGUE-Pipeline to con- struct the Chinese dataset. Nevertheless, MiGUE-Pipeline is language-agnostic and can be also applied to other languages with corresponding prompts. 3 In the task of MiGUE-Prediction, we restrict the documents to those occurring prior to the predicted event, preventing data leakage. From Single- to Cross-Document: Benchmarking Multi-Granularity Event Analysis of Large Language ModelsSIGIR β26, July 20β24, 2026, Melbourne, VIC, Australia. 0346891012 Number of Retrieved Documents (n) 0.2 0.3 0.4 0.5 0.6 Score (Accuracy) Qwen3-Max (Start) Qwen3-Max (End) Kimi-k2 (Start) Kimi-k2 (End) Figure 2: Accuracy on MiGUE-Induction (Temporal Order) with different numbers of retrieved documents. Deepseek-V3.2 and GLM-4.7 rank highest among open-source LLMs. We also have other interesting findings: RAG can generally improve performance across the four tasks, but the gains are unstable. Specifically, RAG benefits strong LLMs such as Gemini-3-Pro on most of the tasks. However, for weaker LLMs like Qwen-3-8B, the noisy retrieved documents may degrade the original performance in turn. How to design spe- cific RAG strategies for event analysis still needs further study. At the MiGUE-Reasoning task, temporal and subevent rela- tions generally benefit from cross-document settings, whereas coreference and causal reasoning exhibit the opposite trend. We conjecture that multi-document contexts may introduce addi- tional temporal signals (e.g., explicit timestamps), which can be easily captured by LLMs. Consequently, temporal-related relations benefit from these additional signals. In contrast, for coreference and causal reasoning, different documents may describe the same event with distinct perspectives. Also, temporally successive events are not necessarily causally related. Such heterogeneity may in- troduces noise and ambiguity, leading to degraded performance in cross-document settings. At the MiGUE-Induction task, LLMs are sensitive to start times of events but exhibit limited understanding of their time spans. We observe from the subtask of interval-based tem- poral ordering that most LLMs perform significantly worse on end-time ordering than on start-time one, with the accuracy often dropping by nearly half. This suggests that identifying relative event start times is much simpler for current LLMs, whereas reasoning about event durations remains challenging. Causal tasks remain challenging for current LLMs. Most of these LLMs achieve unsatisfactory performance on the tasks of causal reasoning and causal graph induction. We conjecture that one reason is the misalignment between model-internal causal rep- resentations and human causal understanding. Moreover, current causal definitions in event analysis are often confined to language- based descriptions and lack rigorous mathematical formalization. We leave the exploration of more fine-grained probabilistic causal modeling at the schema level instead of merely textual logical for- mulations in event analysis as important future work. 5.3 Detailed Analysis Analysis on the Number of Retrieved Documents. To further analyze this effect of additional relevant documents on temporal reasoning, we evaluate Qwen3-Max and Kimi-K2 under UltraRAG 0-300300-900900-15001500-21002100-3000 Instruction Length 0.0 0.5 1.0 Accuracy Qwen3-8BQwen3-30BQwen3-235B Figure 3: Accuracy on MiGUE-Reasoning (Subevent) with different lengths of input instructions. with different numbers of retrieved documents. Figure 2 shows that both models exhibit a riseβfallβstabilization pattern for start- ordering and end-ordering accuracies as the number of retrieved documents increases. Moderate retrieved documents improve per- formance, whereas excessive documents degrade accuracy, possibly due to noise accumulation and context dilution. End-ordering per- formance consistently remains lower than start-ordering, revealing the deficiencies of LLMs in modeling event time spans. Analysis on the Length of Input Instructions. We further ana- lyze the scaling behavior of LLMs with different lengths of input instructions on subevent reasoning. As shown in Figure 3, Qwen3- 235B exhibits steady performance gains as the instruction length increases, indicating strong capabilities of long-context understand- ing. For comparison, Qwen3-30B/8B also benefits from longer in- puts but reaches its peak in the 1500β2100 / 900β1500 range, fol- lowed by a slight decline. This suggest that larger models benefit from richer contextual evidence, whereas smaller models saturate earlier and degrade under extended inputs. 