Event Instruction Tuning

• Event Instruction Tuning is a set of methods that explicitly models events, their structures, and relations to improve model performance.
• It refines tuning by integrating event-specific supervision, including triggers, schemas, and temporal decision rules to clarify context.
• Applications span event reasoning, extraction, detection, and multimodal analysis, enhancing both textual and motion-based tasks.

Event instruction tuning is a family of instruction-tuning approaches in which the supervision explicitly models events, event structure, event relations, schema descriptions, or event-aware temporal decision rules, rather than relying only on generic instruction-following corpora. In recent work, this paradigm appears in event reasoning through event quadruples, in event detection through trigger-centered causal supervision, in event extraction through schema text and annotation guidelines, in event relation extraction through relation-conditioned multiple-answer generation, and in multimodal settings through schema-guided prompting, streaming response-state prediction, and native-motion supervision (Tao et al., 2024, Bethany et al., 2024, Srivastava et al., 22 Feb 2025, Xu et al., 6 Feb 2025, Yuan et al., 2 Dec 2025, Xia et al., 24 Dec 2025, Li et al., 2024).

1. Problem setting and motivation

A common premise across the literature is that generic instruction tuning does not adequately expose models to the structural properties of events. "EVIT: Event-Oriented Instruction Tuning for Event Reasoning" argues that small instruction-tuned models such as Alpaca-7B, Vicuna-7B, WizardLM-7B, and Dolly-v2-7B underperform on event reasoning because their instruction-tuning corpora do not explicitly model events, event structure, or relations between events. "Enhancing Event Reasoning in LLMs through Instruction Fine-Tuning with Semantic Causal Graphs" makes a parallel claim for event detection: standard prompting and generic instruction tuning do not explicitly teach the causal relationship between event triggers and event types. "Instruction-Tuning LLMs for Event Extraction with Annotation Guidelines" frames event extraction as a setting in which fine-grained schema understanding is central, so the instruction itself must carry more than a task verb; it must include textual descriptions of event types and argument roles (Tao et al., 2024, Bethany et al., 2024, Srivastava et al., 22 Feb 2025).

This literature therefore treats event instruction tuning less as a single architecture than as a design principle: event tasks are reformulated so that the model must condition on explicit event units, relation labels, triggers, roles, or temporal response states. A recurrent implication is that gains come from exposing the model to the internal organization of events rather than only to task-output pairs.

2. Representational units and instruction formats

The main methods differ in surface form, but they converge on the idea that instruction tuning should operate over explicit supervisory units that make event structure visible.

Work	Supervisory unit	Representative task
EvIT	event quadruple $(e^h, r, e^t, c)$	event reasoning
SCG Instructions	Semantic Causal Graph causal subgraph	event detection
Guideline-based EE	schema text plus Python class/object output	event extraction
MAQInstruct	event-relation instruction plus multiple answers	event relation extraction
SSGPF	Event Type Schema Guided Prompting and Argument Role Schema Guided Prompting	multimedia event extraction
Streamo	<Silence>, <Standby>, <Response> with temporally grounded dialogue	streaming video understanding
LLaMo	native motion with text-guided frame selection	human action understanding

EvIT defines the event quadruple as $(\mathcal{Q}) = (e^h, r, e^t, c)$, where $e^h$ is the head event, $e^t$ is the tail event, $r$ is the relation, and $c$ is the background context paragraph. SCG-based event detection defines a Semantic Causal Graph $G=(V,E)$ whose node set is partitioned into context nodes, event trigger nodes, and event type nodes, with directed edges for context-to-trigger, trigger-to-trigger, and trigger-to-type relations. Guideline-based event extraction formulates the prompt for an autoregressive LLM as $P=[I \oplus E_e \oplus X]$, and represents schemas as Python classes and extracted instances as Python objects. MAQInstruct retains autoregressive generation but adds a bipartite matching term, $$\mathcal{L}= \mathcal{L}_{CE} + \lambda \mathcal{L}_{BPM}$$, to reduce answer-order sensitivity in set-valued relation extraction (Tao et al., 2024, Bethany et al., 2024, Srivastava et al., 22 Feb 2025, Xu et al., 6 Feb 2025).

