EAGER: Enhancing Generative Event Extraction via Reinforcement Learning with Verifiable Rewards

1 Introduction

• Frontier general-purpose LLMs still fall short of competitive end-to-end event extraction performance across diverse event types and domains.The section identifies coarse supervision as a limitation of prior training strategies.Source: §1 · p. 1
• Standard RL objectives rely on format validity and task-level accuracy, leaving errors such as hallucinated triggers, unsupported arguments, over-generation, and incomplete coverage insufficiently targeted.When completions in a group receive identical prompts, their outputs can be homogeneous, producing near-zero reward variance and uninformative gradients.Source: §1 · p. 2
• EAGER decomposes rewards across validity, extraction accuracy, groundedness, over-generation, coverage, and span precision, and varies negative schemas across completions to increase reward variance.The schema variation is intended to alleviate advantage collapse and enable more informative policy gradients.Source: §1 · p. 2

2 Related Work

• Instruction-based IE, annotation-guided prompting, and large-scale instruction resources aim to improve schema adherence and cross-domain generalization, while code-generation methods use programming-language syntax constraints.The examples include InstructUIE [Wang et al., 2023a], GoLLIE, IEPile, CodeIE [Li et al., 2023b], and KnowCoder [Li et al., 2024].Source: §2 · p. 2

2.1 Event Extraction with Large Language Models

• LLM event-extraction research spans pipeline and joint modeling, automated annotation, and instruction tuning with detailed event and role descriptions to address extraction precision and generalization.Prior work reports improvements for low-frequency and cross-schema event types from detailed descriptions during instruction tuning.Source: §2.1 · p. 2

2.2 Preference and Reinforcement Learning for Structured Extraction

• EventRL [Gao et al., 2024] and multi-objective RL extend reward-based extraction, while ADELIE uses IE-specific preference pairs; these approaches target structured-output reliability.The passage describes ADELIE as combining supervised fine-tuning with Direct Preference Optimization, and reports strong performance across closed, open, and on-demand IE settings.Source: §2.2 · pp. 2–3

3.1 Task Formulation

• Event extraction maps text to schema-constrained representations through trigger identification and classification, plus argument identification and classification.The schema specifies permissible argument roles for the structured event representation.Source: §3.1 · p. 3

3.2 Code-Based Input and Output Representations

• EE is represented as Python code, with each event schema defined as a @dataclass and extracted events represented as class instances.The authors state that Python syntax guarantees well-formed outputs and simplifies parsing.Source: §3.2 · p. 3

3.3 Annotation Guideline Generation

• Each Python class schema is augmented with natural-language guidelines generated through instruction-tuned prompting.The guidelines add semantic detail not present in the compact code-based schemas.Source: §3.3 · p. 3

3.4 Prompt Structure

• Training prompts combine a task instruction, the gold event schema with its guideline, and the input text.Negative schemas for event types absent from the input are also sampled to improve event-type discrimination.Source: §3.4 · p. 3, Figure 2 · p. 3

3.5 Post-Training Framework

• The post-training framework applies DAPO to groups of candidate outputs using verifiable, schema-based rewards.The paper investigates RL with verifiable rewards for end-to-end event extraction.Source: §3.5 · pp. 3–4

3.6 Reward-Decoupled Normalization

• Reward-Decoupled Normalization independently normalizes each reward within the sampled group to balance contributions without manual weighting.The authors state this also alleviates reward hacking by preventing any single component from dominating optimization.Source: §3.6 · p. 4

3.7 SCAE: Schema-Contrastive Advantage Estimation

• SCAE training achieves substantially lower cumulative ACR than the fixed-prompt baseline.It varies negative-schema subsets across completions, changing distractor difficulty and increasing within-group reward variance.Source: §3.7 · pp. 4–5, Figure 3 · p. 5

3.8 Verifiable Rewards Modeling

• Specialized verifiable rewards target complementary behaviors, including output validity, extraction correctness, groundedness, prediction balance, and span precision.The modular design provides more interpretable, fine-grained optimization signals.Source: §3.8 · p. 5

