Paper deep dive

Internalizing Academic Writing Workflows for Introduction Generation via Struct-Aware Policy Learning

Meicong Zhang, Tiancheng Su, Jiahao Cheng, Guoxiu He, Xinqi Tao, Dejia Song

Year: 2026Venue: arXiv preprintArea: cs.CLType: PreprintEmbeddings: 85

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 89%

Last extracted: 8/5/2026, 5:10:19 AM

Summary

The paper introduces StructPO, a struct-aware policy learning framework that internalizes multi-stage academic writing workflows into a single-pass policy for generating paper introductions. By using explicit stage tokens and Struct-aware Relative Advantage (SRA) estimation, StructPO decouples local stage quality from global coherence, improving semantic alignment, structural rationality, and inference efficiency compared to external workflow-based baselines and competitive with large closed-source models like GPT-5.1.

Entities (10)

StructPO · methodframework · 98%SRA · algorithmcomponent · 95%Qwen3-32b · model · 92%GPT-5.1 · model · 90%ACL 2025 · datasetcorpus · 88%GRPO · algorithm · 85%PPO · algorithm · 85%AutoSurvey · baseline-system · 82%SurveyForge · baseline-system · 82%STIG · baseline-system · 80%

Relation Signals (8)

StructPO → internalizes → multi-stage writing workflows

confidence 95% · StructPO internalizes multi-stage academic writing workflows into a single-pass policy

StructPO → uses → SRA

confidence 95% · StructPO introduces a Struct-aware Relative Advantage (SRA) estimator.

StructPO → uses → Qwen3-32b

confidence 92% · when scaled to Qwen3-32B

StructPO → iscompetitivewith → GPT-5.1

confidence 90% · remains competitive with GPT-5.1 in human evaluation when scaled to Qwen3-32B.

StructPO → evaluatedon → ACL 2025

confidence 88% · We evaluate StructPO on 1,176 ACL 2025 papers

StructPO → outperforms → SurveyForge

confidence 85% · StructPO consistently improves semantic alignment... over prompt- and workflow-based baselines... SurveyForge... underperforms

StructPO → outperforms → STIG

confidence 85% · STIG obtains strong structure scores... but suffers from length collapse. In contrast, StructPO preserves structural control

StructPO → outperforms → AutoSurvey

confidence 85% · StructPO consistently improves semantic alignment... over prompt- and workflow-based baselines... AutoSurvey... underperforms

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Generating a rigorous paper introduction with large language models (LLMs) remains challenging, since it requires coordinating background, gap identification, method and contribution within a coherent narrative. Existing solutions externalize this process as multi-stage prompts or agent workflows which are expensive and vulnerable to cross-stage drift. We propose StructPO, a struct-aware policy learning framework that internalizes the entire multi-stage writing workflow into a single-pass policy controlled by explicit stage tokens. StructPO introduces struct-aware credit assignment to decouple local stage quality from global coherence and refinement-guided optimization to internalize revision behavior into the first-pass policy. Experiments show that StructPO improves semantic alignment, structural rationality and inference efficiency over workflow-based baselines, generalizes to out-of-domain settings, and remains competitive with GPT-5.1 in human evaluation when scaled to Qwen3-32B. These results show that internalizing academic writing workflows through fine-grained policy optimization offers a viable alternative to costly external orchestration.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

83,989 characters extracted from source content.