6 Conclusion In this work, we present a comprehensive benchmark called MiGUE- Bench for multi-granularity event analysis, which covers the tasks ranging from fine-grained event detection and reasoning to cross- document event structure induction and forecasting, together with an LLM-driven automatic pipeline named MiGUE-Pipeline for scal- able event data construction. Through extensive experiments on state-of-the-art LLMs and RAG methods, we show that despite their strong general capabilities, these models still exhibit notable limi- tations in multi-granularity event analysis. We hope that MiGUE- Bench and MiGUE-Pipeline can serve as standardized testbeds for diagnosing the capability boundaries of LLMs in event analysis and inspire future advances in this research field. Acknowledgments This work was supported by Noncommunicable Chronic Diseases- National Science and Technology Major Project (No. 2023ZD0501806), Sichuan Science and Technology Program (No. 2025ZNSFSC1488) , Fundamental Research Funds for the Central Universities (No. ZYGX2025XJ041), and CIPS-SMP-Zhipu Large Model Fund (No. CIPS-SMP20250314). References [1] Anthropic. 2025. Introducing Claude Haiku 4.5. https://w.anthropic.com/ news/claude-haiku-4-5. Accessed: 2026-04-25. SIGIR β26, July 20β24, 2026, Melbourne, VIC, Australia.Tao Wen et al. [2]Anthropic. 2025. Introducing Claude Opus 4.5. https://w.anthropic.com/ news/claude-opus-4-5. Accessed: 2026-04-25. [3]Michael Bugert, Nils Reimers, and Iryna Gurevych. 2021. Generalizing Cross- Document Event Coreference Resolution Across Multiple Corpora. Computa- tional Linguistics 47, 3 (2021), 575β614. [4]Chunkit Chan, Cheng Jiayang, Weiqi Wang, Yuxin Jiang, Tianqing Fang, Xin Liu, and Yangqiu Song. 2024. Exploring the Potential of ChatGPT on Sentence Level Relations: A Focus on Temporal, Causal, and Discourse Relations. In Findings of the Association for Computational Linguistics: EACL 2024. 684β721. [5]Meiqi Chen, Yubo Ma, Kaitao Song, Yixin Cao, Yan Zhang, and Dongsheng Li. 2024. Improving large language models in event relation logical prediction. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 9451β9478. [6]Muhao Chen, Hongming Zhang, Qiang Ning, Manling Li, Heng Ji, Kathleen McKeown, and Dan Roth. 2021. Event-centric natural language processing. In Proceedings of the 59th Annual Meeting of the Association for Computational Lin- guistics and the 11th International Joint Conference on Natural Language Processing: Tutorial Abstracts. 6β14. [7]Yuxuan Chen, Dewen Guo, Sen Mei, Xinze Li, Hao Chen, Yishan Li, Yixuan Wang, Chaoyue Tang, Ruobing Wang, Dingjun Wu, Yukun Yan, Zhenghao Liu, Shi Yu, Zhiyuan Liu, and Maosong Sun. 2025. UltraRAG: A Modular and Automated Toolkit for Adaptive Retrieval-Augmented Generation. CoRR abs/2504.08761 (2025). [8]Yubo Chen, Liheng Xu, Kang Liu, Daojian Zeng, and Jun Zhao. 2015. Event extraction via dynamic multi-pooling convolutional neural networks. In Proceed- ings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 167β176. [9]Milind Choudhary and Xinya Du. 2024. QAEVENT: Event Extraction as Question- Answer Pairs Generation. In Findings of the Association for Computational Lin- guistics: EACL 2024. 1860β1873. [10] DeepSeek-AI. 2025. DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models. CoRR abs/2512.02556 (2025). [11]George R. Doddington, Alexis Mitchell, Mark A. Przybocki, Lance A. Ramshaw, Stephanie M. Strassel, and Ralph M. Weischedel. 2004. The Automatic Content Extraction (ACE) Program - Tasks, Data, and Evaluation. In Proceedings of the Fourth International Conference on Language Resources and Evaluation. 