Event Referential Reasoning (EventRR)

• Event Referential Reasoning is a framework that models events as structured semantic units defined by participants, time, context, and inter-event relations.
• It employs graph-based representations and contrastive objectives to capture causal, temporal, and hierarchical dependencies in both textual and video domains.
• Benchmarks like ESTER and EV highlight its ability to enhance precision in event inference and referential grounding tasks.

Searching arXiv for Event Referential Reasoning and closely related event reasoning papers. Event Referential Reasoning (EventRR) denotes a family of problems in which a system must identify, ground, select, generate, or segment the event licensed by context, discourse structure, or explicit inter-event relations. Across the literature, the common premise is that an event is not merely a token sequence: it is a structured semantic unit whose interpretation depends on participants, time, place, contextual grounding, and relations such as causality, temporality, hierarchy, conditionality, counterfactuality, contrast, or co-reference. In one direct formulation, EventRR is the name of a referring video object segmentation framework that parses a referring expression into a Referential Event Graph and reasons over that graph by topological traversal (Xu et al., 10 Aug 2025).

1. Conceptual foundations

A central premise of EventRR research is that an event must be modeled as a semantically integrated object. ClarET defines an event as a semantic unit, “usually a text span composed of a predicate and its arguments,” and frames the core problem as context-to-event reasoning: given a paragraph containing multiple events and discourse connectives, infer a missing event from the remaining context and its correlations to other events. This formulation explicitly emphasizes event-level correlations rather than token-level co-occurrence, and it treats causal, temporal, and contrastive markers as important evidence for event inference (Zhou et al., 2022).

A related formalization appears in EvIT, which represents event reasoning with an event quadruple

$q=(h,r,c,t),$

where $h$ is the head event, $t$ is the tail event, $r$ is the relation between the two events, and $c$ is the context. The paper argues that this structure is “complete” for event reasoning because it jointly captures event semantics and syntax, relational knowledge, and contextual grounding. Its learning targets are

$t=M(h,c,r)$

for generation and

$t=M(h,c,r \mid C)$

for candidate-based discrimination (Tao et al., 2024).

The ontological coreference literature makes the identity conditions even more explicit. One account defines an event as “a happening, that is, an entity directly related to time and (indirectly) related to space via its participants,” and rephrases this as: “An Event identifies a spatio-temporal region, some relevant entities in it, and what happens to these entities.” On that basis, event identity is characterized by four necessary features: event type, participants with thematic roles, time of happening, and global location. This ontology-driven view treats event referential reasoning as identity reasoning over structured event descriptions rather than lexical similarity between triggers (Borgo et al., 2016).

These formulations converge on a shared view: EventRR is fundamentally about mapping from partial or indirect evidence to an event-level representation whose meaning is constrained by relational structure. This suggests that “reference” in EventRR is broader than classical mention resolution. It includes event completion, event selection, relation-sensitive answer retrieval, and grounding of an event description to a visual or textual referent.

2. Task formulations and benchmarks

The most direct text benchmark for EventRR-style reasoning is ESTER, a machine reading comprehension dataset for Event Semantic Relation Reasoning. ESTER reframes event relation extraction as question answering with natural-language queries, covering five relation types: Causal, Conditional, Counterfactual, Sub-event, and Co-reference. It provides more than 6K questions, captures 10.1K event relation pairs, and uses passages from TempEval3 news articles containing 3–4 consecutive sentences with at least 7 event triggers; the average passage length is about 128 tokens. The final split contains 1492 training passages with 4547 questions, 108 development passages with 301 questions, and 287 test passages with 1170 questions. Evaluation uses token-based $F_1^T$, event-based HIT@1, and Exact Match; the best system reported in the benchmark paper, UnifiedQA-large, achieves $22.1$ EM, $63.3$ $h$0, and $h$1 HIT@1, compared with a human baseline of $h$2 EM, $h$3 $h$4, and $h$5 HIT@1 (Han et al., 2021).

A complementary perspective is provided by EV, a benchmark for evaluating event reasoning at both schema level and instance level. EV defines an event schema as a graph $h$6, where $h$7 is the set of event types and $h$8 is the set of relations between events, and an instance-level graph $h$9, where each instance event has an event type but also detailed arguments and context. The benchmark introduces two paradigms at both levels: Contextual Event Classification (CEC), which asks the model to choose an event under a target relation, and Contextualized Relation Reasoning (CRR), which asks the model to determine the relation between two queried events. EV covers six relation types—$t$0, $t$1, $t$2, $t$3, $t$4, and $t$5—and contains four subsets of size 492, 558, 767, and 835 for S-CEC, I-CEC, S-CRR, and I-CRR, with average graph size 3.62 nodes and 2.78 edges (Tao et al., 2024).