Expand or collapse full text

Internalizing Academic Writing Workflows for Introduction Generation via Struct-Aware Policy Learning Meicong Zhang 1 , Tiancheng Su 1 , Jiahao Cheng 1 , Guoxiu He 1∗ , Xinqi Tao 2 , Dejia Song 2 1 School of Economics and Management, East China Normal University, Shanghai, China 2 Xiaohongshu Inc., Shanghai, China mczhang,tcsu,chengjiahao@stu.ecnu.edu.cn, gxhe@fem.ecnu.edu.cn, taoxinqi,dejiasong@xiaohongshu.com Abstract Generating a rigorous paper introduction with large language models (LLMs) remains challenging, since it requires coordi- nating background, gap identification, method and contribution within a coherent narrative. Existing solutions externalize this process as multi-stage prompts or agent workflows which are expensive and vulnerable to cross-stage drift. We pro- pose StructPO, a struct-aware policy learning framework that internalizes the entire multi-stage writing workflow into a single-pass policy controlled by explicit stage tokens. StructPO introduces struct-aware credit assignment to decouple local stage quality from global coherence and refinement-guided optimization to internalize revision behavior into the first-pass policy. Experiments show that StructPO improves seman- tic alignment, structural rationality and inference efficiency over workflow-based baselines, generalizes to out-of-domain settings, and remains competitive with GPT-5.1 in human evaluation when scaled to Qwen3-32B. These results show that internalizing academic writing workflows through fine-grained policy optimization offers a viable alternative to costly external orchestration. Introduction LLMs have recently been applied to many parts of scientific workflows, including literature understanding, hypothesis generation, experimental support and manuscript drafting (Zhang et al. 2025a; Lu et al. 2024; Garikaparthi et al. 2025; Li et al. 2025b; Langley 2024; Wang et al. 2023; Yang et al. 2024b; Zhao et al. 2025; Boiko et al. 2023). In academic paper writing, the introduction is the section where a study first presents its motivation, novelty and contribution to readers. It must connect the study to prior work, identify limitations in existing approaches, motivate the proposed method and preview the paper’s contributions in a coherent rhetorical sequence. We therefore focus on introduction generation, which combines background construction, gap identification, method positioning and contribution framing within a single section. Existing LLM-based academic writing systems have made progress on literature reviews, survey generation, and general ∗ Corresponding author. Copyright © 2027, Association for the Advancement of Artificial Intelligence (w.aaai.org). All rights reserved. manuscript assistance (Wang et al. 2024b; Yan et al. 2025; Liu et al. 2025). To handle long-form writing, these systems often rely on multi-stage prompts or agent workflows that separate outlining, drafting, revising and polishing. However, such external orchestration is usually handcrafted, incurs repeated context replay and may introduce cross-stage drift, where later stages deviate from earlier plans or repeat incompatible rhetorical content. A natural alternative is to internalize the writing workflow into the model, enabling structured introduction generation in a single pass. Yet supervised fine-tuning provides limited direct feedback on long-range rhetorical organization, while standard policy optimization methods such as PPO and GRPO (Schulman et al. 2017; Shao et al. 2024) typically assign one scalar reward to the whole output. For structured writing, this coarse feedback makes it difficult to identify which rhetorical section succeeds or fails, leading to noisy credit assignment and unstable optimization. In this work, we propose StructPO, a struct-aware policy optimization framework for academic introduction generation. StructPO internalizes multi-stage academic writing workflows into a single-pass policy controlled by explicit stage tokens. Specifically, an introduction is represented as eight structural units corresponding to four rhetorical sections: background, problem, method and contributions. Each section consists of an outline and corresponding content. This representation exposes stage boundaries during training and avoids external workflow orchestration during inference. To optimize this structured generation process, StructPO introduces a Struct-aware Relative Advantage (SRA) estima- tor. Rather than assigning the same advantage to all tokens in an introduction, SRA computes stage-level advantages and combines them with a global alignment signal. This allows LLMs to reinforce high-quality behavior within each rhetorical stage while still preserving document-level coher- ence. In addition, StructPO incorporates refinement-guided optimization. During training, LLMs generate draft–revision trajectories and drafts that are substantially improved by their revisions receive an additional penalty. We evaluate StructPO on 1,176 ACL 2025 papers and compare it against prompt-based systems, workflow-based baselines and a suite of strong closed-source LLMs including GPT-4o, GPT-5.1, GPT-5.5, Claude Opus 4.8, and Gemini-3.1- Pro-Preview. StructPO achieves near-perfect structure scores, arXiv:2608.03138v1 [cs.CL] 4 Aug 2026 strong semantic similarity and substantially lower inference overhead. When scaled to a Qwen3-32B backbone, StructPO further becomes competitive with GPT-5.1 in blind human evaluation and matches or surpasses several closed-source LLMs on automated metrics. Further analyses show that StructPO mitigates length collapse, reduces reward hacking and transfers to CVPR papers. Our contributions are summarized as follows: •We formulate academic introduction generation as a single- pass structured policy and propose StructPO to internalize multi-stage writing workflows. • We develop Struct-aware Relative Advantage estimation and a refinement-guided penalty to support stage-level credit assignment and training-time distillation of revision behavior. •We provide empirical evaluations on ACL and CVPR papers, showing that StructPO improves structural control and inference efficiency over prompt-based and workflow- based baselines. When scaled to a Qwen3-32B backbone, StructPO surpasses all evaluated closed-source LLMs on automated metrics and remains competitive with GPT- 5.1 and Claude Opus 4.8 under judge-based and human evaluation. Related Work LLM-based Scientific Writing Systems LLM-based scientific writing systems support idea generation, literature review, experimental design, result analysis, citation assistance and drafting (Su et al. 2025; Wu et al. 2023; Li et al. 2025a; Zimmermann et al. 2024; Agarwal et al. 2024; Schmidgall et al. 2025; Ge et al. 2021; Wang et al. 2021). AutoSurvey and SurveyForge further show the effectiveness of workflow decomposition for literature review (Wang et al. 2024b; Yan et al. 2025). However, introduction writing requires tighter rhetorical coordination among background construction, gap articula- tion, method positioning and contribution framing. Recent systems such as ResearchAgent support broader research- assistance capabilities (Baek et al. 2025), but still largely rely on multi-stage orchestration (Liu et al. 2025). In contrast, we aim to internalize this writing logic within the policy. Step-Level Policy Optimization Reinforcement learning has become a key paradigm for align- ing LLMs with human preferences and task-specific objectives (Ouyang et al. 2022). Policy-gradient methods such as RE- INFORCE (Williams 1992) and PPO (Schulman et al. 2017) optimize policies from scalar rewards, while GRPO (Shao et al. 2024) removes the critic by normalizing rewards within sampled groups. However, trajectory-level rewards are often too coarse for long-horizon generation. Recent work therefore explores finer- grained credit assignment, including hierarchical grouping, Shapley-value decomposition, information-gain-based super- vision and calibrated step-wise advantages (Feng et al. 2025; Li et al. 2026; Huang et al. 2026; Wang et al. 2026; Fei et al. 2025). These methods demonstrate the value of dense optimization signals, but they mainly target homogeneous reasoning tasks such as mathematics or coding. Structured academic writing poses a different challenge because planning and realization stages are heterogeneous yet interdependent. Outlines require coverage and organization, whereas content paragraphs require semantic fidelity, length control, and rhetorical consistency. StructPO addresses this gap by assigning struct-aware advantages to different writing stages while maintaining global introduction-level coherence. Methodology We formulate introduction writing as a structured policy- learning problem. As shown in Figure 1, StructPO com- bines explicit stage decomposition, modality-aware stage rewards, and SRA-based dense optimization. Training uses draft–revision trajectories so the model acquires revision behavior during training while requiring only a single pass at inference time. Task Definition: Explicit Structural Decomposition We formulate introduction generation as conditional sequence generation from inputx =x tit ,x abs ,x exp , where the three terms denote the title, abstract, and extracted experimental metadata. To reduce structural drift, we decompose each introduction intoK = 8units defined by four rhetorical sections and two modalities: S =P bg ,P prob ,P meth ,P contrib ×M out ,M con . (1) Each stage is generated between explicit control tokens, and the full trajectory is y = K M k=1 (⟨s k ⟩⊕ w k ⊕⟨e k ⟩),(2) wherew k = (w k,1 ,...,w k,T k )is the generated text of stage k. This grammar exposes stage boundaries for reward com- putation and enables single-pass structured generation at inference time. Rollout Strategy: Interactive Trajectory Generation To distill revision behavior into the policy, training uses draft– revision trajectories. For each input, the model first generates a drafty d and then a revisiony r , formingτ = [y d ,y r ] = [u 1 ,...,u K ,u K+1 ,...,u 2K ], where the firstKunits belong to the draft phase and the remainingKunits to the revision phase. Following GRPO, we sampleNtrajectories fromπ θ old for each input and compute relative advantages within the sampled group. Reward Design StructPO scores each structural unit separately to provide dense supervision for both planning and content realiza- tion. For dense content stages, we apply Gaussian length regularization,r gau len (w k ) = exp[−(l(w k ) − L ∗ k ) 2 /(2δ 2 )], wherel(w k )is the word count,L ∗ k is the target length, andδis the bandwidth. We setδ = 6for section-level rewards andδ = 12for the global length reward. For Title Abstract Research Outcomes Input Space X StructPO Policy π θ (Reβinement-Guided Introduction Generation) [Background Outline] [Background Content] [Contributions Outline] [Contributions Content] s 1 s 2 s 7 s 8 [Background Outline] [Background Content] [Contributions Outline] [Contributions Content] s 1 s 2 s 7 s 8 · · Explicit Structural Decomposition & Rollout with Reβinement Multi-Dimensional Hierachical Reward Engine Outline Reward Bidirectional Matching (F1-Sim: P sem , R sem ) Linear Length Penalty Content Reward Latent Sementic Consistency (cos(휙(w k ), 휙(w k gt ))) Gaussian Length Reg. (RBF) Structural Rationality Score Dense Reward Vector r = [r 1 1 , . . . , r 1 8 , r 2 1 , . . . , r 2 8 ] Global Reward Vector R global , R reβinement Structured Relative Advantage (SRA) Algorithm Global Advantage Lobal AdvantageFinal Adgantage Segment Advantage A (i) k Dual-Stream Fusion Compute Advantage Dense Advantage Projection A (i) t Drafting Phase y draft Drafting Phase y reβinement Policy Update trajectory Figure 1: Overview of the StructPO training pipeline. Given paper metadata, the policy generates draft–revision trajectories