token-level alignment

Existing CLIP-style graph–text alignment methods suffer from two key limitations: (1) they enforce rigid one-to-one mapping assumptions, ignoring intrinsic many-to-many semantic relationships in graphs; and (2) they rely on static alignment objectives, yielding poor robustness under noisy supervision. This paper proposes ADAligner—the first dynamic supervision-aware graph–text alignment framework. ADAligner adaptively selects either subgraph-level many-to-many alignment (for high-quality data) or node-level one-to-one alignment (under noise) via batch-level alignment reliability estimation and dynamic filtering of low-confidence samples. It introduces a soft subgraph alignment loss and a tunable optimization objective, with theoretical guarantees establishing it as a stable negative-feedback system. Evaluated on nine text-attributed graph datasets, ADAligner achieves significant gains in zero-shot/few-shot classification, link prediction, and cross-modal retrieval, accelerates training by 2–3×, and demonstrates exceptional robustness to label noise.

Current text-to-image generation models struggle to achieve smooth transitions between semantically similar prompts due to substantial differences in token sequences—particularly in wording, ordering, and conceptual positioning—which hinders effective image blending and continuous editing. This work proposes a Token-to-Token Alignment framework that, without modifying the underlying model, employs a two-stage strategy: first aligning the semantic structures of prompts and then aligning their token embedding representations. By reconstructing diverse prompts into a shared structural form, the method reveals that the latent continuous semantic structure within the text embedding space can be effectively leveraged through representation alignment. Consequently, linear interpolation in this aligned space yields coherent semantic transitions, significantly enhancing the quality of image semantic mixing and continuous editing.

Traditional metrics—such as character overlap or distributional similarity—fail to characterize cross-lingual knowledge transfer capability between language pairs with highly divergent writing systems. Method: This paper introduces *token alignability*, a novel subword-level metric quantifying the degree to which subword tokens across languages can be reliably aligned. We formally define and empirically validate token alignability as a core predictor of multilingual tokenization quality and cross-lingual transfer performance. Our methodology integrates subword-level alignment modeling, cross-lingual embedding space analysis, controlled encoder-decoder architecture comparisons, and ablation studies varying training data scale. Results: Experiments demonstrate that token alignability significantly outperforms conventional metrics in predicting cross-lingual performance on non-overlapping script pairs (e.g., Chinese–English, Japanese–German). It enhances interpretability and practicality in tokenizer design and language-pair selection. The codebase and full reproducibility package are publicly released.

This work addresses the low efficiency of key information identification and semantic retrieval in text. We first discover that large language model (LLM) text embeddings naturally align with salient input tokens in the latent space—a phenomenon empirically validated across diverse model architectures, training paradigms, and embedding methods. Leveraging this insight, we propose a principal-component-guided alignment enhancement method that decomposes the embedding geometry and explicitly steers representations toward critical tokens. Building upon this, we design a lightweight sparse retrieval paradigm that retains only ~20% of token embeddings while achieving 80% of dense retrieval performance. Experiments across eight mainstream LLM embedders confirm the robustness of the alignment mechanism. Our findings provide an interpretable foundation for sparse retrieval and instruction-tuned embeddings, and advance the understanding of the intrinsic nature of semantic relevance.

Existing image-text alignment benchmarks rely on rule-based perturbations or short captions, failing to assess fine-grained semantic alignment. Method: We introduce the first fine-grained evaluation benchmark built on synthetically generated image-text pairs from diverse generative models (both image-to-text and text-to-image), along with the first fine-grained alignment metric specifically designed for synthetic data. Contribution/Results: Through systematic human annotation of sentence-level correctness and comprehensive evaluation, we uncover pervasive deficiencies in mainstream models—including CLIP—such as fine-grained perceptual blind spots, spatial localization errors, and generator-induced self-biases. Our benchmark significantly improves discriminative power for evaluating local semantic alignment in vision-language models, establishing a new paradigm and reproducible standard for fine-grained alignment modeling and assessment.

This work addresses severe parsing errors in dense document pages caused by unstable layout assumptions, which lead to mismatches between detector outputs and the input sequence expected by the parser. To resolve this, the authors introduce a lightweight structural refinement module between a DETR-style detector and the parser, performing set-level reasoning using query features, semantic cues, bounding box geometry, and visual evidence. This module jointly decides which instances to retain, refines bounding boxes, and predicts the correct parsing order. By integrating retention-oriented supervision with a difficulty-aware ordering objective, the method significantly enhances layout–parsing interface consistency under complex layouts, reducing the reading order edit distance to 0.024 on OmniDocBench while consistently improving page-level layout quality.

This work proposes a novel approach to address the challenge novice modelers often face in ensuring semantic alignment between domain models and textual specifications during early software engineering phases. The method first employs natural language processing to preprocess specification texts and generates human-authored natural language descriptions for each model element. It then leverages a large language model (LLM) to compare these descriptions against the original specifications, automatically classifying their alignment status as aligned, misaligned, or uncertain, while providing interpretable evidence for each judgment. By uniquely integrating LLM capabilities with human-crafted model descriptions, the approach achieves high-precision semantic alignment verification, demonstrating near-perfect precision (≈100%) and 78% recall across multiple domain datasets. Individual element analysis requires between 18 seconds and one minute, indicating strong potential for integration into modeling tools.

This work proposes a novel metric for evaluating the morphological plausibility of subword segmentations without relying on scarce and inconsistent gold-standard word segmentation data. Instead, it leverages widely available morphosyntactic resources—such as Universal Dependencies and UniMorph—and employs IBM Model 1 to probabilistically align subword units with morphological features. The resulting alignment scores serve as a proxy for morphological well-formedness. The proposed method demonstrates strong correlation with traditional morpheme boundary recall across diverse languages and substantially improves applicability in morphologically complex or low-resource settings, thereby enabling a cross-lingually generalizable framework for morphological evaluation of subword segmentations.

This work addresses the scarcity of large-scale, high-quality sentence-aligned corpora for text simplification in non-English languages by systematically constructing and publicly releasing a multilingual simplification corpus covering Catalan, English, French, Italian, and Spanish. Leveraging crowdsourcing, the authors collect simplified texts from comparable documents and implement a document-to-sentence alignment mechanism to produce high-quality sentence pairs. This resource fills a critical gap in non-English simplification data and provides a foundational benchmark for training and evaluating multilingual text simplification systems.

This study investigates the self-iterative refinement mechanisms of document-level large language models (LLMs) in literary translation, where such processes remain poorly understood. The authors systematically evaluate nine LLMs across seven language pairs, examining various multi-granularity translation-and-refinement configurations and prompting strategies through large-scale human assessment and model strength perturbation experiments. Findings indicate that document-level translation followed by segment-level refinement yields robust performance. Generic prompting consistently outperforms error-focused or evaluate-then-refine approaches. Refinement primarily enhances fluency, stylistic coherence, and terminological consistency, with limited gains in fidelity. Moreover, the refinement process tends to align outputs with the optimizer’s own distribution rather than accurately correcting errors.