Other benchmarks instantiate EventRR through generation or selection rather than QA. ClarET is evaluated on 9 benchmarks: 5 generation tasks and 4 classification tasks spanning abductive, counterfactual, and ending reasoning, with relation types such as causal, temporal, and contrast. The tasks most closely aligned with EventRR are APSI, MCNC, and $t$6NLG, because they require recovering or selecting an event from contextual evidence. EvIT evaluates on 8 tasks derived from ECARE, MCTACO, SocialIQA, and SCT, in both Open and Close forms, treating ECARE and MCTACO as held-in tasks and SocialIQA and SCT as held-out tasks. This variety indicates that EventRR is not tied to a single benchmark style; it appears as generation, discrimination, multiple-choice inference, free-form answer generation, and structured grounding (Zhou et al., 2022, Tao et al., 2024).

3. Neural textual models

A major line of work approaches EventRR through event-aware pre-training. EventBERT continues pre-training from RoBERTa-large on an unsupervised corpus of BookCorpus paragraphs filtered to contain multiple correlated events. Its basic training unit is a tuple $t$7, where $t$8 is a paragraph, $t$9 an extracted event span, and $r$0 a discourse relation in the paragraph but outside the event. The model learns with three contrastive objectives: Correlation-based Event Ranking, Contradiction Event Tagging, and Discourse Relation Ranking, combined as

$r$1

EventBERT is not presented as an explicit EventRR model, but it is closely related because it trains representations that judge whether an event fits its paragraph, whether an event span is contradictory, and whether a discourse relation is plausible (Zhou et al., 2021).

ClarET makes the event-recovery aspect more explicit. Its backbone objective is whole event recovering, which masks exactly one event in a paragraph and learns the conditional generation

$r$2

To strengthen event-correlation learning and reduce the difficulty of reconstructing a long event span, ClarET adds Contrastive Event-correlation Encoding and Prompt-based Event Locating, with total pre-training objective

$r$3

The paper is explicit that ClarET is a correlation-aware context-to-event transformer for general event-centric reasoning and should be viewed as a broader precursor framework for EventRR rather than a direct instance of event reference resolution (Zhou et al., 2022).

Instruction tuning has produced a second family of methods. EvIT fine-tunes Llama-7B on 212,538 event quadruples mined from BookCorpus and converted into instruction-following data. Its training combines generation and multiple-choice discrimination over the quadruple $r$4, with candidate sets built from the gold tail event plus two negative events. The reported training setup uses 4 $r$5 NVIDIA A100 GPUs, 3 epochs, DeepSpeed, ZeRO-2, fp16 mixed precision, AdamW, max sequence length 512, batch size 32, gradient checkpointing, and about 3 hours of training. On Close tasks, EvIT reaches an overall average of 75.69, which is 4.34 points higher than Alpaca; on Open tasks, it reports an average BERTScore of 29.63, 7.51 points higher than Alpaca. The paper interprets these gains as evidence that explicit event structure, relation-aware training, candidate discrimination, context grounding, and instruction-style prompting improve event reasoning and reduce hallucination (Tao et al., 2024).

TranCLR targets EventRR most directly in narrative QA. In ESTER, the relevant event set is

$r$6

where the question contains a single event trigger $r$7 and the answer depends on passage events linked to it by one of পাঁচ semantic relations. TranCLR uses an invertible transformation matrix $r$8 to project question events, answer events, and other events into a common event embedding space,

$r$9

and then applies contrastive learning plus relation-type classification with total loss

$c$0

On ESTER, the generative setting improves UnifiedQA-large from $c$1 to $c$2 in $c$3, with EM increasing from $c$4 to $c$5; in the extractive setting, TranCLR reaches $c$6 $c$7, $c$8 HIT@1, and $c$9 EM. The paper highlights over $t=M(h,c,r)$0 gain in token-level $t=M(h,c,r)$1 and about $t=M(h,c,r)$2 gain in EM under the multi-answer setting (Lu et al., 2022).