with explicit stage tokens. A modality-aware reward engine scores outline and content stages, and SRA projects fused local-global advantages back to tokens for fine-grained policy optimization. outline stages, we use a linear sentence-count penalty, r lin len (w k ) = 1−|N gen − N ref |/ max(N ref , 1). For content stages, semantic fidelity is measured by embedding similarity,r sim (w k ) = cos(φ(w k ),φ(w gt k )) , whereφ(·)is a pretrained dense encoder such as Qwen- Embedding (Zhang et al. 2025b). For outline stages, we use bidirectional semantic matching. Given generated and refer- ence outline sentence embeddingss i m i=1 andr j n j=1 , with M ij = cos(s i ,r j ), we compute P sem = 1 m m X i=1 max j M ij , R sem = 1 n n X j=1 max i M ij , r match = 2P sem R sem P sem + R sem + ε 0 . (3) To enforce rhetorical purity, we train a DeBERTa-v3 clas- sifier (He, Gao, and Chen 2021) to predict the intended rhetorical section of each sentence. For stagek, the structural rationality score isr struct (w k ) = 1− C mis /C tot , whereC mis is the number of sentences assigned to the wrong section and C tot is the total number of sentences. The final local reward depends on the modality of stagek: r k = ( r match (w k )r lin len (w k ),Outline, ω 1 r sim (w k )r struct (w k ) + ω 2 r gau len (w k ), Content. (4) The reward vectorr = [r 1 ,...,r 2K ]provides dense local supervision, while a global rewardR global measures whole- introduction alignment with the reference. Struct-aware Relative Advantage (SRA) and Dense Optimization Standard GRPO assigns a single trajectory-level signal to every token: ˆg GRPO = E ∇ θ logπ(y t )· R(τ )− μ glob σ glob ,(5) whereR(τ ) = P 2K k=1 r k is the cumulative trajectory reward, andμ glob ,σ glob are the mean and standard deviation ofR(τ ) within the sampled group. For long-form generation, this formulation introduces attribution noise because rewards from unrelated sections affect every token, and the gradient variance grows with cross-section interference: Var(ˆg GRPO )∝ Var(r k ) + X j̸=k Var(r j ).(6) Consequently, a high-quality background paragraph can be penalized because of errors in the method paragraph. To mitigate this issue, StructPO normalizes stage rewards and fuses local and global signals. We first map each local reward to a comparable scale: r k,norm = r k r max k ,(7) wherer max k denotes the theoretical upper bound ofr k derived from its component metrics. We then decouple credit assign- ment into two complementary streams. For trajectoryiand stagek, the local advantage normalizes rewards within the same stage across the group, A (i,k) local = r (i) k − μ k σ k ,(8) whereμ k andσ k are the within-stage mean and standard deviation across theNsampled trajectories. The global advantageA (i) global is defined analogously by normalizing R (i) global within the group. The fused advantage is ˆ A (i) k = λA (i) global + (1− λ)A (i,k) local ,(9) whereλ∈ [0, 1]controls the trade-off between global coher- ence and local stage quality. To incorporate revision supervision, we define the stage- level refinement gap as ∆ (i) k = max 0, r k (τ (i) r )− r k (τ (i) d ) ,(10) whereτ (i) d andτ (i) r denote the draft and revision phases of trajectoryi. The struct-aware relative advantage with refinement guidance is then A (i) k = ˆ A (i) k − I τ (i) ∈ Draft · η ∆ (i) k σ k ,(11) whereη ≥ 0is the refinement penalty coefficient andI(·)is the indicator function. This term explicitly penalizes draft stages that are substantially improved by their revisions. We then project stage-level advantages back to tokens: A (i) t = 2K X k=1 I t (i,k) start ≤ t≤ t (i,k) end ·A (i) k ,(12) wheret (i,k) start andt (i,k) end are the token positions immediately following⟨s k ⟩and immediately preceding⟨e k ⟩in trajectory i. The final objective is L StructPO (θ) = − 1 N N X i=1 T i X t=1 I t∈S (i) val · min ρ (i) t A (i) t , clip ρ (i) t , 1− ε c , 1 + ε c A (i) t + β KL D KL π θ ∥π ref , (13) whereT i is the length of trajectoryi,S (i) val is the set of token positions strictly within valid structural boundaries,ρ (i) t = MethodSem. Sec. Len. Struc. Qwen2.5-7B-Instruct Backbone Pure Prompt0.861 0.672 0.566 0.610 Elaborate Prompt0.813 0.742 0.442 0.725 AutoSurvey0.875 0.743 0.523 0.647 STIG0.858 0.769 0.573 0.884 Refinement w/o training 0.878 0.746 0.693 0.800 SurveyForge0.872 0.745 0.670 0.707 StructPO0.913 0.825 0.708 0.971 Qwen3-8B Backbone Pure Prompt0.871 0.773 0.666 0.761 Elaborate Prompt0.862 0.741 0.703 0.802 AutoSurvey0.884 0.749 0.609 0.727 STIG0.870 0.789 0.487 0.840 Refinement w/o training 0.897 0.753 0.603 0.808 SurveyForge0.878 0.721 0.547 0.755 StructPO0.915 0.825 0.699 0.987 Table 1: Quantitative results of introduction generation on the ACL 2025 dataset. We compare StructPO against open-source prompt-based, workflow-based, and refinement baselines on automated metrics. Sem.: Semantic Similarity; Sec.: Section Similarity; Len.: Length Score; Struc.: Structure Score. π θ (y (i) t | y (i) <t )/π θ old (y (i) t | y (i) <t )is the importance sampling ratio,ε c is the PPO clipping range,β KL ≥ 0controls the KL regularization strength, andπ ref is a frozen reference policy. Unlike standard GRPO, SRA allows the gradient sign to vary within a single trajectory, enabling the model to reinforce desirable local behavior while penalizing undesirable local behavior within the same update. Experiments We construct the experimental corpus from approximately 3,200 ACL conference papers published between 2021 and 2025. PDFs are parsed with MinerU (Wang et al. 2024a), from which we extract titles, abstracts, introductions, figure/table captions, table contents, and available experimental metadata. ACL 2025 papers are held out exclusively for testing, yielding 1,176 test instances; the remaining ACL 2021–2024 papers are split into approximately 300 SFT papers, 1,550 RL-training papers, and 150 validation papers. Stage-level supervision is obtained by using GPT-4o (Achiam et al. 2023) with a fixed decomposition prompt to segment each introduction into background, problem, method, and contribution sections, together with corresponding outlines. For out-of-domain evaluation, we additionally collect 141 unseen CVPR papers, which are used only for zero-shot transfer evaluation and never for training or validation. Baselines and Models We instantiate StructPO on three open-source backbones of increasing scale: Qwen2.5-7B-Instruct, Qwen3-8B, and Qwen3-32B (Yang et al. 2024a, 2025), using Verl as the training framework and SGLang for rollout and inference MethodSem. Sec. Len. Struc. Auto Overall AWQ SFC Sound. Pre. Con. Judge Overall StructPO (Qwen2.5-7B) 0.913 0.825 0.708 0.9713.4173.095 4.439 2.553 2.981 2.8743.188 StructPO (Qwen3-8B) 0.915 0.825 0.699 0.9873.4263.491 4.629 2.757 3.437 3.1263.488 StructPO (Qwen3-32B) 0.914 0.825 0.701 0.9823.4224.098 4.934 3.178 3.533 3.3463.818 GPT-4o0.866 0.625 0.745 0.8883.1242.959 4.993 2.951 3.039 2.9133.371 GPT-5.10.915 0.822 0.656 0.9413.3344.063 4.970 3.187 3.317 3.6973.847 GPT-5.50.908 0.807 0.480 0.8773.0724.199 4.992 3.117 3.077 3.1233.702 Claude Opus 4.80.922 0.828 0.660 0.9583.3684.073 4.983 3.429 3.394 3.2913.834 Gemini-3.1-Pro-Preview 0.890 0.787 0.703 0.9603.3404.108 4.983 3.233 3.618 3.0203.792 Table 2: Comprehensive evaluation of StructPO and strong closed-source LLMs. Auto Overall is the sum of four automated metrics: Sem., Sec., Len., and Struc. Judge Overall is the average of five LLM-as-a-judge metrics: AWQ, SFC, Soundness, Presentation, and Contribution. Soundness, Presentation, and Contribution are reviewer-style rubric metrics used only for evaluation and are not used during StructPO training. (Sheng et al. 2025; Zheng et al. 2024). The 7B and 8B variants are trained on 8 H800 GPUs with a global batch size of 16 and a rollout size of 8; the 32B variant follows the same training recipe with adjusted parallelism to fit the larger model. We compare against Pure Prompt, ELABORATE Prompting (Garg et al. 2025), AutoSurvey (Wang et al. 2024b), STIG (Zhang, su, and He 2025), SurveyForge (Yan et al. 2025), refinement without training and a suite of strong closed-source LLMs including GPT-4o, GPT-5.1, GPT-5.5, Claude Opus 4.8 and Gemini-3.1-Pro-Preview. All comparable open-source baselines use the same test set, input metadata, and decoding protocol. Evaluation Metrics We evaluate generated introductions with automated and LLM-as-a-judge metrics. Automated metrics include full- introduction semantic similarity (Sem.), section-level similar- ity (Sec.), length score (Len.), structure score (Struc.), and Auto Overall, the sum of the four normalized scores. Fol- lowing Zheng et al. (2023), judge metrics include Academic Writing Quality (AWQ), Scientific and Factual Consistency (SFC), Soundness, Presentation, and Contribution; their average is reported as Judge Overall. Soundness assesses accurate and efficient reflection of core paper information, Presentation assesses rhetorical placement and coherent flow, and Contribution assesses whether the main contributions are stated clearly, specifically, and concisely. These three reviewer-style metrics are used only for evalu- ation, not for StructPO training. For the main comparison in Table 2 we report all five judge metrics, while the ablation in Table 3 reports only AWQ and SFC to keep the table compact, as these two metrics most directly capture writing quality and factual grounding. All methods are evaluated with the same fixed judge prompt and scoring rubric. Main Results Table 1 shows that StructPO consistently improves semantic alignment, section-level quality, and structural rationality over prompt- and workflow-based baselines on both open-source backbones. Although AutoSurvey decomposes writing into stages, its external workflow still underperforms a single-pass policy trained with stage-level credit assignment, suggesting that decomposition must be internalized to avoid cross-stage drift. STIG obtains strong structure scores through stage- token generation, but suffers from length collapse. In contrast, StructPO preserves structural control while producing richer and more semantically aligned content. Table 2 compares StructPO with strong closed-source LLMs. On automated metrics, all StructPO variants achieve higher Auto Overall scores than the closed-source models, mainly through stronger section alignment and structural rationality. On judge metrics, closed-source LLMs retain advantages in factual reliability and contribution framing, likely due to stronger pretrained knowledge and instruction following. However, scaling StructPO from 8B to 32B sub- stantially improves Judge Overall, surpassing GPT-5.5 and Gemini-3.1-Pro-Preview and nearly matching GPT-5.1 and Claude Opus 4.8. These results suggest that StructPO pro- vides robust structural control, while larger backbones further improve content quality. These findings suggest that StructPO is complementary to backbone scaling, with policy optimiza- tion providing reliable rhetorical structure and larger models improving factual grounding and fluency. Analysis Ablation Study Table 3 shows the contribution of each component. SFT learns the stage-token format but suffers from length collapse, especially on Qwen3-8B where the length score is only 0.529, indicating that supervised imitation captures surface structure but not long-form rhetorical completeness. Standard GRPO improves several automatic metrics but is less stable because it assigns one trajectory-level advantage to all tokens; on Qwen2.5-7B, this causes a large drop in factual and citation consistency. Removing struct-aware ad- vantage assignment also degrades performance, particularly length control on Qwen3-8B, suggesting that heterogeneous stages require separate credit assignment. The global reward MethodSem. (↑) Sec. (↑) Len. (↑) Struc. (↑) AWQ (↑) SFC (↑) Overall (↑) Base Model: Qwen2.5-7B-Instruct Refinement w/o training0.8780.7460.6930.8002.6523.29074.214 SFT0.8850.7860.5470.8322.9294.41975.682 GRPO0.9210.8290.6940.9863.0802.25879.290 StructPO w/o Refinement0.9120.8280.6970.9593.0604.36682.764 StructPO w/o Global Reward0.9100.8210.7060.9753.0094.35782.976 StructPO w/o Struct-Awareness 0.9080.8230.6440.9853.2624.50282.734 StructPO (Ours)0.9130.8250.7080.9713.0954.43983.398 Base Model: Qwen3-8B Refinement w/o training0.8970.7530.6030.8082.7724.97476.674 SFT0.8860.7900.5290.8303.0214.55775.849 GRPO0.9040.7880.6630.9843.3374.97283.387 StructPO w/o Refinement0.9150.8270.6630.9833.5044.72484.217 StructPO w/o Global Reward0.9080.8210.6380.9813.6064.88483.932 StructPO w/o Struct-Awareness 0.9040.8200.5230.9553.4214.85680.590 StructPO (Ours)0.9150.8250.6990.9873.4914.62984.734 Table 3: Ablation study on different components of StructPO. Refinement w/o training: Inference-time refinement using SFT model; SFT: Supervised Fine-Tuning baseline; GRPO: Standard trajectory-level policy optimization; w/o Refinement: Removed refinement-guided penalty; w/o Global: Removed global semantic alignment reward; w/o Struct-Aware: Removed stage-level local advantage assignment. For brevity, we report AWQ and SFC as representative judge metrics; Overall is a composite score defined as the sum of the four automated metrics (rescaled to[0, 100]) and the two judge metrics, so its scale differs from the Auto Overall in Table 2. The best results are highlighted in bold. and refinement-guided penalty provide complementary gains: without the global reward, semantic alignment weakens; with- out refinement guidance, the model loses part of the benefit from draft–revision comparisons. StructPO achieves the best overall performance on both backbones, confirming that lo- cal stage rewards, global alignment, and refinement-guided training jointly improve structured introduction generation. The ablation results also indicate that no single component dominates across all metrics; instead, StructPO benefits from balancing local controllability, global semantic alignment, and revision-aware regularization. We omit full Qwen3-32B ablations due to training cost, but trends across 7B and 8B suggest similar conclusions at larger scale. Qualitative Case Study We examine a representative ACL 2025 test case, the Speech- Fake paper (Huang et al. 2025). AutoSurvey generates detailed content but exhibits cross-stage drift, such as revealing the proposed dataset too early and producing markdown-style meta-commentary. STIG preserves the four-part skeleton, but its output is short, repetitive, and weak in contribution synthesis. StructPO produces a more balanced introduction. It follows the background–problem–method–contribution progression with clearer transitions, while incorporating key details such as dataset characteristics, equal error rate, and cross-speaker evaluation. Compared with AutoSurvey, it avoids workflow fragmentation; compared with STIG, it provides denser con- tent and stronger contribution framing. This qualitative pattern is consistent with the quantitative results, where StructPO improves structure without collapsing output length. A full side-by-side example is included in the Appendix . Token Efficiency Analysis We compare inference-time token consumption using Qwen3- 8B as the common backbone. As shown in Figure 2, total usage is divided into overhead tokens (input context, prompts, and intermediate workflow states) and effective tokens (final introduction content), with the effective rate defined as the ratio of effective tokens to total tokens. Workflow-based systems incur large overhead due to repeated context replay: AutoSurvey generates 778 effective tokens but requires 19,469 overhead tokens, yielding only a 3.84% effective rate. In contrast, single-pass baselines reduce orchestration cost but may sacrifice completeness; STIG produces only 422 effective tokens, indicating length collapse. StructPO achieves a better efficiency–completeness trade- off. Since refinement behavior is distilled into the policy during training, StructPO performs single-pass inference without external revision loops. It generates 523 effective tokens with 4,976 overhead tokens, reducing overhead by 74.4% compared with AutoSurvey while producing richer content than STIG. This shows that internalizing writing workflows can substantially improve deployment efficiency without costly multi-turn orchestration. Generalization Study To assess cross-domain generalization, we evaluate ACL- trained checkpoints on 141 unseen CVPR papers without domain-specific tuning. As shown in Figure 3, StructPO Pure Prompt AutoSurvey STIGStructPO 0 5,000 10,000 15,000 20,000 25,000 Token Count Effective Token (Visual x3) Overhead Token Effective Rate (%) Best Efficiency 593 5,905 6,498 778 19,469 20,247 422 4,435 4,857 523 4,976 5,499 0% 5% 10% 15% 20% 25% Effective Rate (%) 9.13% 3.84% 8.69% 9.51% Figure 2: Inference-token comparison across generation paradigms. StructPO achieves a favorable Pareto balance between informative content and computational overhead, avoiding both the token bloat of AutoSurvey and the length collapse of STIG. Effective token bars are visually scaled 3× for clarity. Sem.Sec.Len.Struc.AWQSFC 55 65 75 85 95 105 Normalized Score (0–100) 91.2 (0.912) 81.8 (0.818) 63.8 (0.638) 98.2 (0.982) 61.8 (3.09) 90.2 (4.51) 91.5 (0.915) 82.2 (0.822) 63.0 (0.630) 98.7 (0.987) 68.8 (3.44) 92.9 (4.64) 91.3 (0.913) 82.4 (0.824) 64.7 (0.647) 97.6 (0.976) 80.8 (4.04) 99.4 (4.97) 91.4 (0.914) 81.2 (0.812) 71.5 (0.715) 94.2 (0.942) 80.1 (4.01) 99.7 (4.99) StructPO (Qwen2.5-7B-Instruct)StructPO (Qwen3-8B)StructPO (Qwen3-32B)GPT-5.1 Figure 3: Zero-shot transfer results from ACL to 141 unseen CVPR papers. StructPO variants preserve strong structural control under domain shift, consistently outperforming GPT-5.1 on structure score and section-level alignment. GPT-5.1 retains an advantage in length control and factual consistency, while StructPO with Qwen3-32B achieves the best academic writing quality score. transfers its rhetorical scaffolding effectively: all variants preserve strong structural rationality and outperform GPT-5.1 on structure and section-level alignment, despite GPT-5.1’s broader exposure to vision-domain text. On content-oriented judge metrics, StructPO-32B is close to GPT-5.1, matching factual consistency and slightly im- proving academic writing quality. Semantic similarity is also comparable across systems. These results suggest that stage- level rewards learn a domain-general introduction structure with strong transferability, while remaining weaknesses under domain shift are mainly stylistic and length-related. Adaptive length targets and retrieval augmentation may further improve transfer. Human Evaluation We conduct a blind human evaluation on 30 randomly sam- pled ACL test papers, comparing StructPO with Qwen3-32B against GPT-5.1. Three NLP researchers judge anonymized outputs based on coherence, structural completeness and academic writing quality. As shown in Table 4, StructPO achieves a 53.3% win rate by majority voting and receives 55.6% of all annotator votes. These results indicate that, with a stronger open-source backbone, StructPO can be pre- ferred over GPT-5.1 in human evaluation, although the margin remains moderate. Comparison StructPO Win GPT-5.1 Win Win Rate53.3% (16/30) 46.7% (14/30) Win Votes55.6% (50/90) 44.4% (40/90) Table 4: Human evaluation results comparing StructPO with Qwen3-32B against GPT-5.1. Win Rate is computed by majority voting over 30 samples, and Win Votes are computed from all 90 annotator votes. Conclusion We proposed StructPO, a structure-aware policy optimization framework that internalizes multi-stage academic introduction writing into a single-pass stage-token policy. By combining stage-level credit assignment with refinement-guided opti- mization, StructPO improves local rhetorical quality, global coherence, and inference efficiency. Experiments on ACL papers show consistent gains over prompt- and workflow- based baselines, and zero-shot evaluation on CVPR papers demonstrates effective transfer. StructPO still relies on a fixed eight-stage template and may require adaptation for theoretical, survey, or non-standard papers. Future work may explore adaptive structural planning and retrieval-augmented grounding within the SRA reward loop (Lewis et al. 2020). References Achiam, J.; Adler, S.; Agarwal, S.; Ahmad, L.; Akkaya, I.; Aleman, F. L.; Almeida, D.; Altenschmidt, J.; Altman, S.; Anadkat, S.; et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774. Agarwal, S.; Sahu, G.; Puri, A.; Laradji, I. H.; Dvijotham, K. D.; Stanley, J.; Charlin, L.; and Pal, C. 2024. Litllm: A toolkit for scientific literature review. arXiv preprint arXiv:2402.01788. Baek, J.; Jauhar, S. K.; Cucerzan, S.; and Hwang, S. J. 2025. ResearchAgent: Iterative Research Idea Generation over Scien- tific Literature with Large Language Models. In Chiruzzo, L.; Ritter, A.; and Wang, L., eds., Proceedings of the 2025 Confer- ence of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technolo- gies (Volume 1: Long Papers), 6709–6738. Albuquerque, New Mexico: Association for Computational Linguistics. ISBN 979-8-89176-189-6. Boiko, D. A.; MacKnight, R.; Kline, B.; and Gomes, G. 2023. Autonomous chemical research with large language models. Nature, 624(7992): 570–578. Fei, W.; Kong, H.; Liang, S.; Lin, Y.; Yang, Y.; Tang, J.; Chen, L.; and Hua, X. 2025. Self-Guided Process Reward Optimization with Redefined Step-wise Advantage for Process Reinforcement Learning. arXiv:2507.01551. Feng, L.; Xue, Z.; Liu, T.; and An, B. 2025. Group-in-Group Policy Optimization for LLM Agent Training. In The Thirty- ninth Annual Conference on Neural Information Processing Systems. Garg, K.; Shaikh, F.; Bandyopadhyay, S.; and Caragea, C. 2025. Let’s Use ChatGPT To Write Our Paper! Benchmarking LLMs To Write the Introduction of a Research Paper. arXiv preprint arXiv:2508.14273. Garikaparthi, A.; Patwardhan, M.; Kanade, A. S.; Hassan, A.; Vig, L.; and Cohan, A. 2025. MIR: Methodology Inspira- tion Retrieval for Scientific Research Problems. In Che, W.; Nabende, J.; Shutova, E.; and Pilehvar, M. T., eds., Proceed- ings of the 63rd Annual Meeting of the Association for Compu- tational Linguistics (Volume 1: Long Papers), 28614–28659. Vienna, Austria: Association for Computational Linguistics. ISBN 979-8-89176-251-0. Ge, Y.; Dinh, L.; Liu, X.; Su, J.; Lu, Z.; Wang, A.; and Diesner, J. 2021. BACO: A Background Knowledge- and Content- Based Framework for Citing Sentence Generation. In Zong, C.; Xia, F.; Li, W.; and Navigli, R., eds., Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), 1466– 1478. Online: Association for Computational Linguistics. He, P.; Gao, J.; and Chen, W. 2021. DeBERTaV3: Improving DeBERTa using ELECTRA-Style Pre-Training with Gradient- Disentangled Embedding Sharing. arXiv:2111.09543. He, P.; Liu, X.; Gao, J.; and Chen, W. 2021. DEBERTA: DECODING-ENHANCED BERT WITH DISENTANGLED ATTENTION. In International Conference on Learning Representations. Huang, M.; Zhang, L.; Li, Y.; Wu, Y.; and Liu, J. 2026. SketchVL: Policy Optimization via Fine-Grained Credit Assignment for Chart Understanding and More. arXiv:2601.05688. Huang, W.; Gu, Y.; Wang, Z.; Zhu, H.; and Qian, Y. 2025. SpeechFake: A Large-Scale Multilingual Speech Deepfake Dataset Incorporating Cutting-Edge Generation Methods. In Che, W.; Nabende, J.; Shutova, E.; and Pilehvar, M. T., eds., Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 9985–9998. Vienna, Austria: Association for Computational Linguistics. ISBN 979-8-89176-251-0. Langley, P. 2024. Integrated Systems for Computational Scientific Discovery. Proceedings of the AAAI Conference on Artificial Intelligence, 38(20): 22598–22606. Lewis, P.; Perez, E.; Piktus, A.; Petroni, F.; Karpukhin, V.; Goyal, N.; Küttler, H.; Lewis, M.; Yih, W.