Baselines

• EAGER is compared with few-shot proprietary LLMs, GoLLIE, supervised fine-tuning, guideline-augmented instruction tuning, ADELIEDPO, and EventRL.Reproduced baselines marked with † use the same evaluation protocols and a unified experimental setup.Source: §4 · p. 5

Evaluation Datasets

• Experiments use seven standard end-to-end event-extraction datasets spanning different domains, with the “split 1” data split.The datasets are WikiEvents, PHEE, CASIE, Genia2011, Genia2013, MLEE, and M2E2.Source: §4 · p. 5

Evaluation Metrics

• Evaluation reports F1 for trigger identification and classification and argument identification and classification, including attached argument versions AI+ and AC+.Mean-F1 across six subtask metrics is the main aggregate measure of end-to-end extraction quality.Source: §4 · pp. 5–6

Evaluation Materials

• The six verifiable rewards target structural invalidity, extraction inaccuracy, hallucination, over-generation, under-coverage, and boundary imprecision.The reward set is summarized in Table 1; Table 2 reports test-split Mean-F1 results across the seven datasets, and Table 3 evaluates individual reward contributions.Source: Table 1 · p. 6, Table 2 · p. 6, Table 3 · p. 6
• Table 4 evaluates the impact of augmenting event schemes with annotation guidelines.The caption states the comparison's scope but does not provide result values.Source: Table 4 · p. 6

5.1 Performance Analysis

• Even GPT-5.4-mini’s 11.70 falls below GoLLIE-7B’s 16.35; few-shot GPT-4o, o1, and GPT-5.4 score 9.20, 9.52, and 9.47, respectively.The passage uses these comparisons to underscore the difficulty of end-to-end event extraction and the gap between prompting and task-adapted models.Source: §5.1 · p. 6
• 30.79 average Mean-F1 gives EAGER the highest result, surpassing the strongest prior reinforcement-learning baseline by +5.15 across datasets.The reported gain is consistent across datasets and annotation schemes.Source: §5.1 · p. 7
• Removing annotation guidelines reduces average F1 from 30.79 to 22.61 for EAGER and from 16.35 to 5.81 for GoLLIE-7B.The authors connect this pattern to the descriptive grounding guidelines provide for schema generalization.Source: §5.1 · p. 7

5.2 Ablation Study

• Removing SCAE drops average F1 from 30.79 to 22.54; the REE + Rfmt ablation likewise falls from 27.36 to 23.56.Varying negative schemas across sampled completions increases advantage variance and supplies more informative gradient updates, helping activate fine-grained rewards.Source: §5.2 · p. 7, Table 6 · p. 7
• 7.4 points for argument metrics versus 2.1 for trigger metrics: the full reward improves arguments more, while Rcov and Rspan are the strongest single rewards for triggers and arguments, respectively.Adding any individual specialized reward improves overall performance over REE + Rfmt, and the full combination of six components performs best.Source: §5.2 · p. 8, Table 7 · p. 8

5.3 Error Analysis

• All reduces total errors by 6,639, the second-largest reduction, while missing arguments rise by +956; reward configurations reduce false positives but increase missing arguments and roles.The baseline’s dominant error category is over-generation; span-boundary and parsing errors remain marginal and stable. Error categories are not exhaustive.Source: §5.3 · p. 8, Figure 5 · p. 8

6 Conclusion

• Across seven event extraction benchmarks, EAGER consistently improves over strong baselines, particularly on argument-level extraction quality.The authors emphasize combining fine-grained reward modeling with optimization that preserves reward diversity.Source: §6 · p. 8
• EAGER combines task-aligned verifiable rewards with SCAE to provide more informative optimization signals for structured extraction.The reward decomposes extraction objectives, while schema-contrastive grouping is introduced to mitigate reward variance collapse.Source: §6 · p. 8
• The conclusion highlights fine-grained reward modeling and reward-diversity-preserving optimization as important for generative information extraction.Source: §6 · p. 8