A more mechanistic direction is represented by ROLE and ABLE for zero-shot event-relational reasoning. ROLE locates the key module $t=M(h,c,r)$3 in Flan-T5-large using average indirect effect and edits the last linear layer of the most influential encoder MLP or decoder cross-attention module. ABLE transfers the resulting location and weight-edit patterns across tasks by analogy, with

$t=M(h,c,r)$4

and

$t=M(h,c,r)$5

The paper reports that ROLE improves interpretability and reasoning performance with reduced computational cost, while ABLE achieves state-of-the-art results in zero-shot reasoning across causal classification and sub-event extraction benchmarks (Tang et al., 1 Jan 2025).

4. Symbolic, constrained, and schema-level reasoning

Not all EventRR research is neural or end-to-end. A symbolic line of work derives event coreference from ontological analysis and RDF reasoning. One framework analyzes six event types from ECB—Arresting, Killing, Dying, Charging, Shooting, and Attacking—and encodes certain and possible coreference criteria as SPARQL rules over RDF event graphs extracted by Pikes. For Killing, for example, certain coreference includes $t=M(h,c,r)$6 and $t=M(h,c,r)$7, while possible coreference includes combinations of shared killer, tool, time, and place. Evaluated on ECB/EventCorefBank, with 482 news documents, 1744 annotated event mentions, and 339 distinct events, the method substantially improves BLANC $t=M(h,c,r)$8: from 66.14 for the lemma baseline to 93.29 when combining possible and certain rules. The paper explicitly characterizes this as a precision-oriented, ontology-driven event coreference method, while also noting recall limitations, dependence on extraction quality, and manual ontological analysis effort of about 20 minutes per event type (Borgo et al., 2016).

Another structured approach models EventRR as inference over an event complex rather than isolated event pairs. Joint constrained learning for event-event relation extraction treats temporal relations $t=M(h,c,r)$9 and hierarchy relations $t=M(h,c,r \mid C)$0 jointly, and converts annotation, symmetry, and conjunction constraints into differentiable objectives:

$t=M(h,c,r \mid C)$1

This design allows logical propagation across triples of events and across task families, unifying temporal chains and subevent structure. On MATRES, the model reaches $t=M(h,c,r \mid C)$2 versus $t=M(h,c,r \mid C)$3 for the best baseline; on HiEve, it reports micro-average $t=M(h,c,r \mid C)$4. The paper argues that learning-time logical regularization can replace a more expensive global inference process and better induce coherent event complexes (Wang et al., 2020).

Schema-level evaluation introduces a further distinction between knowing event structures and using them in a human-aligned way. EV reports that LLMs can perform event reasoning but that their performance remains “far from satisfactory,” with CEC easier than CRR and causality easier than temporal or hierarchy reasoning. Averaged over selected models, I-CEC is 58.91 while I-CRR is 46.18. The same study also argues that models have event schema knowledge but are not aligned with humans on how to utilize it, and shows that direct injection of schema knowledge improves instance-level reasoning—for example, GPT-4 rises from 63.80 to 69.89 on CEC and from 61.20 to 63.11 on CRR under direct guidance (Tao et al., 2024).

Taken together, these results define an important axis within EventRR: explicit structure can be represented as ontology, logical constraints, or schema memory, and each choice changes what counts as evidence for an event reference. This suggests that EventRR is as much a representational problem as a prediction problem.

5. EventRR as a direct RVOS framework

In 2025, EventRR became the name of a concrete model for referring video object segmentation. The framework decouples RVOS into object summarization and referent reasoning. Given a video clip

$t=M(h,c,r \mid C)$5

with objects $t=M(h,c,r \mid C)$6 and a referring expression $t=M(h,c,r \mid C)$7, the goal is to predict the referent mask

$t=M(h,c,r \mid C)$8

On the visual side, EventRR uses frame-level summarization and video-level summarization with Sliding Window Attention, producing temporal queries $t=M(h,c,r \mid C)$9. On the language side, it converts an AMR parse into a Referential Event Graph

$F_1^T$0

a single-rooted directed acyclic graph whose nodes are concepts and whose edges are semantic roles. Because raw AMR may have the wrong root and may contain cycles from co-reference, the method re-roots the graph on the referent concept and inverts edges when needed to guarantee a DAG. The model further adds Refer-Aware Positional Encoding based on shortest-path distance to the root (Xu et al., 10 Aug 2025).