-t.; Rocktäschel, T.; et al. 2020. Retrieval-augmented generation for knowledge- intensive nlp tasks. Advances in neural information processing systems, 33: 9459–9474. Li, L.; Xu, W.; Guo, J.; Zhao, R.; Li, X.; Yuan, Y.; Zhang, B.; Jiang, Y.; Xin, Y.; Dang, R.; Rong, Y.; Zhao, D.; Feng, T.; and Bing, L. 2025a. Chain of Ideas: Revolutionizing Research Via Novel Idea Development with LLM Agents. In Christodoulopoulos, C.; Chakraborty, T.; Rose, C.; and Peng, V., eds., Findings of the Association for Computa- tional Linguistics: EMNLP 2025, 8971–9004. Suzhou, China: Association for Computational Linguistics. ISBN 979-8- 89176-335-7. Li, S.; Huang, J.; Zhuang, J.; Shi, Y.; Cai, X.; Xu, M.; Wang, X.; Zhang, L.; Ke, G.; and Cai, H. 2025b. SciLitLLM: How to Adapt LLMs for Scientific Literature Understanding. In The Thirteenth International Conference on Learning Representations. Li, Y.; Cao, Z.; Qiao, J.; and Hu, S. 2026. SSVPO: Effective Step-Level Credit Assignment for RL Training of Language Models. In The Fourteenth International Conference on Learning Representations. Liu, C.; Wang, C.; Cao, J.; Ge, J.; Wang, K.; Zhang, L.; Cheng, M.-M.; Zhao, P.; Li, T.; Jia, X.; Li, X.; Li, X.; Liu, Y.; Feng, Y.; Huang, Y.; Xu, Y.; Sun, Y.; Zhou, Z.; and Xu, Z. 2025. A Vision for Auto Research with LLM Agents. CoRR, abs/2504.18765. Lu, C.; Lu, C.; Lange, R. T.; Foerster, J. N.; Clune, J.; and Ha, D. 2024. The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery. CoRR, abs/2408.06292. Ouyang, L.; Wu, J.; Jiang, X.; Almeida, D.; Wainwright, C.; Mishkin, P.; Zhang, C.; Agarwal, S.; Slama, K.; Ray, A.; et al. 2022. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35: 27730–27744. Schmidgall, S.; Su, Y.; Wang, Z.; Sun, X.; Wu, J.; Yu, X.; Liu, J.; Moor, M.; Liu, Z.; and Barsoum, E. 2025. Agent Laboratory: Using LLM Agents as Research Assistants. arXiv:2501.04227. Schulman, J.; Wolski, F.; Dhariwal, P.; Radford, A.; and Klimov, O. 2017. Proximal Policy Optimization Algorithms. arXiv:1707.06347. Shao, Z.; Wang, P.; Zhu, Q.; Xu, R.; Song, J.; Bi, X.; Zhang, H.; Zhang, M.; Li, Y. K.; Wu, Y.; and Guo, D. 2024. DeepSeek- Math: Pushing the Limits of Mathematical Reasoning in Open Language Models. arXiv:2402.03300. Sheng, G.; Zhang, C.; Ye, Z.; Wu, X.; Zhang, W.; Zhang, R.; Peng, Y.; Lin, H.; and Wu, C. 2025. Hybridflow: A flexible and efficient rlhf framework. In Proceedings of the Twentieth European Conference on Computer Systems, 1279–1297. Su, H.; Chen, R.; Tang, S.; Yin, Z.; Zheng, X.; Li, J.; Qi, B.; Wu, Q.; Li, H.; Ouyang, W.; Torr, P.; Zhou, B.; and Dong, N. 2025. Many Heads Are Better Than One: Improved Scientific Idea Generation by A LLM-Based Multi-Agent System. In Che, W.; Nabende, J.; Shutova, E.; and Pilehvar, M. T., eds., Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 28201–28240. Vienna, Austria: Association for Computational Linguistics. ISBN 979-8-89176-251-0. Wang, B.; Xu, C.; Zhao, X.; Ouyang, L.; Wu, F.; Zhao, Z.; Xu, R.; Liu, K.; Qu, Y.; Shang, F.; Zhang, B.; Wei, L.; Sui, Z.; Li, W.; Shi, B.; Qiao, Y.; Lin, D.; and He, C. 2024a. MinerU: An Open-Source Solution for Precise Document Content Extraction. arXiv:2409.18839. Wang, G.; Dai, S.; Ye, G.; Gan, Z.; Yao, W.; Deng, Y.; Wu, X.; and Ying, Z. 2026. Information Gain-based Policy Opti- mization: A Simple and Effective Approach for Multi-Turn Search Agents. In The Fourteenth International Conference on Learning Representations. Wang, H.; Fu, T.; Du, Y.; Gao, W.; Huang, K.; Liu, Z.; Chandak, P.; Liu, S.; Van Katwyk, P.; Deac, A.; et al. 2023. Scientific discovery in the age of artificial intelligence. Nature, 620(7972): 47–60. Wang, Q.; Xiong, Y.; Zhang, Y.; Zhang, J.; and Zhu, Y. 2021. Autocite: Multi-modal representation fusion for contextual citation generation. In Proceedings of the 14th ACM In- ternational Conference on Web Search and Data Mining, 788–796. Wang, Y.; Guo, Q.; Yao, W.; Zhang, H.; Zhang, X.; Wu, Z.; Zhang, M.; Dai, X.; Zhang, M.; Wen, Q.; Ye, W.; Zhang, S.; and Zhang, Y. 2024b. AutoSurvey: Large Language Models Can Automatically Write Surveys. In Globerson, A.; Mackey, L.; Belgrave, D.; Fan, A.; Paquet, U.; Tomczak, J.; and Zhang, C., eds., Advances in Neural Information Processing Systems, volume 37, 115119–115145. Curran Associates, Inc. Williams, R. J. 1992. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning, 8(3): 229–256. Wu, Q.; Bansal, G.; Zhang, J.; Wu, Y.; Zhang, S.; Zhu, E.; Li, B.; Jiang, L.; Zhang, X.; and Wang, C. 2023. Autogen: En- abling next-gen llm applications via multi-agent conversation framework. arXiv preprint arXiv:2308.08155, 3(4). Yan, X.; Feng, S.; Yuan, J.; Xia, R.; Wang, B.; Zhang, B.; and Bai, L. 2025. SurveyForge: On the Outline Heuristics, Memory-Driven Generation, and Multi-dimensional Evalua- tion for Automated Survey Writing. CoRR, abs/2503.04629. Yang, A.; Li, A.; Yang, B.; Zhang, B.; Hui, B.; Zheng, B.; Yu, B.; Gao, C.; Huang, C.; Lv, C.; Zheng, C.; Liu, D.; Zhou, F.; Huang, F.; Hu, F.; Ge, H.; Wei, H.; Lin, H.; Tang, J.; Yang, J.; Tu, J.; Zhang, J.; Yang, J.; Yang, J.; Zhou, J.; Zhou, J.; Lin, J.; Dang, K.; Bao, K.; Yang, K.; Yu, L.; Deng, L.; Li, M.; Xue, M.; Li, M.; Zhang, P.; Wang, P.; Zhu, Q.; Men, R.; Gao, R.; Liu, S.; Luo, S.; Li, T.; Tang, T.; Yin, W.; Ren, X.; Wang, X.; Zhang, X.; Ren, X.; Fan, Y.; Su, Y.; Zhang, Y.; Zhang, Y.; Wan, Y.; Liu, Y.; Wang, Z.; Cui, Z.; Zhang, Z.; Zhou, Z.; and Qiu, Z. 2025. Qwen3 Technical Report. arXiv:2505.09388. Yang, A.; Yang, B.; Hui, B.; Zheng, B.; Yu, B.; Zhou, C.; Li, C.; Li, C.; Liu, D.; Huang, F.; Dong, G.; Wei, H.; Lin, H.; Tang, J.; Wang, J.; Yang, J.; Tu, J.; Zhang, J.; Ma, J.; Xu, J.; Zhou, J.; Bai, J.; He, J.; Lin, J.; Dang, K.; Lu, K.; Chen, K.; Yang, K.; Li, M.; Xue, M.; Ni, N.; Zhang, P.; Wang, P.; Peng, R.; Men, R.; Gao, R.; Lin, R.; Wang, S.; Bai, S.; Tan, S.; Zhu, T.; Li, T.; Liu, T.; Ge, W.; Deng, X.; Zhou, X.; Ren, X.; Zhang, X.; Wei, X.; Ren, X.; Fan, Y.; Yao, Y.; Zhang, Y.; Wan, Y.; Chu, Y.; Liu, Y.; Cui, Z.; Zhang, Z.; and Fan, Z. 2024a. Qwen2 Technical Report. arXiv preprint arXiv:2407.10671. Yang, Z.; Du, X.; Li, J.; Zheng, J.; Poria, S.; and Cambria, E. 2024b. Large Language Models for Automated Open-domain Scientific Hypotheses Discovery. In Ku, L.-W.; Martins, A.; and Srikumar, V., eds., Findings of the Association for Com- putational Linguistics: ACL 2024, 13545–13565. Bangkok, Thailand: Association for Computational Linguistics. Zhang, M.; su, T.; and He, G. 2025. Eliminating Agentic Workflow for Introduction Generation with Parametric Stage Tokens. arXiv:2601.09728. Zhang, Y.; Khan, S. A.; Mahmud, A.; Yang, H.; Lavin, A.; Levin, M.; Frey, J.; Dunnmon, J.; Evans, J.; Bundy, A.; et al. 2025a. Exploring the role of large language models in the scientific method: from hypothesis to discovery. npj Artificial Intelligence, 1(1): 14. Zhang, Y.; Li, M.; Long, D.; Zhang, X.; Lin, H.; Yang, B.; Xie, P.; Yang, A.; Liu, D.; Lin, J.; Huang, F.; and Zhou, J. 2025b. Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models. arXiv preprint arXiv:2506.05176. Zhao, Y.; Chen, W.; Xu, Z.; Patwardhan, M.; Wang, C.; Liu, Y.; Vig, L.; and Cohan, A. 2025. AbGen: Evaluating Large Language Models in Ablation Study Design and Evaluation for Scientific Research. In Che, W.; Nabende, J.; Shutova, E.; and Pilehvar, M. T., eds., Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 12479–12491. Vienna, Austria: Association for Computational Linguistics. ISBN 979-8- 89176-251-0. Zheng, L.; Chiang, W.-L.; Sheng, Y.; Zhuang, S.; Wu, Z.; Zhuang, Y.; Lin, Z.; Li, Z.; Li, D.; Xing, E.; et al. 2023. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in neural information processing systems, 36: 46595–46623. Zheng, L.; Yin, L.; Xie, Z.; Sun, C.; Huang, J.; Yu, C. H.; Cao, S.; Kozyrakis, C.; Stoica, I.; Gonzalez, J. E.; et al. 2024. Sglang: Efficient execution of structured language model programs. Advances in neural information processing systems, 37: 62557–62583. Zimmermann, R.; Staab, M.; Nasseri, M.; and Brandtner, P. 2024. Leveraging large language models for literature review tasks-a case study using chatgpt. In International Confer- ence on Advanced Research in Technologies, Information, Innovation and Sustainability, 313–323. Springer. Structural Rationality Illustration Auxiliary classifier. The Structural Rationality Score relies on an auxiliary classifier that predicts the intended rhetorical role of each sentence in a generated introduction. We use GPT-4o to annotate the rhetorical structure of approximately 1,000 introductions, drawn from both model-generated drafts and ACL Findings papers, and then train a DeBERTa-v3- large (He, Gao, and Chen 2021; He et al. 2021) classifier on the resulting sentence-level labels. The trained classifier achieves an accuracy of 94.76% on a held-out test set, which is sufficient for providing reliable structural feedback during reinforcement learning. Score computation. Given a generated paragraphw k as- signed to a target rhetorical section (e.g., Background or Problem Statement), the classifier labels each constituent sentence with one of the four canonical sections. Sentences whose predicted label differs from the target are marked as structurally irrational. The Structural Rationality Score for w k is then computed asr struct (w k ) = 1− C mis /C tot , where C tot is the total number of sentences inw k andC mis is the number of mislabeled sentences. This formulation directly penalizes cross-stage content leakage and rewards rhetorical purity within each structural unit. Illustrative example. Figure 4 visualizes the calculation process on two paragraphs generated by AutoSurvey, where the first paragraph is labeled as Background and the sec- ond as Problem Statement. Sentences highlighted in dark red are classified as structurally erroneous. Concretely, the Background paragraph contains 7 sentences, of which 2 pre- maturely describe methodological details, while the Problem Statement paragraph contains 8 sentences. Aggregating across both paragraphs yields a final Structural Rationality Score of0.873. The case also exposes a typical failure mode of multi-stage workflows: the model first states the problem, immediately proposes a dataset that resolves the issue, and then re-opens existing problems in the second paragraph, leading to repetitive mentions and structural confusion that StructPO is explicitly trained to avoid. Additional Experimental Details This section provides additional implementation details that complement the experimental setup described in Section Experiments. These details include dataset splits, training configurations, baseline implementations, metric aggregation and token accounting. Dataset Splits and Construction We construct the experimental corpus from approximately 3,200 ACL conference papers published between 2021 and 2025. PDF files are parsed into structured text with MinerU (Wang et al. 2024a). For each paper, we extract the title, abstract, original introduction, figure captions, table cap- tions and table contents. We also extract baseline-reference information from experimental sections when available. The ACL 2025 papers are held out exclusively for testing, yielding 1,176 test instances. The remaining ACL papers are used for training and validation. Specifically, we use approximately 300 papers for supervised fine-tuning, around 1,550 papers for reinforcement learning training, and about 150 papers for validation during training. The held-out ACL 2025 set is not used in either SFT or RL training. To obtain stage-level supervision, we decompose each orig- inal introduction into four rhetorical sections: Background, Problem and Limitations of Existing Methods, Brief Method Overview and Summary of Main Results, and Our Contri- butions. For each section, we further extract an outline. The decomposition is performed with GPT-4o (Achiam