This suggests that event instruction tuning is characterized not by a uniform loss or backbone, but by a shift toward structurally constrained instruction spaces: relation-aware generation, schema-conditioned decoding, or temporally aligned interactive supervision.

3. Event-oriented instruction tuning for event reasoning

EvIT is the clearest event-centric formulation in the textual reasoning setting. Its core learning problem is event-relation learning: the model predicts the tail event from the head event, relation, and context,

$e^t = \mathrm{M}(e^h, c, r),$

and also learns a discriminative multiple-choice version,

$(\mathcal{Q}) = (e^h, r, e^t, c)$0

where $(\mathcal{Q}) = (e^h, r, e^t, c)$1 contains the gold tail event and negative candidate events. The training instances are rendered as standard instruction-input-response triples, with templates generated by ChatGPT for every combination of 6 relations, 2 learning forms, and 2 context conditions, yielding $(\mathcal{Q}) = (e^h, r, e^t, c)$2 template types, and 100 prompts per type. The event quadruples are mined heuristically from BookCorpus using discourse connectives from the Penn Discourse TreeBank, an end-to-end relation parser inspired by ASER, a relation set $(\mathcal{Q}) = (e^h, r, e^t, c)$3, two negative candidates per positive example, event-length filters of 2 to 10 words, and context-length filters of 10 to 50 words. The final training set contains 212,538 event quadruple instances. Llama-7B is fine-tuned on 4 NVIDIA A100 GPUs for 3 epochs with DeepSpeed Zero-2, fp16, AdamW, gradient checkpointing, maximum sequence length 512, and batch size 32, with a reported runtime of about 3 hours. On eight event reasoning tasks from E-CARE, MCTACO, SocialIQA, and SCT, EvIT is reported as the best among the 7B models on nearly all settings: average close-task score 75.69, 4.34 points above Alpaca; held-out close-task average 71.46; and average open-task BERTScore 29.63, 7.51 above Alpaca. Human evaluation rates EvIT best overall, especially on format, though the model slightly underperforms Alpaca in content on intentional reasoning, which the paper attributes to intention being a held-out relation (Tao et al., 2024).

The significance of EvIT is methodological as much as empirical. It shows that event reasoning can be made instruction-following-compatible without discarding structured supervision: generation and discrimination are both encapsulated as instructions, and the contextual background $(\mathcal{Q}) = (e^h, r, e^t, c)$4 is treated as part of the event unit rather than as incidental text.

4. Event extraction, event detection, and event relation extraction

For event extraction, "Instruction-Tuning LLMs for Event Extraction with Annotation Guidelines" studies whether annotation guidelines function as effective instructional signals. Event extraction is framed as structured generation under schema constraints, with trigger extraction and argument extraction both decoded in code form. The experiments use ACE05 and RichERE, with LLaMA-3.1-8B-Instruct as the primary model and rsLoRA via Unsloth. The paper compares human-written Guideline-H with machine-generated variants Guideline-P, Guideline-PN, Guideline-PS, Guideline-PN-Int, and Guideline-PS-Int. In full-data training without negative sampling, guidelines consistently help: on ACE, Guideline-P is best, with about +10 TC and +5 AC over NoGuideline; on RichERE, Guideline-PN is strongest, with about +5 TC and +2 AC. Machine-generated guidelines can outperform human-written ones by up to 11% in trigger classification and 7% in argument classification. The gains are especially notable for cross-schema generalization, low-resource settings such as 2,000 training samples, and rare event types. The paper also reports that five diverse guidelines are better than one integrated guideline, and that Guideline-PN is the most stable overall (Srivastava et al., 22 Feb 2025).