The central reasoning module is Temporal Concept-Role Reasoning. TCRR traverses REG in topological order so that children are processed before parents and each node’s score can be interpreted as the referent distribution of the sub-expression rooted at that node. For a parent concept $F_1^T$1 and temporal query $F_1^T$2, the score is

$F_1^T$3

where Object-Concept Align measures direct compatibility between query and concept, and Temporal Referent-Context Align measures whether the child-derived context and the role embedding support the parent-child relation in the video. The paper emphasizes that each reasoning step can be interpreted as a question-answer pair derived from concept-role relations in REG. Supervision combines DETR-style segmentation losses with Pseudo-Referent Reasoning Loss:

$F_1^T$4

Empirically, EventRR reports 66.9/65.0/68.8 $F_1^T$5 on Ref-Youtube-VOS with Video-Swin-B, 65.4/62.7/68.1 on DAVIS17-RVOS, overall IoU 81.4 and mean IoU 73.1 on A2D-Sentences, and oIoU 73.9 and mIoU 72.6 on JHMDB-Sentences (Xu et al., 10 Aug 2025).

An adjacent multimodal formulation is ENTER for VideoQA, which also reasons over event graphs, but uses generated Python code over APIs such as find_node, get_temporal_after, get_temporal_before, get_caused_by, get_resulted_in, get_children, and get_parent. ENTER defines event graphs whose nodes are events and whose edges are temporal, causal, and hierarchical relations, and adds a hierarchical iterative update mechanism with denser graph generation, denser caption generation, and multimodal retrieval. The paper reports 75.1 accuracy on NExT-QA, 71.5 on IntentQA, and 54.4 on EgoSchema, positioning event-graph reasoning as an interpretable alternative to both top-down modular and bottom-up opaque VideoQA systems (Ayyubi et al., 24 Jan 2025).

6. Empirical status, misconceptions, and open issues

A recurrent misconception is that EventRR is equivalent to event coreference. The literature is broader. ESTER includes causal, conditional, counterfactual, sub-event, and co-reference relations; EV spans event classification and relation reasoning at both schema and instance levels; EvIT jointly trains generation and discrimination; ClarET and EventBERT address event recovery and correlation; the RVOS EventRR framework targets object grounding through eventful referring expressions rather than textual coreference alone (Han et al., 2021, Tao et al., 2024, Tao et al., 2024, Xu et al., 10 Aug 2025).

A second misconception is that event reasoning is adequately captured by generic token-level language modeling. ClarET explicitly identifies two failures of whole event recovering alone—implicit event-correlation and learning difficulty—and introduces contrastive and prompt-based objectives to compensate. EventBERT similarly argues that MLM-style pretraining emphasizes local token co-occurrence rather than event-level causal, temporal, and discourse consistency. TranCLR’s gains on ESTER, and EvIT’s gains over instruction-tuned 7B baselines, reinforce the same point: relation-sensitive event representations and candidate discrimination improve referential precision (Zhou et al., 2022, Zhou et al., 2021, Lu et al., 2022, Tao et al., 2024).

The empirical picture also shows persistent gaps. In ESTER, even the benchmark’s strongest system remains well below human performance. EV finds that LLMs possess event schema knowledge but are not aligned with humans on how to use it, and that CRR is substantially harder than CEC. This suggests that many systems can recognize plausible events without reconstructing the relational structure humans rely on (Han et al., 2021, Tao et al., 2024).

Methodological limits remain heterogeneous. The ontology-driven coreference method handles only six event types and requires manual ontological analysis. ROLE and ABLE focus mainly on binary reasoning tasks. The RVOS EventRR model depends on AMR parsing, syntactic preprocessing, and pseudo-referent supervision because concept-level referent annotations are unavailable. ENTER remains sensitive to caption quality, graph completeness, and consistent referencing across event nodes (Borgo et al., 2016, Tang et al., 1 Jan 2025, Xu et al., 10 Aug 2025, Ayyubi et al., 24 Jan 2025).

A plausible implication is that EventRR is best understood not as a single architecture or benchmark, but as a structured inference regime with several recurring commitments: events are first-class units; event references are mediated by typed relations and context; explicit structure—whether graph-based, contrastive, ontological, or instruction-tuned—usually improves performance; and evaluation must distinguish local plausibility from genuine event-level relational reasoning. Under that interpretation, ClarET and EventBERT are strong precursor frameworks, TranCLR and ESTER are direct text-side testbeds, EvIT and ROLE/ABLE show how structure can be injected into LLMs, symbolic and constrained methods demonstrate the value of explicit reasoning, and the named EventRR framework extends the same logic to video grounding (Zhou et al., 2022, Zhou et al., 2021, Lu et al., 2022, Tao et al., 2024, Wang et al., 2020, Xu et al., 10 Aug 2025).