The paper introduces ColRel, a two-stage approach that uses large language models to discover relationships between columns in data lakes. It first creates column embeddings from available metadata and data at ingestion, then refines these embeddings with business dictionaries to generate concise natural-language descriptions. Experiments on public benchmarks and an industrial ERP dataset demonstrate ColRel’s effectiveness, especially in scenarios with weak signals and semantically related columns.
By Ahlame Diouan (ERIC, UL2), Eric Ferey (ERIC, UL2), Sabine Loudcher (ERIC, UL2), J\'er\^ome Darmont (ERIC, UL2)
The paper introduces GROUND, a framework that limits large language model (LLM) analytics to a governed semantic layer for enterprise data warehouses. GROUND supplies approved metrics, dimensions, join paths, filters, and security rules, then validates generated SQL against these constraints before execution, retrying or abstaining on violations. In benchmarks, GROUND eliminates hallucinations across all evaluated categories and prevents row‑level security breaches, outperforming schema‑only, schema‑RAG, and semantic‑only approaches.
By Aravind Sasidharan Pillai
The paper introduces the Relational Hypergraph Transformer (RHT), a unified architecture that models relational databases as hypergraphs and learns pentadimensional embeddings (PentE). RHT applies sparse relational attention whose complexity scales with the average relational degree, making it computationally efficient for large, high‑dimensional, and high‑cardinality datasets. Experiments on the Synthea synthetic electronic health record dataset show that RHT produces more semantically coherent embeddings than tabular, relational, and temporal graph baselines, while remaining scalable, and the authors provide an open‑source implementation and plan clinical validation on MIMIC‑IV.
By Edouard Lansiaux, Hugo Kazzi, Aur\'elien Loison, Slim Hammadi, Emmanuel Chazard
The paper introduces the Multi-Instruction Multi-Shot Long-Video Editing (MMLVE) task, aiming to enable consistent editing of long videos with multiple instructions. It proposes an agentic editing framework that combines Large Language Models and Vision-Language Models for shot-level decoupling and precise instruction parsing. The authors also present the MMLVE-Bench dataset and evaluation metrics, showing that their MMLVE-Agent outperforms existing state‑of‑the‑art methods by eliminating hallucinations and preserving temporal consistency.
By Chenyang Wu, Fuchen Long, Binyuan Huang, Xinlong Sun, Xi Chen, Chun-Le Guo, Chongyi Li
The paper introduces BIND, a framework that binds a human’s biometric data to an AI agent’s identity and task-specific authority, enabling secure, real‑time delegation of control. By generating a token that an AI agent presents to an Identity Auditor, the system performs biometric authentication and recovers the agent’s ID and scope, providing non‑repudiable proof of human oversight. A practical implementation using face features and a fuzzy commitment scheme with turbo error‑correcting codes achieves a 96% true match rate at zero false match rate and supports 1024‑bit agent tokens.
By Joseph Geo Benjamin, Anil K Jain, Karthik Nandakumar
The paper introduces a decoupled, I/O-aware pipeline for extracting embeddings from large whole-slide images (WSIs). It splits the workflow into patch generation, parallel embedding inference, and sharded vector database ingestion, isolating data movement from compute to reduce I/O overhead. The resulting distributed vector database stores embeddings with rich metadata, enabling efficient filtering, retrieval, and downstream tasks such as classification and few-shot learning, especially in low-resource settings.
By Mayanka Chandrashekar, Xi Zhang, Ethan Seefried, Tirthankar Ghosal, John Gounley, Heidi Hanson
Vagdhenu is a Sanskrit shloka‑to‑chant text‑to‑speech system that preserves meter (vrutta) and phonological nuances. It builds on an off‑the‑shelf flow‑matching backbone and a large‑scale neural vocoder, adding a Kannada‑based frontend to avoid schwa deletion, a phonology‑aware frontend handling visarga sandhi and sibilant distinctions, and a vrutta‑aware reference selection mechanism. The authors report that a text‑side prosody conditioner is ineffective in their architecture, while reference clips and voice‑steering retraining provide the necessary prosody control, and they demonstrate the system’s performance on a 32‑chapter video corpus and an audio app covering 18,000 verses.
whyItMatters":"The system delivers high‑fidelity, meter‑aware Sanskrit chanting, enabling large‑scale deployment of authentic recitations for educational and cultural preservation purposes."