et al. 2023) using the prompt in Appendix . The resulting section-content and section-outline pairs are used for supervised training, reward computation and evaluation. For out-of-domain evaluation, we additionally collect 141 unseen CVPR papers. These papers are used only for zero- shot transfer evaluation. No CVPR paper is used during SFT, RL training, reward-model construction or validation. SplitSource#Usage SFTACL 21–24∼300SFT RLACL 21–24∼1,550StructPO Val.ACL 21–24∼150Validation TestACL 20251,176Evaluation OODCVPR141Transfer Table 5: Dataset splits used in our experiments. Training Configuration StructPO is implemented with the Verl reinforcement learn- ing framework (Sheng et al. 2025), and SGLang (Zheng et al. 2024) is used for rollout and inference. We instantiate StructPO on Qwen2.5-7B-Instruct and Qwen3-8B (Yang et al. 2024a, 2025). All reinforcement learning experiments use 8 H800 GPUs, a global batch size of 16 and a rollout size of 8. For the dual-stream advantage fusion in Struct-aware Rela- tive Advantage estimation, we set the global advantage weight toλ = 0.3and the local struct-aware advantage weight to1− λ = 0.7. This setting places stronger emphasis on fine-grained stage-level control while still preserving whole- introduction semantic alignment. During RL training, we also apply a KL-divergence constraint to regularize the updated policy against excessive deviation from the reference model. The length bandwidths follow the reward design:δ = 6 for section-level content length rewards andδ = 12for the global length reward. During training, only tokens located within valid structural boundaries are optimized. Baseline Implementation Details We compare StructPO with prompt-based, workflow-based, supervised stage-token and closed-source LLM baselines. Background The rise of speech deepfakes represents a significant security and ethical challenge in the digital age, driven by rapid advancements in artificial intelligence and speech synthesis technologies. These deepfakes, capable of generating highly realistic synthetic speech, have proliferated across various domains, from social media to political discourse, raising concerns about misinformation, identity deception, and the manipulation of public opinion. The ease with which speech deepfakes can be created and disseminated has outpaced the development of effective detection mechanisms, leaving critical vulnerabilities in digital communication. Recent breakthroughs in text-to-speech synthesis, voice conversion, and neural vocoder techniques have enabled the production of high-fidelity audio that is increasingly difficult to distinguish from authentic speech. This growing threat underscores the urgent need for robust detection systems and comprehensive datasets to train and evaluate such models. In response, the SpeechFake dataset was developed as a large-scale, multilingual resource encompassing over 3 million deepfake samples across 46 languages. By incorporating diverse generation methods, SpeechFake provides a benchmark for advancing detection capabilities and understanding the factors influencing model performance in real-world scenarios. Current speech deepfake detection systems face significant limitations that hinder their effectiveness in real-world applications. Many existing datasets suffer from insufficient scale, limited diversity, and inadequate representation of modern generation techniques, resulting in models that lack generalizability and robustness. These shortcomings are exacerbated by the rapid evolution of speech synthesis methods, which often outpace the capabilities of detection systems trained on outdated or narrowly focused data. Furthermore, the distinction between natural speech and synthetic audio remains challenging due to the increasing sophistication of deepfake generation tools, which can closely mimic human speech patterns and prosody. As a result, detection models frequently exhibit poor performance on unseen or adversarially generated samples. The lack of multilingual support in many datasets further restricts their applicability in global contexts, where speech deepfakes may be produced and consumed across diverse linguistic environments. These limitations not only impede the development of reliable detection systems but also slow progress in understanding the underlying characteristics that differentiate synthetic from authentic speech. Addressing these challenges is critical for advancing the field and setting the stage for the development of a comprehensive dataset capable of meeting the evolving needs of speech deepfake detection. Problem Statement Structural Rationality Score = 1 - Cmis / Ctotal = 0.873 Cmis = 2Ctotal = 15 Figure 4: Illustration of the calculation of the structure score. Pure Prompt. The model directly generates the introduc- tion from the title, abstract, figure information, table informa- tion and baseline-reference information. No explicit outline guidance, stage tokens or refinement loop is used. ELABORATE Prompting. We adapt ELABORATE prompting (Garg et al. 2025) to introduction generation by enforcing a four-part rhetorical structure covering context, research gap, proposed solution and contribution. Unlike StructPO, this baseline relies only on prompt-level structure control and does not update the model parameters. AutoSurvey. AutoSurvey (Wang et al. 2024b) is adapted as a workflow-based baseline. The original survey-generation workflow is modified for introduction generation. The system first produces structured outlines and then expands them into dense introduction content. This baseline represents explicit multi-stage orchestration. SurveyForge. Same as AutoSurvey, SurveyForge is also adapted as a workflow-based baseline. STIG. STIG (Zhang, su, and He 2025) uses the same stage- token format as StructPO, but it is trained only with supervised fine-tuning. It does not use Struct-aware Relative Advantage estimation, global-local reward fusion or refinement-guided optimization. This baseline isolates the effect of stage-token supervision without reinforcement learning. Refinement without Training. This baseline performs inference-time refinement using the SFT model. The model first generates an initial draft and then revises it with the refinement prompt in Appendix . Unlike StructPO, the revision behavior is not distilled into the policy during training, so the method requires an additional inference pass. Closed-source LLMs. We evaluate GPT-4o and GPT-5.1 through API access. They are prompted with the same struc- tured input materials used by StructPO, including the title, abstract, figure information, table information and baseline- reference information. These models serve as strong closed- source references for semantic quality, writing fluency and factual consistency. Token Accounting Protocol For the token-efficiency analysis in Section Analysis, we separate inference-time token consumption into overhead ConfigurationValue Training frameworkVerl Rollout / inference backendSGLang Hardware8 H800 GPUs Global batch size16 Rollout size8 SFT data size∼300 papers RL data size∼1,550 papers Validation data size∼150 papers Global advantage weight λ0.3 Local advantage weight 1− λ0.7 Section length bandwidth6 Global length bandwidth12 KL regularizationEnabled Table 6: Main training configurations for StructPO. tokens and effective tokens. Overhead tokens include input context, system prompts, formatting instructions, intermediate outlines, intermediate drafts, revision instructions and other workflow states that are consumed during generation but are not part of the final delivered introduction. Effective tokens refer to the final introduction content presented to the user after removing structural markers and intermediate workflow artifacts. For workflow-based systems such as AutoSurvey, token usage is accumulated across all intermediate calls. Therefore, repeated context replay and intermediate outputs are counted each time they are consumed by the workflow. For single- pass systems such as StructPO and STIG, the token count is computed from the single inference call. The effective rate is defined as: Effective Rate = Effective Tokens Effective Tokens + Overhead Tokens . (14) This protocol allows us to compare multi-stage workflows and single-pass policies under the same accounting framework. Qualitative Test Case To complement the brief discussion in the main text, we provide a more detailed qualitative diagnosis of the three competing paradigms on a representative ACL 2025 test paper, SpeechFake (Huang et al. 2025). Figure 5 visualizes the generated introductions side by side. AutoSurvey: cross-stage drift and meta-commentary. Be- cause AutoSurvey decomposes generation into independent agent calls, the resulting text exhibits clear cross-stage drift. As highlighted in blue, the model prematurely reveals the proposed dataset (“the SpeechFake dataset was developed...”) at the end of the very first background paragraph, breaking the rhetorical suspense expected in an academic introduc- tion. In the contribution section, the agent further halluci- nates a markdown-style subtitle and inserts unnatural meta- commentary, treating each stage as a self-contained blog post rather than a coherent component of a unified introduction. These artifacts illustrate how external multi-stage orches- tration weakens the macroscopic narrative when individual agents lose access to the document-level context. STIG: length collapse and shallow synthesis. STIG suc- cessfully partitions the output into a four-paragraph skeleton through stage-token supervision, but suffers from the well- known length-collapse failure mode of pure SFT. The full introduction contains only 318 words, which is markedly insufficient for a top-tier conference paper. The second para- graph repeatedly uses the word “limited”, indicating shallow synthesis, and the final contribution bullets are essentially copy-pasted from the preceding method paragraph rather than abstracted at a higher level. This confirms that behavioral cloning alone, even with explicit stage tokens, is unable to encourage rich and well-elaborated academic content. StructPO: balanced structure and dense content. In con- trast, StructPO produces a balanced introduction that follows the background–problem–method–contribution progression with explicit rhetorical anchors (e.g., “With the rapid advance- ment...”, “However...”, “In this paper...”, “In summary...”), as marked in bold. Unlike AutoSurvey, StructPO avoids cross- stage drift; unlike STIG, it generates dense, paper-specific content, referencing concrete artifacts such as Table 1 and Figure 2, and reporting specific evaluation metrics including equal error rate (EER) and cross-speaker results. The contri- bution section is condensed into three well-formatted bullet points that synthesize, rather than duplicate, the preceding paragraphs. This case suggests that struct-aware policy opti- mization can internalize the workflow-style organization of writing while preserving coherent single-pass generation. Human Evaluation Details We provide additional details of the blind human evaluation. We randomly sampleN = 30papers from the ACL test set. For each paper, we present the source abstract and two generated introductions toE = 3human evaluators with NLP research experience. The two introductions are generated by StructPO and GPT-5.1, respectively. Model identities are hidden from the evaluators and the presentation order is randomized to reduce position bias. Evaluators are instructed to choose the better introduction based on a holistic assessment of logical coherence, structural completeness and academic writing quality. For each paperi ∈ 1,...,Nand evaluatorj ∈ 1,...,E, we define a binary preference variablev (M) i,j ∈ 0, 1, wherev (M) i,j = 1if evaluatorjprefers model M ∈StructPO, GPT-5.1for paperiand0otherwise. We compute two metrics. Win Votes measures the overall vote share across all annotations: WinVotes (M) = 1 N · E N X i=1 E X j=1 v (M) i,j × 100%. (15) Win Rate measures the proportion of papers for which a model wins by majority vote: WinRate (M) = 1 N N X i=1 I   E X j=1 v (M) i,j > E 2   × 100%. (16) Paper ID Human 1 Human 2 Human 3 Winner 1StructPO GPTStructPO StructPO 104GPTStructPO GPTGPT 106StructPO GPTGPTGPT 109StructPO StructPO StructPO StructPO 110GPTGPTStructPO GPT 1002GPTGPTStructPO GPT 1005GPTGPTStructPO GPT 1007GPTStructPO GPTGPT 1009GPTStructPO StructPO StructPO 1020StructPO StructPO GPTStructPO 1025GPTGPTGPTGPT 1029StructPO GPTStructPO StructPO 1032GPTGPTStructPO GPT 1033StructPO GPTStructPO StructPO 1035GPTStructPO StructPO StructPO 1036StructPO StructPO GPTStructPO 1037StructPO GPTStructPO StructPO 1044StructPO GPTGPTGPT 1045StructPO GPTStructPO StructPO 1047StructPO StructPO StructPO StructPO 1058StructPO GPTGPTGPT 1068GPTStructPO StructPO StructPO 1069StructPO StructPO GPTStructPO 1072StructPO StructPO StructPO StructPO 1081GPTStructPO GPTGPT 1089StructPO StructPO StructPO StructPO 1107StructPO GPTGPTGPT 1122GPTGPTStructPO GPT 1123StructPO StructPO StructPO StructPO 1124StructPO GPTGPTGPT Table 7: Raw human-annotation data for 30 sampled papers. Table 7 reports the raw annotation results for all 30 sampled papers. “StructPO” and “GPT” indicate the model preferred by each annotator and the final winner is determined by majority vote. Prompt for Structural Decomposition The following prompt is used to decompose ACL introduc- tions into section-level content and outlines. 