7 Limitations

• Automatically generated annotation guidelines may vary in quality across LLMs, domains, and schemas, and the assessed error categories do not fully model coreference.The authors identify cross-sentence arguments and nested event structures as phenomena coreference modeling may help address.Source: §7 · p. 9
• Evaluation remains limited to English event extraction with predefined schemas, while balancing precision and recall remains challenging.The authors suggest curriculum-based reinforcement learning as a possible way to better balance precision and recall during optimization.Source: §7 · p. 9

A Code-based Representation

• Event schemas are represented as Python classes and extracted events as instances, yielding syntactically well-formed outputs that facilitate parsing.The representation is described as human-readable and as mitigating ambiguities in natural-language instructions.Source: A Code-based Representation · pp. 11–12, Figure 6 · p. 11

B Guidelines Annotation Generation

• Guidelines are generated per event type from 10 positive and 15 negative examples, then incorporated into schema docstrings and comments.The guidelines specify event definitions, argument roles, semantic functions, representative mentions, and edge cases.Source: B Guidelines Annotation Generation · p. 12

B.1 Additional Annotation Guidelines Analysis

• Augmenting event schemas with annotation guidelines improves event extraction performance for the GoLLIE backbone.Source: §B.1 · p. 12

C Supervised Fine-Tuning

• Supervised fine-tuning adapts the model to event extraction using code-style schemas and annotated prompt-response pairs, with a standard autoregressive likelihood objective.The described outputs are intended to be syntactically valid and semantically grounded Python event representations.Source: C Supervised Fine-Tuning · p. 12

D Inference

• At inference, all event schemas are provided jointly so the model can produce instantiated event objects in one forward pass.Source: D Inference · p. 12

E DAPO Background

• DAPO modifies GRPO with asymmetric clipping, token-level loss, dynamic sampling, and removal of the KL penalty.The passage describes these changes as addressing entropy collapse, reward noise, and sequence-level length bias.Source: E DAPO Background · p. 12

F Verifiable Rewards Formulation; F.1 Validity Reward; F.2 Extraction Accuracy Reward

• The reward formulation combines parseability and schema compatibility with extraction accuracy and generation behavior, including grounding, avoiding spurious predictions, and event coverage.Validity is binary; extraction accuracy sums F1 across six metrics, but remains coarse for specific generation issues.Source: F Verifiable Rewards Formulation · pp. 12–13, §F.1 · p. 13, §F.2 · p. 13

F.3 Groundedness Reward; F.4 Over-generation Reward

• Groundedness rewards whether predicted mentions occur verbatim in the input, while an over-generation penalty decreases as predictions exceed gold event and argument counts.The groundedness signal is agnostic to event-type correctness; the over-generation reward is 1 with no excess predictions and saturates at 0.Source: §F.3 · p. 13, §F.4 · p. 13

F.5 Coverage Reward

• The coverage reward measures predicted schema-valid events and arguments relative to gold counts, with clamping so over-prediction cannot inflate the score.Source: §F.5 · p. 13

F.6 Span Precision Reward

• The span precision reward gives partial credit to argument spans that overlap gold spans, addressing exact-match rewards’ treatment of supersets as fully incorrect.It averages scores for aligned predicted arguments and is gated to zero when no overprediction occurs.Source: §F.6 · pp. 13–14

G Implementation Details

• Experiments use GoLLIE-7B with QLoRA adapters and DAPO optimization, selecting the best validation checkpoint and using greedy decoding at inference.The reported setup uses rank 8 adapters, batch size 4 for RL, and up to 10 training epochs.Source: G Implementation Details · p. 14

H Training Protocol

• Training batches concatenate and shuffle the seven datasets for cross-domain and cross-schema batches; the experiments use DAPO with SCAE and the proposed rewards without an SFT warm start.SFT yielded only marginal validation improvements; preliminary tests found G=4 completions outperformed G=2, and K=10 was selected under memory and prompt-length constraints.Source: H Training Protocol · p. 14

I Pseudo Code of DAPO with SCAE

• The paper provides pseudocode outlining the DAPO training process with SCAE.Source: I Pseudo Code of DAPO with SCAE · p. 14