By Prathosh A P
The paper presents a method for cross‑architecture knowledge distillation from a fine‑tuned DINOv2 Vision Transformer teacher to a lightweight bidirectional Visual State Space Model (LVSSM) student for tea leaf disease classification. By addressing training‑stability issues with a progressive convolutional stem and gated selective‑scan block, the 4.45 M‑parameter student achieves a mean test accuracy of 95.41%—a 3.09‑point improvement over the teacher’s 92.32%—while using only one‑fifth of the teacher’s parameters. Ablation studies show that simple logit‑level distillation outperforms intermediate feature alignment, and the gains are specific to students that start below the teacher’s performance.
By Zibo Zhou, Zongsen Qiu, Rui Chen, Yujie Yao, Yue Zhou, Jianjun Wang
The paper investigates how to effectively pre‑train language models when the data budget is limited but compute is plentiful. It shows that increasing model size only improves performance up to an optimal point, after which overfitting degrades generalization, and that this optimal size varies with both the data budget and downstream tasks. To overcome the inefficiencies of standard Transformers in this regime, the authors propose recursive Transformers that reuse a shared block across depth and employ factorized embeddings, achieving better results than standard models on 10M–100M word pre‑training budgets and competitive performance with BabyLM Challenge 2025 winners.
By Serdar G\"ulbahar, Lukas Edman, Alexander Fraser
The paper introduces Successive Capacity Growth (SCG), a method for adaptively expanding Vision Transformer encoders in Joint-Embedding Predictive Architectures (JEPAs). SCG starts with a minimal encoder and incrementally increases width or depth based on a task‑agnostic test‑and‑verify mechanism, while a Sketched Isotropic Gaussian Regularizer (SIGReg) keeps learned semantic dimensions independent. Experiments on multi‑object dynamics and 2D navigation tasks show that SCG achieves up to 20.3% better prediction loss than fixed small baselines and 23% better than fixed large models, with far greater parameter efficiency and no false‑positive expansions.
By Frederik Berenz
The paper introduces Composed Gait Retrieval (CoGR), a task that retrieves a target gait sequence using a reference sequence and a natural language modification query. To support this, the authors create the first gait-language datasets—Language‑Augmented CCPG and Language‑Augmented CASIA‑B—via an automated annotation pipeline powered by large vision‑language models. They propose ComposeGait, an identity‑anchored composition framework with a Part‑aware Identity Adapter that injects identity tokens into a shared Q‑Former, achieving state‑of‑the‑art retrieval performance on both benchmarks.
By Jingchen Fei, Zengbin Wang, Yukun Liu, Muyi Sun, Shibiao Xu, Man Zhang
TopiCLEAR is a framework that clusters document or sentence embeddings using adaptive dimensionality reduction to uncover low‑dimensional geometric structures that correspond to human‑interpretable topics. The method is evaluated on four benchmark datasets, showing strong agreement with human annotations, especially for short and informal texts. A Twitter case study demonstrates that TopiCLEAR yields more interpretable topics than LDA, recovering both annotated topic structure and coherent sub‑topics.
By Aoi Fujita, Taichi Yamamoto, Yuri Nakayama, Ryota Kobayashi
The study compares an English-only and a bilingual decoder-only model, each 310 M parameters, trained on eight diverse languages while controlling for English exposure, compute, and document overlap. After aligning on shared English vocabulary, the authors find that token embeddings appear similar, but the deeper hidden states used for prediction diverge across models. This hidden‑state mismatch grows through middle transformer layers and persists despite controls, indicating that contextual processing differs between the models.
"whyItMatters":"The findings show that embedding alignment can conceal significant internal representation differences, which is crucial for any downstream work that assumes aligned multilingual models are interchangeable."
By Anjishnu Mukherjee, Ziwei Zhu, Antonios Anastasopoulos
The paper introduces LitEm, a neural regression model that allows transductive knowledge graph embedding models to predict numerical attributes. LitEm achieves top or near‑top performance on most attributes across datasets such as FB15K‑237, YAGO15K, DB15K, and Mutagenesis. A co‑training framework further improves link prediction for bilinear models while enabling them to predict numerical attributes, demonstrating literal‑aware encoding of attribute information.