1 Please break down the introduction section of the following academic paper into a structured outline format for academic discussion purposes. 2 The content should be divided into the following four sections , extracting key points for each: 3 4 1. Background: Basic background and significance of the research field 5 - Number of points: 2-4 6 2. Problem and Limitations of Existing Methods: Current issues , challenges and limitations of existing methods 7 - Number of points: 2-6 8 3. Brief Method Overview and Summary of Main Results: Overview of the proposed method , main experimental results and findings 9 - Number of points: 4-8 10 4. Our Contributions: Main contributions and innovations of the paper 11 - Number of points: 2-3 12 13 Please output in the following JSON format , including outline points and paragraphs classified by section: 14 15 16 "sections ": 17 "Background ": "Combine all paragraphs and sentences belonging to the background section", 18 "Problem and Limitations of Existing Methods ": "Combine all paragraphs and sentences belonging to the problems and limitations section", 19 "Brief Method Overview and Summary of Main Results ": " Combine all paragraphs and sentences belonging to the method overview and main results sections", 20 "Our Contributions ": "Combine all paragraphs and sentences belonging to the contributions section" 21 , 22 "outline ": 23 "Background ": [" Point 1", "Point 2", "..."] , 24 "Problem and Limitations of Existing Methods ": [" Point 1", "Point 2", "..."] , 25 "Brief Method Overview and Summary of Main Results ": [" Point 1", "Point 2", "..."] , 26 "Our Contributions ": [" Point 1", "Point 2", "..."] 27 28 29 30 Introduction content: 31 text 32 33 Notes: 34 1. Analyze the content coherently and categorize it into the appropriate sections , strictly controlling the number of points for each section. 35 2. When assigning sections , ensure continuity; for example , Background must be at the beginning of the article and if there is an Our Contributions section , it must be at the end. There should be no section 1, section 2, section 1 sequences. 36 3. Some papers may not have a section similar to Our Contributions; if so , generate an empty Our Contributions field. 37 4. First , divide the sections , then perform an outline analysis to identify key points. 38 5. Do not use demonstrative pronouns like "this" or "the model" in the key points; use specific names if available. Inference and Refinement Prompts The following prompts are used for inference and refinement. 1 As an academic writing expert who has completed a research project and is currently in the paper -writing stage , please draft the introduction section based on the available materials. 2 Follow the format of Outline to Content , first drafting the outline of this section and then the content. 3 When writing , ensure logical coherence and smooth transitions and use fluent and standard academic English. 4 5 Style and Content Requirements: 6 - Maintain a formal academic tone. 7 - Be as coherent and concise as possible and directly related to the title and abstract. 8 - Use transitional phrases effectively. 9 10 Citation Instructions: 11 - Do not mention any citations. For example , "( Smith et al.)". 12 - Do not use reference formats such as \ ref. 13 14 Compose the Introduction of an ACL paper based on the corresponding research materials. For each sub -section , first list the outline and then write the corresponding content of that section. You need to write four sections: 15 1. Background: Provide the research background and the current status of the field. (The content text is about 90 words) 16 2. Problem and Limitations of Existing Methods: Describe the research problem and the limitations of existing methods. (The content text is about 180 words) 17 3. Brief Method Overview and Summary of Main Results: Briefly introduce the proposed method and summarize the main results. (The content text is about 230 words) 18 4. Our Contributions: Summarize the contributions of this paper. (The content text is about 70 words) 19 20 Please write introduction with the following writing format and use < STAGE > and <END > markers to represent writing stages: 21 <STAGE0 > Outline for Background: 22 (Outline for Background) <END0 > 23 <STAGE1 > Contents for Background: 24 (Contents for Background) <END1 > 25 <STAGE2 > Outline for Problem and Limitations of Existing Methods: 26 (Outline for Problem and Limitations of Existing Methods) <END2 > 27 <STAGE3 > Contents for Problem and Limitations of Existing Methods: 28 (Contents for Problem and Limitations of Existing Methods) <END3 > 29 <STAGE4 > Outline for Brief Method Overview and Summary of Main Results: 30 (Outline for Brief Method Overview and Summary of Main Results) <END4 > 31 <STAGE5 > Contents for Brief Method Overview and Summary of Main Results: 32 (Contents for Brief Method Overview and Summary of Main Results) <END5 > 33 <STAGE6 > Outline for Our Contributions: 34 (Outline for Our Contributions) <END6 > 35 <STAGE7 > Contents for Our Contributions: 36 (Contents for Our Contributions) <END7 > 37 38 Research materials: 39 Title: data['title'] 40 Abstract: data['abstract'] 41 Figure information: data['figures'] 42 Table information: data['tables'] 43 Baseline references: data['ref'] 1 Please carefully review your writing according to revision suggestions and make revisions to improve logical coherence , clarity of expression and adherence to academic standards. Follow the original writing format and provide the revised content directly. Evaluation Prompts for AWQ and SFC The following prompts are used for AWQ and SFC evaluation. 1 system_prompt = """ You are a senior academic writing expert evaluating the writing quality of a research paper Introduction , particularly for CS/NLP venues (e.g., ACL , EMNLP , NAACL). 2 3 CRITICAL INSTRUCTION: 4 - Evaluate ONLY the writing quality , structure , and academic tone. 5 - COMPLETELY IGNORE all citations and references. 6 - Do NOT consider whether citations exist , are correct , or are properly formatted. 7 - Treat the text as if all citation markers (e.g., "[1]" , "(Author , Year) ") do not exist. 8 - Focus purely on: vocabulary , sentence structure , logical flow , content balance , and professional expression .""" 9 10 user_prompt = f""" Generated Introduction : 11 generated_introduction 12 13 ### Evaluation Criteria (IGNORE ALL CITATIONS) 14 15 1. Academic Vocabulary & Formal Tone 16 - Uses precise , formal terminology appropriate for scholarly writing 17 - Avoids colloquial expressions , slang , or overly casual language 18 - Employs appropriate hedging (e.g., "may suggest", "appears to", " potentially ") 19 - Uses domain -specific technical terms correctly 20 - Avoids vague expressions (e.g., " many studies", "significant improvement" without specifics) 21 - GOOD: Using concrete numbers and specific comparisons (e.g., "10 times larger", "15% improvement ") 22 - BAD: Empty modifiers without support (e.g., "pioneering", " comprehensive", "thorough" without evidence) 23 24 2. Logical Structure & Argumentation 25 - Clear problem statement and research motivation 26 - Logical progression: Background - Research Gap - Proposed Solution - Contributions 27 - Well -organized paragraphs with clear topic sentences 28 - Smooth transitions between ideas and sections 29 - Coherent narrative that guides the reader 30 31 CRITICAL - Paragraph Balance: 32 - Paragraphs should have UNEQUAL lengths reflecting their importance 33 - Core contributions/methods should be MORE prominent (longer/more detailed) 34 - Background should be CONCISE (not a lengthy tutorial) 35 - RED FLAG: All paragraphs having similar length (~ equal words) indicates poor prioritization 36 37 What to PENALIZE (max 2 points): 38 - Mechanical section preview: " Section 3 presents ..., Section 4 describes ..., Section 5 shows ..." 39 - Method/experiment paragraphs as task lists: "We do A. We then do B . We further do C. Additionally , we do D." 40 - Repetitive content across paragraphs (e.g., same limitation stated twice) 41 42 What is ACCEPTABLE (especially in CS/ NLP): 43 - Contributions listed with bullet points or numbered items (e.g., " First , ... Second , ...") 44 - "Our contributions are as follows: (1) ... (2) ... (3) ..." 45 - This is STANDARD practice in ACL , EMNLP , NAACL papers - do NOT penalize 46 47 3. Content Proportion & Emphasis 48 - Background: Should be CONCISE ( ideally 15 -25% of introduction), NOT a textbook tutorial 49 - Research Gap: Clearly and specifically articulated with concrete examples or evidence 50 - Proposed Solution: Should be prominent , with specific details ( numbers , comparisons) 51 - Contributions: Should be SPECIFIC and CONCRETE , not vague claims 52 53 Balance Guidelines: 54 - Background > 40% - likely too long 55 - Contributions < 15% - likely too brief or weak 56 - All paragraphs about 10% of each other - poor structure (lack of emphasis) 57 58 Good Contributions Example: 59 "We introduce X, comprising 3M samples across 46 languages -10 times larger than prior benchmarks ." 60 61 Bad Contributions Example: 62 "We introduce X, a pioneering and comprehensive dataset for the research community ." 63 64 4. Sentence Quality & Readability 65 - Varied sentence structures (not repetitive patterns) 66 - Appropriate sentence length (not too long or choppy) 67 - Clear and unambiguous expressions 68 - Proper grammar and syntax 69 - No awkward phrasing or unclear constructions 70 - PENALIZE: Repetitive sentence patterns (e.g., "as shown in Figure X" used 3+ times) 71 - PENALIZE: Consecutive "We + verb" sentences in non -contribution sections 72 73 5. Introduction -Specific Requirements 74 - Opening: Engaging first paragraph that establishes importance and relevance 75 - BAD Opening: "With the rapid advancement of X..." / "In recent years , X has attracted ..." 