For event detection, the SCG-based approach explicitly decomposes prediction into trigger identification followed by type inference. The central factorization is

$(\mathcal{Q}) = (e^h, r, e^t, c)$5

contrasted with standard instruction tuning that directly learns $(\mathcal{Q}) = (e^h, r, e^t, c)$6. Appendix A reports 20 instruction variations, each asking the model to identify event triggers and classify each trigger into its event type. LoRA is used with rank 256, alpha 256, 6 epochs, batch size 16, AdamW in 8-bit precision, learning rate $(\mathcal{Q}) = (e^h, r, e^t, c)$7, cosine scheduler, and warmup ratio 0.1. Across MAVEN, FewEvent, CASIE, M$(\mathcal{Q}) = (e^h, r, e^t, c)$8E$(\mathcal{Q}) = (e^h, r, e^t, c)$9, and MLEE in the TextEE benchmark, SCG Instructions outperform standard instruction fine-tuning by an average of 35.69% on Event Trigger Classification. Mistral-LoRA-SCG Instruct reaches average EC/TI/TC of 90.08/65.39/61.02, compared with GPT-4 six-shot RAG at 77.37/49.89/44.41. The paper reports only a minimal average drop of 2.03 points across six general benchmarks, and a context-mutation ablation causes an average TC drop of only 3.58 points, which the authors interpret as evidence that the model is learning the trigger-event relationship rather than overfitting to surface context (Bethany et al., 2024).

For event relation extraction, MAQInstruct reformulates the task from event-event prompting to event-relation prompting. Instead of asking, for every pair of event mentions, which relation holds, it asks: given one event mention and one event relation instruction, list all matching event mentions. This reduces the number of samples from $e^h$0 to $e^h$1, where $e^h$2 is the number of event mentions and $e^h$3 the number of relation types. The framework uses marked mentions, a dependency parsing chain, and a bipartite matching loss with $e^h$4 so that the order of multiple answers does not dominate training. On MAVEN-ERE, HiEve, MATRES, and MECI, MAQInstruct improves over InstructERE across ChatGLM3, Qwen, and Llama2 backbones. For Llama2, the reported gains over InstructERE are +4.9% on coreference, +4.1% on temporal, +4.2% on causal, and +3.9% on sub-event. The efficiency comparison on MAVEN-ERE coreference reports 631,486 queries and 813 minutes for InstructERE versus 17,780 queries and 25 minutes for MAQInstruct. Ablations show that removing markers, removing the dependency parsing chain, or removing the bipartite matching loss all degrade performance (Xu et al., 6 Feb 2025).

5. Multimedia, streaming, and motion-centered variants

Event instruction tuning has expanded beyond text into multimedia event extraction, streaming video, and human motion. In multimedia event extraction, SSGPF uses LLaVA-v1.5-7B as backbone and decomposes the task into ETSGP for event detection and ARSGP for argument extraction. ETSGP prompts the model with candidate event types and definitions and asks it to output event type and trigger. ARSGP retrieves the role schema $e^h$5 for each predicted event type and runs one prompt per role. For visual arguments, the framework uses text-bridged grounding: the MLLM generates a textual region description, and SEEM converts that description into a bounding box. Because M2E2 does not provide a multimodal training set, the paper constructs a weakly aligned instruction-tuning dataset of 2,316 image-text pairs from ACE2005 and SWiG, matched by a CLIP-based crossmodal news retriever further trained on news image-caption pairs. LoRA is applied with rank 128, alpha 64, 15 epochs, batch size 96, and learning rate $e^h$6. On M2E2, SSGPF reports MED $e^h$7 and MEAE $e^h$8, improving by 5.8 F1 on MED and 8.4 F1 on MEAE over prior baselines. The paper also states that zero-shot SSGPF performs poorly, so instruction tuning is essential (Yuan et al., 2 Dec 2025).

In streaming video, Streamo extends event instruction tuning to continuous, temporally open-ended interaction. Streamo-Instruct-465K contains 465K samples, including 400K manually/re-annotated valid samples, drawn from 135,875 videos across Koala, LLaVA-Video, ActivityNet, QVHighlight, YouCook2, HACS, EgoTimeQA, DiDeMo, and COIN. The supervision covers real-time narration, action understanding, event captioning, temporal event grounding, and time-sensitive question answering, all normalized into a dialogue where the model predicts <Silence>, <Standby>, or <Response>. The special state tokens are trained with a focal + frequency-balanced loss because the empirical ratio is roughly Silence : Standby : Response = 12 : 3 : 2. Streamo-7B is reported to beat Dispider by +13.83% average on OVO-Bench, to gain an extra +4.66% when trained at 1 fps and tested at 2 fps, and to improve over ET-Instruct-164K by +7.1% on forward active responding and +11.79% overall. The paper further states that Streamo-7B improves its Qwen2.5-VL-7B base by about +3.4% average on offline benchmarks and that Streamo-Bench contains 300 videos and 3,000 instructions/tasks (Xia et al., 24 Dec 2025).