J Training Dynamics Analysis

• Figure 8 tracks mean completion length, mean reward, and generation entropy, alongside event extraction validation performance.The authors describe monitoring intermediate metrics as useful for diagnosing undesired behavior and checking reward components.Source: J Training Dynamics Analysis · p. 15, Figure 8 · p. 15

J.1 Mean Completion Length Analysis

• Mean completion length increases steadily during training, consistent with the model producing more complete event objects and argument slots rather than under-populated outputs.The analysis reports no prolonged stagnation or decline in length and suggests the training signal discourages overly conservative decoding.Source: §J.1 · p. 15

J.2 Mean Reward Analysis

• Mean reward rises monotonically without instability, but its imperfect correlation with held-out extraction performance makes validation-set evaluation important for detecting overfitting.The authors describe Rfull as a consistent training signal for fitting the training-set distribution.Source: §J.2 · p. 15

K Training Computational Cost

• Table 9 reports EAGER’s training cost on four NVIDIA H100 GPUs, with wall time covering the full training data for all seven datasets.Source: Table 9 · p. 15, K Training Computational Cost · p. 15

J.3 Generation Entropy Analysis

• Generation entropy declines slowly, stabilizes, then rises slightly; no erratic fluctuations are observed after removing KL regularization (β=0).The authors note that exploration supports diverse span-selection and role-assignment strategies before convergence on schema-grounded outputs.Source: §J.3 · p. 15

• Seven benchmark datasets span diverse domains, annotation schemes, and event ontologies, providing a testbed for robustness across heterogeneous schemas and extraction difficulty.Source: L Datasets Details · pp. 15–16
• WikiEvents is a large-scale news benchmark with richly annotated real-world event mentions and complex argument structures (Li et al., 2021).Source: L Datasets Details · p. 15
• PHEE targets pharmacovigilance extraction from medical case reports, while CASIE covers cybersecurity incident reports with specialized schemas and technical vocabulary.Source: L Datasets Details · p. 15
• GENIA2011 and GENIA2013 use PubMed abstracts and include nested events with biologically grounded arguments; MLEE extends biomedical extraction to molecular-level interactions.Source: L Datasets Details · p. 16
• M2E2 was designed for joint text-image event understanding, but this evaluation uses only its textual component (Li et al., 2020).Source: L Datasets Details · p. 16

M Additional Results and Analysis

• Tables 12–14 provide detailed event-extraction results for WikiEvents, PHEE, CASIE, GENIA2011, GENIA2013, MLEE, and M2E2.Source: M Additional Results and Analysis · p. 16, Table 12 · p. 19, Table 13 · p. 21

M.1 Trigger and Argument Performance Analysis

• Across most datasets, all rewards yield the strongest TI/TC and highest AI+/AC+ scores, with especially pronounced gains on M2E2 and PHEE.Source: §M.1.1 · p. 16, §M.1.2 · p. 16

M.1.1 Event Trigger Performance (TI,TC)

• On triggers, full rewards improve M2E2 by TI +11.17 and TC +14.29 over baseline; Rovr reduces spurious triggers but can hurt TI when the baseline under-generates.Source: §M.1.1 · p. 16

M.1.2 Argument Performance (AI, AC, AI+, AC+)

• Argument metrics lag trigger metrics, and full rewards produce the highest AI+/AC+ scores on most datasets.On M2E2, full rewards yield AI+ +14.87 and AC+ +17.65 over baseline; the authors also report that anchored metrics expose trigger-argument misalignment hidden by mean-F1.Source: §M.1.2 · p. 16

N Examples Event Schema with Generated Annotation Guidelines

• Figures 14–20 illustrate one event schema augmented with annotation guidelines for each dataset.Source: N Examples Event Schema with Generated… · p. 17

M.2 Additional Error Analysis

• Figures 9–13 compare error categories across datasets for individual rewards and the full reward set against REE + Rfmt.Source: §M.2 · p. 16, Figure 9 · p. 20, Figure 10 · p. 20