By Rupesh Sapkota, Louis Mozart Kamdem Teyou, Moshood Yekini, Caglar Demir, Axel-Cyrille Ngonga Ngomo
The paper demonstrates that a self‑supervised Vision Transformer (ViT) pretrained on a fast, low‑cost semi‑numerical simulator can produce data summaries that transfer across different simulators without retraining. In 21cm cosmology, the ViT—named SKATR—pretrained on 67,000 21cmFAST lightcones is applied unchanged to hydrodynamical Loreli II lightcones, enabling accurate inference of five astrophysical parameters with fewer radiative‑transfer simulations than a fully‑supervised baseline. SKATR remains accurate, informative, and calibrated even under realistic SKA antenna array noise, outperforming supervised models retrained on noisy data.
By Yannic Pietschke, Caroline Heneka, Ayodele Ore, Romain Meriot
KinyaEmbed is the first sentence‑embedding model specifically designed for Kinyarwanda, built on KinyaBERT‑large and trained through a four‑stage curriculum that incorporates paraphrase pairs, translated MNLI triplets, OPUS‑100 translation pairs, and high‑quality KinyaCOMET pairs. It outperforms existing multilingual embeddings on the SemRel2024‑rw benchmark, achieving a Spearman ψ of 0.7298, and introduces the Wiki‑RW‑STS benchmark of 300 contamination‑free Kinyarwanda sentence pairs. All model checkpoints, filtered pairs, and the new benchmark are publicly released.
By Ireddi Rakshitha, Devavarapu Yashwanth, Ntakirutimana Pierre
The paper presents a scalable graph neural network (GNN) system for friend recommendation on a massive social graph. It introduces two key design choices: multi-hash ID embeddings that shrink the embedding table by over 98% without hurting ranking quality, and a timestamp-sorted compressed sparse row (CSR) storage with binary search that reduces temporal neighbor sampling from linear to logarithmic time. Experiments on a 194‑million‑user, 28‑billion‑edge graph show that these techniques enable production‑grade performance, boosting friend additions by 16% and unique friend adders by 11.5% in an online A/B test.
By Maksim Utushkin, Andrei Ovsiannikov, Alexander D'yakonov
The pro-team at LLMs4OL 2026 presented a system for ontology learning that tackles both the End-to-End Flagship Task (Task A) and the Ontology Extension Reuse Task (Task B). Their approach uses an offline retrieval‑augmented few‑shot prompting pipeline with Qwen2.5‑14B‑Instruct and MiniLM‑L6‑v2 for retrieval, selecting top‑5 examples for Task A and top‑2 for Task B, and applies a left‑truncated context‑windowing strategy to keep task instructions in long prompts. For Task B, generated triples are filtered deterministically by a vocabulary constraint, keeping triples that involve at least one term from the closed vocabulary and removing duplicates of the initial ontology, achieving high scores in Semantic Graph Similarity, Term‑Typing F1, and Taxonomy Discovery F1, though no non‑taxonomic relations were extracted.
By Shivam Mishra, Dhannu Ram Meena, Muneendra Ojha, Krishna Pratap Singh, Kuldeep Singh
The study investigates how large language models (LLMs) handle diverse Indian oral traditions, using the Rajasthani Pabuji epic, Tamil Sangam poetry, and Bengali folk tales as case studies. By prompting Claude Sonnet and Gemini with 54 generation requests across generic, culturally specific, and regional-language prompts, the authors measured reference drift and cross-tradition convergence using Sentence‑BERT embeddings. Results show that while outputs stay closer to their own tradition than to others, there is significant cross‑tradition similarity (0.52–0.66), indicating partial homogenisation; moreover, regional‑language prompting consistently reduced fidelity to authentic traditions.
By Paarth Singh Rathore
MIMO: Multilingual Information Retrieval via Monolingual Objectives proposes a two‑stage framework that first aligns a student model to a stable English semantic space using knowledge distillation, then jointly optimizes distillation and cross‑lingual contrastive learning to improve retrieval discrimination while preserving alignment. The approach addresses language clustering and the trade‑off between cross‑lingual alignment and embedding uniformity, outperforming existing cross‑lingual training baselines on both multilingual and multi‑monolingual benchmarks. MIMO also remains competitive with larger off‑the‑shelf models and its alignment‑uniformity analysis clarifies the distinct roles of the two loss components.
whyItMatters":"The study demonstrates a practical method to enhance multilingual information retrieval performance by balancing alignment and uniformity, which is crucial for real‑world search environments where queries and documents span multiple languages."
By Youngjoon Jang, Seongtae Hong, Heuiseok Lim