Human Motion Instruction Tuning, implemented in LLaMo, argues that event-like human actions should not be collapsed into language tokens before tuning. Motion is kept in native continuous form, and the model uses a Cross Talker module for text-guided frame selection, reducing sequence complexity from $e^h$9 to $e^t$0. The training mixture includes MoVid, Swing with 20,000 annotated golf swing videos and expert QA instructions, HumanML3D, KIT-ML, and Mo-RepCount. On MoVid-Bench-Motion, LLaMo reports 55.32 accuracy / 3.67 score, compared with MotionLLM at 49.50 / 3.49 and MotionGPT at 36.86 / 3.11. On BABEL-QA, it reports 0.458 overall. On the Swing dataset, it reports 24.80 accuracy / 2.48 score, compared with MotionLLM at 16.53 / 1.57 and MotionGPT at 14.35 / 1.40. The paper explicitly notes that it does not use the exact phrase “event instruction tuning” as a formal method title, but it presents native-motion instruction tuning as relevant whenever the event is an embodied action unfolding in time (Li et al., 2024).

6. Cross-cutting methodological issues

One cross-cutting issue is that instruction data quality depends not only on task content but also on instruction format. "Exploring Format Consistency for Instruction Tuning" is not event-specific, but it is directly relevant to event instruction tuning datasets assembled from heterogeneous sources. Its Unified Instruction Tuning framework converts source instructions into a shared target format using GPT-3.5, considers Ni-v2-style unified formats such as DP, DPE, DPN, and DPNE, and introduces a perplexity-based denoising method that samples $e^t$1 candidates with temperature 1.0, scores them with GPT-J, and keeps the lowest-perplexity transfer. On T5-LM-xl, UIT reports average improvements on unseen-instruction generalization of about 9.3% EM and 7.6% Rouge-L, and it shows that high task diversity plus unified formatting is better than adding more tasks without format unification. The paper also reports that an offline GPT-J model trained on about 3,000 parallel examples performs much better than heuristic rules and only slightly worse than GPT-3.5 overall (Liang et al., 2023).

A second issue concerns what instruction tuning changes internally. "From Language Modeling to Instruction Following: Understanding the Behavior Shift in LLMs after Instruction Tuning" reports three effects: instruction tuning helps models recognize the instruction parts of prompts and condition responses on them, encourages self-attention heads to capture more word-word relationships about instruction verbs, and encourages feed-forward networks to rotate pre-trained knowledge toward user-oriented tasks. The paper supports this with gradient-based prompt-response attribution, self-attention head interpretation, and PCA-based feed-forward analysis, comparing pre-trained LLaMA and Mistral to Vicuna and Mistral-Instruct. In the context of event instruction tuning, these findings help explain why schema text, trigger-first prompting, relation-aware templates, and temporally explicit response states can have large effects: the tuned model is being encouraged to parse and condition on the event instruction itself rather than only on the surrounding text (Wu et al., 2023).

Several misconceptions are therefore rejected by current evidence. More tasks are not necessarily enough without format consistency. Human-written guidelines are not always superior to machine-generated ones. One-shot generation is not always the best formulation for structured event outputs. Zero-shot multimodal prompting may be inadequate for multimedia event extraction. Generic instruction following is not equivalent to event competence. These claims are repeatedly supported, though in task-specific ways, by the event reasoning, extraction, detection, and multimedia studies cited above (Srivastava et al., 22 Feb 2025, Bethany et al., 2024, Yuan et al., 2 Dec 2025).

The broader picture is that event instruction tuning is becoming a unifying strategy for tasks where the supervision must preserve event identity, event structure, or event timing. The literature does not yet offer a single canonical formulation. Instead, it offers a set of convergent design choices: represent events explicitly; make triggers, roles, or relations visible in the prompt; decompose structured outputs when sequence order is artificial; and align the instruction format with the event semantics of the task.