837β840. [12]Chuang Fan, Daoxing Liu, Libo Qin, Yue Zhang, and Ruifeng Xu. 2022. Towards Event-level Causal Relation Identification. In SIGIR β22: The 45th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1828β1833. [13]John Foley, Michael Bendersky, and Vanja Josifovski. 2015. Learning to Extract Local Events from the Web. In Proceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval. 423β432. [14]Jie Gong, Biaoshuai Zheng, and Qiwang Hu. 2025. EventRelBench: A Comprehen- sive Benchmark for Evaluating Event Relation Understanding in Large Language Models. In Findings of the Association for Computational Linguistics: EMNLP 2025. 9084β9099. [15] Google. 2025. Gemini 3: A new era of intelligence. https://blog.google/products- and-platforms/products/gemini/gemini-3/. Accessed: 2026-04-25. [16] Yong Guan, Hao Peng, Xiaozhi Wang, Lei Hou, and Juanzi Li. 2024. OpenEP: Open-Ended Future Event Prediction. CoRR abs/2408.06578 (2024). [17]Zirui Guo, Lianghao Xia, Yanhua Yu, Tu Ao, and Chao Huang. 2025. LightRAG: Simple and Fast Retrieval-Augmented Generation. In Findings of the Association for Computational Linguistics: EMNLP 2025. 10746β10761. [18] Yu Hong, Tongtao Zhang, Tim OβGorman, Sharone Horowit-Hendler, Heng Ji, and Martha Palmer. 2016. Building a Cross-document Event-Event Relation Corpus. In Proceedings of the 10th Linguistic Annotation Workshop held in conjunction with ACL 2016. 1β6. [19]Zhilei Hu, Zixuan Li, Xiaolong Jin, Long Bai, Jiafeng Guo, and Xueqi Cheng. 2025. Large Language Model-Based Event Relation Extraction with Rationales. In Proceedings of the 31st International Conference on Computational Linguistics. 7484β7496. [20]Kuan-Hao Huang, I-Hung Hsu, Tanmay Parekh, Zhiyu Xie, Zixuan Zhang, Prem Natarajan, Kai-Wei Chang, Nanyun Peng, and Heng Ji. 2024. TextEE: Benchmark, Reevaluation, Reflections, and Future Challenges in Event Extraction. In Findings of the Association for Computational Linguistics: ACL 2024. 12804β12825. [21]Zijie Huang, Jeehyun Hwang, Junkai Zhang, Jinwoo Baik, Weitong Zhang, Do- minik Wodarz, Yizhou Sun, Quanquan Gu, and Wei Wang. 2024. Causal graph ode: Continuous treatment effect modeling in multi-agent dynamical systems. In Proceedings of the ACM Web Conference 2024. 4607β4617. [22]Kimi Team. 2025. Kimi K2: Open Agentic Intelligence. CoRR abs/2507.20534 (2025). [23]Bobo Li, Xudong Han, Jiang Liu, Yuzhe Ding, Liqiang Jing, Zhaoqi Zhang, Jinheng Li, Xinya Du, Fei Li, Meishan Zhang, Min Zhang, Aixin Sun, Philip S. Yu, and Hao Fei. 2025. Event Extraction in Large Language Model. CoRR abs/2512.19537 (2025). [24]Manling Li, Sha Li, Zhenhailong Wang, Lifu Huang, Kyunghyun Cho, Heng Ji, Jiawei Han, and Clare Voss. 2021. The Future is not One-dimensional: Complex Event Schema Induction by Graph Modeling for Event Prediction. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. 5203β5215. [25]Rui Li, Wenlin Zhao, Cheng Yang, and Sen Su. 2022. A Dual-Expert Framework for Event Argument Extraction. In The 45th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1110β1121. [26] Jinzhi Liao, Xiang Zhao, Xinyi Li, Lingling Zhang, and Jiuyang Tang. 2021. Learn- ing Discriminative Neural Representations for Event Detection. In The 44th International ACM SIGIR Conference on Research and Development in Information Retrieval. 644β653. [27] Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. 2023. G-Eval: NLG Evaluation using Gpt-4 with Better Human Alignment. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2511β2522. [28] Llama Team. 2024. The Llama 3 Herd of Models. CoRR abs/2407.21783 (2024). [29] Chenwei Lou, Jun Gao, Changlong Yu, Wei Wang, Huan Zhao, Weiwei Tu, and Ruifeng Xu. 2022. Translation-Based Implicit Annotation Projection for Zero-Shot Cross-Lingual Event Argument Extraction. In The 45th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2076β2081. [30] Yunshan Ma, Chenchen Ye, Zijian Wu, Xiang Wang, Yixin Cao, and Tat-Seng Chua. 2023. Context-aware event forecasting via graph disentanglement. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 1643β1652. [31] Yunshan Ma, Chenchen Ye, Zijian Wu, Xiang Wang, Yixin Cao, Liang Pang, and Tat-Seng Chua. 2023. Structured, Complex and Time-complete Temporal Event Forecasting. CoRR abs/2312.01052 (2023). [32]Anne-Lyse Minard, Manuela Speranza, Eneko Agirre, Itziar Aldabe, Marieke Van Erp, Bernardo Magnini, German Rigau, and Ruben Urizar. 2015. Semeval- 2015 task 4: Timeline: Cross-document event ordering. In proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015). 778β786. [33]Paramita Mirza and Sara Tonelli. 2014. An Analysis of Causality between Events and its Relation to Temporal Information. In Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers. 2097β 2106. [34] Nishanth Nakshatri, Siyi Liu, Sihao Chen, Dan Roth, Dan Goldwasser, and Daniel Hopkins. 2023. Using LLM for improving key event discovery: Temporal-guided news stream clustering with event summaries. In Findings of the Association for Computational Linguistics: EMNLP 2023. 4162β4173. [35] Thien Huu Nguyen, Kyunghyun Cho, and Ralph Grishman. 2016. Joint event extraction via recurrent neural networks. In Proceedings of the 2016 conference of the North American chapter of the association for computational linguistics: human language technologies. 300β309. [36]Thien Huu Nguyen and Ralph Grishman. 2018. Graph Convolutional Networks With Argument-Aware Pooling for Event Detection. In Proceedings of the Thirty- Second AAAI Conference on Artificial Intelligence. 5900β5907. [37] OpenAI. 2026. OpenAI GPT-5 System Card. CoRR abs/2601.03267 (2026). [38] Qwen Team. 2025. Qwen3 Technical Report. CoRR abs/2505.09388 (2025). [39] Rashmi Sankepally. 2019. Event Information Retrieval from Text. In Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval (Paris, France) (SIGIRβ19). 1447. [40]Lei Sha, Feng Qian, Baobao Chang, and Zhifang Sui. 2018. Jointly Extracting Event Triggers and Arguments by Dependency-Bridge RNN and Tensor-Based Argument Interaction. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Applications of Artificial Intelligence (IAAI-18), and the 8th AAAI Symposium on Educational Advances in Artificial Intelligence (EAAI-18), New Orleans, Louisiana, USA, February 2-7, 2018. 5916β5923. [41]Zhengwei Tao, Zhi Jin, Yifan Zhang, Xiancai Chen, Haiyan Zhao, Jia Li, Bin Liang, Chongyang Tao, Qun Liu, and Kam-Fai Wong. 2025. A Comprehensive Evaluation on Event Reasoning of Large Language Models. In The Thirty-Ninth AAAI Conference on Artificial Intelligence. 25273β25281. [42]Team GLM. 2024. ChatGLM: A Family of Large Language Models from GLM-130B to GLM-4 All Tools. CoRR abs/2406.12793 (2024). [43]Vincent A Traag, Ludo Waltman, and Nees Jan Van Eck. 2019. From Louvain to Leiden: guaranteeing well-connected communities. Scientific reports 9, 1 (2019), 1β12. [44] Xiaozhi Wang, Yulin Chen, Ning Ding, Hao Peng, Zimu Wang, Yankai Lin, Xu Han, Lei Hou, Juanzi Li, Zhiyuan Liu, Peng Li, and Jie Zhou. 2022. MAVEN-ERE: A Unified Large-scale Dataset for Event Coreference, Temporal, Causal, and Subevent Relation Extraction. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. 926β941. [45]Xiaozhi Wang, Ziqi Wang, Xu Han, Wangyi Jiang, Rong Han, Zhiyuan Liu, Juanzi Li, Peng Li, Yankai Lin, and Jie Zhou. 2020. MAVEN: A Massive General Domain Event Detection Dataset. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). 1652β1671. From Single- to Cross-Document: Benchmarking Multi-Granularity Event Analysis of Large Language ModelsSIGIR β26, July 20β24, 2026, Melbourne, VIC, Australia. [46]Zairun Yang, Yilin Wang, Zhengyan Shi, Yuan Yao, Lei Liang, Keyan Ding, Emine Yilmaz, Huajun Chen, and Qiang Zhang. 2025. EventRAG: Enhancing LLM Generation with Event Knowledge Graphs. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 16967β16979. [47]Zhihan Zhang, Yixin Cao, Chenchen Ye, Yunshan Ma, Lizi Liao, and Tat-Seng Chua. 2024. Analyzing Temporal Complex Events with Large Language Models? A Benchmark towards Temporal, Long Context Understanding. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 1588β1606.