Skip to content

โœ๏ธ Text Generation

๐Ÿ’ฌ EMNLP2026 ยท 15 paper notes

๐Ÿ“Œ Same area in other venues: ๐Ÿ”ฌ ICLR2026 (12) ยท ๐Ÿ’ฌ ACL2026 (17) ยท ๐Ÿงช ICML2026 (2) ยท ๐Ÿค– AAAI2026 (3) ยท ๐Ÿ“น ICCV2025 (1) ยท ๐Ÿงช ICML2025 (1)

๐Ÿ”ฅ Top topics: Diffusion Models ร—5 ยท Summarization ร—3 ยท Personalized Generation ร—3 ยท Reasoning ร—2

ArgSum-X: A Structure-Aware Benchmark for Argument Summarization and Key Point Analysis

Addressing the long-standing fragmentation between argumentative text summarization (ATS) and key point analysis (KPA) along with the absence of structural fidelity metrics, ArgSum-X presents the first unified multi-stage benchmark and structure-aware evaluation framework, revealing that while general language models achieve high semantic faithfulness, they suffer from substantial deficits in preserving argumentative flow and global reasoning topology.

Beyond a Single Story: Meta-Reviewing Sparse and Incomplete User-Generated Content

Addressing both macro review sparsity and micro attribute incompleteness in user-generated content, MOSAIC adapts academic meta-review principles to aggregate attribute-sentiment evidence from neighbor users, combining an MMoE multi-task backbone with personalized attention to deliver refined rating predictions and attribute-level explanations.

Beyond Prefix Answering: Parallel Infilling Reading for Long Contexts with Diffusion Language Models

Addressing the underutilization of native bidirectional infilling in diffusion language models caused by adopting autoregressive prefix-answering paradigms, this paper proposes Parallel Infilling Reading (PIR) to insert extraction slots between context chunks for parallel evidence denoising and subsequent answer reduction, while resolving cross-slot attention interference through slot-isolation masking and single-active-slot fine-tuning.

curp codebook based continuous user representation for personalized generation w

CURP proposes a plug-and-play personalization framework based on a discrete prototype codebook and a dual-aggregator, factorizing user interaction history into orthogonal subspace prototypes to capture query-adaptive dynamic preferences and stable intrinsic traits, achieving state-of-the-art personalized generation while training only 28M parameters (0.36% of the base LLM).

Evaluating Style-Personalized Text Generation: Challenges and Directions

This paper introduces a comprehensive benchmark spanning eight writing domains and three evaluation settings to scrutinize style discrimination metrics, empirically demonstrating that ensembles of diverse evaluation metrics consistently outperform individual evaluators.

Evo: Autoregressiveโ€“Diffusion Large Language Models with Evolving Balance

This paper introduces Evo, a duality latent flow model that unifies autoregressive next-token prediction and diffusion denoising into a continuous probability trajectory, adaptively balancing global planning and local refinement via learned per-token progression variables while preserving near-AR decoding throughput (52 tokens/s).

FAVoR: Measuring and Mitigating Author-Style Homogenization in Federated Personalized Generation

Addressing author-style homogenization in federated personalized text continuationโ€”where central parameter aggregation erodes individual stylistic nuance toward a shared registerโ€”FAVoR proposes a decoupled shared-private PEFT residual mechanism that isolates local author-specific style adjustments outside aggregation, substantially improving cross-author style discriminability while maintaining continuation utility.

POLARIS: Guiding Small Models to Write Long Stories

POLARIS introduces a compute-efficient creative-writing RL recipe that integrates a structured 16-dimension LLM-as-judge Story Quality rubric with teacher-forced human-reference injection (HRI) inside GRPO groups, allowing a 9B model trained on 4 A100 GPUs for 48 hours to preserve narrative coherence and close length adherence across stories up to 12,000 words.

Provenance Before Prose: Claim-Locked Reporting for Statistical Text Generation

Addressing numerical drift, direction inversions, and rhetorical overstatements when LLMs draft statistical reports for functional neuroimaging (fMRI) and randomized controlled trials (RCTs), this paper introduces Claim-Locked Reporting, a protocol that locks reportable claimsโ€”including evidence sources, exact numbers, effect directions, and permitted language strengthsโ€”in a typed ledger before generation, achieving 98.5%โ€“100.0% cross-seed reproducibility while eliminating harmful statistical fabrications.

SGTR: Learning Style-Generalizable Trajectory Reasoning for Zero-Shot Text Style Transfer

SGTR formulates text style transfer as the joint modeling of structured reasoning trajectories and final stylized outputs, decoupling the process into three interpretable stages alongside latent contrastive regularization to achieve superior zero-shot generalization on unseen styles.

Summarization Is Not Dead Yet

Challenging the widespread narrative that large language models have rendered summarization research obsolete, this paper introduces a rigorous multi-track evaluation across 5 diverse domains and 5 frontier LLMs, demonstrating that human summaries maintain decisive advantages in informativeness and faithfulness while LLM outputs are preferred primarily for surface-level fluency.

Towards Efficient Reasoning: Learning Causal Shortcuts for Diffusion Language Models

To alleviate the exponentially expanded exploration space and error accumulation in diffusion language models under bidirectional attention, this paper proposes the Causal Shortcut Learning (CSL) framework, which extracts global reasoning-guiding token chains via conditional mutual information and applies parallel prioritized masking during supervised fine-tuning, significantly accelerating decoding convergence and improving math and code reasoning accuracy.

Training-Free Character-Length Control in Summarization with Diffusion Language Models

This paper proposes DELTA (Diffusion Expected Length Targeting Algorithm), an inference-time training-free method for masked diffusion language models (such as LLaDA and Dream) that dynamically inserts and deletes mask tokens based on continuous expected character-length estimation, reducing character-length error by an order of magnitude across XSum, CNN/DailyMail, and arXiv while maintaining competitive summary quality.

What Are We Measuring in NLG? A Meta-Analysis of Evaluation Trends (2020-2025)

By developing a multi-LLM extraction and verification pipeline over 14,171 papers across four major ACL conferences from 2020 to 2025, this paper presents the largest meta-analysis of NLG evaluation practices to date, exposing pervasive metric inertia, severe metric-criteria decoupling, and a critical validation gap in LLM-as-a-judge adoption.

When i.i.d. Factorization Breaks Language: The Structural Failure of Discrete Diffusion Language Models

This paper demonstrates that discrete diffusion language models (DDLMs) suffer from severe degeneration and repetition because their standard i.i.d. corruption and factorized sampling assume token exchangeability, which fundamentally conflicts with natural language's structured dependencies and causes marginal collapse under semantic sparsity; it formalizes the Dependency-Coupling Principle (DCP) to guarantee stable parallel generation.