76 - GOOD Opening: Start with a concrete problem , striking fact , or specific context 77 - Closing: Clear statement of contributions (bullet points acceptable in CS/NLP) 78 - Self -contained: Reader understands the paper's purpose without needing other sections 79 - References to Tables/Figures: Good introductions often reference Table 1 or Figure 1 for support 80 81 6. Professional Academic Style 82 - Appropriate use of active/passive voice 83 - Objective and impersonal tone where appropriate 84 - Concise writing without unnecessary redundancy 85 - PENALIZE: Redundant phrases like " contributions can be summarized as follows" (just say "Our contributions :") 86 87 ### Common Defects Checklist 88 - [ ] Excessive background (>40% of introduction) 89 - [ ] All paragraphs roughly equal length (no emphasis) 90 - [ ] Vague or missing problem statement 91 - [ ] Generic contributions without specific numbers/findings 92 - [ ] "Firstly/Secondly/Thirdly/Lastly" mechanical listing (worse than bullet points) 93 - [ ] Repetitive content between paragraphs 94 - [ ] Method paragraph as task list ("We do A. We then do B...") 95 - [ ] Normal opening ("With the rapid advancement ...") 96 - [ ] No reference to Tables/Figures for evidence 97 - [ ] Empty modifiers (" comprehensive", "thorough", "pioneering ") without support 98 99 ### Scoring Guide (1-5) 100 - 5: Excellent. Publication -ready for top venues. Well -balanced content with clear emphasis on contributions , engaging opening , logical flow , specific claims with evidence , appropriate length. Contributions are concrete with numbers/findings. 101 - 4: Good. Solid academic writing suitable for submission. Minor issues : slightly unbalanced content , some repetitive patterns , or contributions could be more specific. References Tables/Figures appropriately. 102 - 3: Acceptable. Recognizably academic but has noticeable problems: somewhat imbalanced paragraphs , normal opening , vague contributions , or some awkward sentences. Needs revision before submission. 103 - 2: Poor. Significant issues: clearly imbalanced (e.g., 50% background), mechanical section listing , task -list method paragraphs , repetitive content , empty claims. Needs substantial revision. 104 - 1: Unacceptable. Severely flawed: no clear structure , mechanical listing throughout , disorganized , or unprofessional writing. 105 106 ### Output Format 107 " vocabulary ": "<brief comment >", " structure ": "<brief comment , note paragraph balance >", "readability ": "<brief comment >", "overall ": "< overall assessment with main issues >", "score ": <integer 1 -5 >""" 1 system_prompt = """ You are a lenient academic reviewer checking for hallucinations. 2 3 ONE STRICT RULE: Fabricated citations are unacceptable (Score 1-2). 4 5 EVERYTHING ELSE: Be lenient. If content can be reasonably inferred from source materials or is common academic knowledge , it is NOT hallucination. 6 7 Academic introductions naturally expand beyond abstracts. This is expected , not hallucination. 8 9 Respond with JSON only .""" 10 11 user_prompt = f""" Source Abstract: 12 abstract 13 14 Source Table Information: 15 table_info 16 17 Source Figure Captions: 18 figure_info 19 20 Generated Introduction: 21 generated_introduction 22 23 ### Evaluation Criteria 24 25 STRICT: Fake Citations (Zero Tolerance) 26 - Any citation "(Author , Year)", "[1]" , etc. MUST exist in sources 27 - Fake citation found - Score 1-2 28 29 LENIENT: Everything Else 30 Accept as valid if content: 31 - Can be inferred or derived from source materials 32 - Is general background or domain knowledge 33 - Is reasonable elaboration or explanation 34 - Is logical implication of what sources describe 35 - Uses approximate numbers (e.g., "about 3M" = "3 ,000 ,000") 36 37 Only flag as hallucination if: 38 - Directly contradicts source materials 39 - Completely fabricates results/methods with no basis in sources 40 41 ### Scoring Guide (1-5) 42 43 - 5: Good - No fake citations. Content aligns with or is inferable from sources. 44 45 - 4: Fine - No fake citations. Reasonable expansions. Perhaps minor liberties but acceptable. 46 47 - 3: Borderline - No fake citations. Some content seems loosely connected to sources but not contradictory. 48 49 - 2: Problematic - Contains fake citation(s), OR directly contradicts sources. 50 51 - 1: Severe - Multiple fake citations , OR fabricates core content entirely. 52 53 Default Assumption: If no fake citations and no obvious contradictions - Score 4-5 54 55 ### Output Format 56 " fake_citations ": "<citations NOT in sources , or'None'>", " factuality_issues ": "<only obvious contradictions/fabrications , or'None '>", "score ": <integer 1 -5 >""" AutoSurvey The rise of speech deepfakes represents a significant security and ethical challenge in the digital age, .... These deepfakes, .... Recent breakthroughs in text-to- speech synthesis, voice conversion, and neural vocoder techniques have enabled the .... This growing threat underscores the urgent need for robust detection systems and comprehensive datasets to train and evaluate such models. In response, the SpeechFake dataset was developed as a large-scale, multilingual resource encompassing over 3 million deepfake samples across 46 languages. By incorporating diverse generation methods, SpeechFake provides a benchmark for advancing detection capabilities and understanding the factors influencing model performance in real- world scenarios. ... (Two Paragraphs) Advancing Speech Deepfake Detection: Contributions and Future Directions This section highlights the key contributions of this work, centered on the development of SpeechFake, a large-scale, multilingual speech deepfake dataset. By incorporating over 3 million deepfake samples—.... These findings underscore the importance of diverse and representative data in training robust deepfake detection models. Looking ahead, future research should focus on improving model adaptability .... SpeechFake serves as a foundational resource to advance the field, fostering innovation in AI ethics and secure speech technologies. (Total 749 words) With the development of deep learning, speech generation technology has advanced rapidly. However, the progress has also led to the rise of deepfake speech, which poses a significant threat to security, privacy, and trust. Deepfake speech can be used to deceive voice authentication systems and may cause substantial harm to individuals, organizations, and even the public. Speech deepfake datasets are crucial for training detection models. However, existing datasets are limited in size and diversity. First, the representation of speech generation methods in current datasets is limited. For example, current datasets primarily focus on text-to-speech (TTS) and voice conversion (VC) techniques. However, with the rapid development of neural vocoder, more advanced generation methods are emerging, such as neural vocoder (NV) and data-driven synthetic voice (DDSV). Thus, existing datasets may not adequately represent modern generation methods. Second, the diversity of generation techniques and languages in current datasets is limited. The types of generation techniques and languages are limited, and there is a lack of in-the-wild scenarios, such as speaking in different environments, using different devices, or incorporating emotions, which can affect the generalizability of detection models. Data diversity and generality are important for robust detection models. To address these issues, we introduce SpeechFake, a large-scale multilingual speech deepfake dataset. SpeechFake is built using 40 advanced speech synthesis tools covering a wide range of generation methods and 46 languages, and it includes 3 million deepfake samples and 3,000 hours of audio. In addition to providing bilingual test sets for robust detection, it also includes multilingual test sets covering 46 languages. SpeechFake covers a variety of generation techniques, including text-to-speech (TTS), voice conversion (VC), and neural vocoder (NV). Our contributions can be summarized as follows: • SpeechFake is the largest speech deepfake dataset to date, covering a wide range of generation methods and 46 languages. • SpeechFake includes bilingual and multilingual test sets for evaluating detection models and studying deepfake speech characteristics. (Total 318 words) STIG StructPO With the rapid advancement of speech generation technology, synthetic speech has reached a level of sophistication that makes it increasingly difficult to distinguish from real speech. Speech deepfake detection has become an important research field, with numerous studies devoted to identifying synthetic speech. Speech deepfake detection has significant practical importance, as it can help detect and mitigate risks of deepfake audio abuse, such as deepfake election interference, deepfake customer service, and other applications involving voice manipulation. However, the development of robust detection models is still hindered by several challenges. Many existing speech deepfake datasets, including recently released datasets, are limited in scale and diversity, as shown in Table 1. As a result, it is difficult to generate sufficient and diverse deepfake samples with distinct characteristics, making it hard to train detection models that can generalize well to unseen deepfakes. Another limitation is that existing datasets often only contain deepfake samples generated using one or two specific generation methods, which limits the generalization capability of the detection models. For example, in the ASVspoof2019-LA dataset, only text-to-speech (TTS) and voice conversion (VC) methods are used, while recent studies often only focus on TTS methods. Moreover, there is a lack of large-scale multilingual datasets, making it difficult to train models for detecting multilingual speech deepfakes, especially for languages other than English. Finally, there is a lack of studies that comprehensively examine the effects of generation methods, language diversity, and speaker variation on detection performance. In this paper, we introduce SpeechFake, a large-scale multilingual speech deepfake dataset specifically designed for speech deepfake detection, as shown in Figure 1. SpeechFake consists of over 3 million deepfake samples, totaling more than 3,000 hours of audio. It contains speech generated by 40 different speech synthesis tools, covering a wide range of speech generation methods, including text- to-speech (TTS), voice conversion (VC), and neural vocoder (NV) methods, as shown in Figure 2. We provide detailed statistics and analysis of SpeechFake, including the distribution of generation methods, speaker genders, languages, and durations. We then conduct baseline experiments to evaluate detection models trained on SpeechFake, providing EER results on both SpeechFake’s test sets and other public benchmark datasets. Additionally, we provide a detailed analysis of the detection performance, including how different generation methods, language diversity, and speaker variation affect the detection accuracy of speech deepfake detection models. We further conduct cross-speaker experiments to assess the generalization capability of detection models trained on SpeechFake and conduct additional experiments to explore the effect of generation methods on detection performance. In summary, our contributions can be summarized as follows: • Introduction of SpeechFake, a large-scale multilingual speech deepfake dataset, which is the largest speech deepfake dataset to date. • Comprehensive analysis of the impact of different generation methods, language diversity, and speaker variation on detection performance. • Evaluation of speech deepfake detection models trained on SpeechFake, demonstrating its potential for detecting speech deepfakes on multiple test sets. (Total 483 words) Figure 5: Qualitative comparison among AutoSurvey, STIG and StructPO on a real ACL test case (Huang et al. 2025). Blue marks structurally inappropriate wording in AutoSurvey, the bottom annotation reports the insufficient total word count of STIG, and bold transitional anchors highlight the rhetorical bridges produced by StructPO.

Instant research discovery