Explore methods for understanding and generating human language. Read research on linguistic tasks, multilingual systems, and evaluation.
26 papers · Latest edition October 2, 2026
Three of the latest briefs in this collection. Read the evidence and the original papers alongside them.
AdaTutoRank teaches rerankers to assemble complementary evidence sets rather than merely picking individually relevant documents, improving RAG and deep-research retrieval with fewer calls.
The paper argues that sentence meaning equivalence is not reliably stored in separate embeddings but is computed when models process both sentences together.
SlopShape detects and identifies AI-written commercial content by recognizing its underlying organizational style, even after the text has been reworded.
Newest editions first.
Kailin Jiang, Lei Liu, Jian Xi, Yangqi Chen, Hui Xu, Hongwei Zhao, Bin Li, Yu Lu, Haibo Shi
AdaTutoRank teaches rerankers to assemble complementary evidence sets rather than merely picking individually relevant documents, improving RAG and deep-research retrieval with fewer calls.
Jiaqi Deng
The paper argues that sentence meaning equivalence is not reliably stored in separate embeddings but is computed when models process both sentences together.
Jochen Madler
SlopShape detects and identifies AI-written commercial content by recognizing its underlying organizational style, even after the text has been reworded.
Ej Zhou, Suchir Salhan, Catherine Arnett, Anna Korhonen
Independently trained language models may spontaneously learn representations that can be rotated and transferred across languages, suggesting multilingual capabilities without joint training.
Ahmetcan Yavuz, Clara Meister, Tiago Pimentel
This study shows that how a tokeniser searches may matter more for language-model efficiency than what objective it optimises.
Ivan Silajev
This paper shows that two classic search-ranking methods, TF-IDF and BM25, can be understood precisely as measuring KL divergence between probability models.
Raphaël Sourty, Antoine Chaffin, Paulo Roberto Moura Junior, Amélie Chatelain
This paper openly trains strong dense and late-interaction search models and shows that token-level matching can improve multilingual and unseen-script retrieval.
David Lowry-Duda, Matteo Cargnelutti, Catherine Brobston, Salwa Ismail, Greg Leppert, Amanda Watson, Jonathan Zittrain
This work turns nearly a million scanned books into a flexible, multilingual research resource by preserving rich annotations instead of making one irreversible cleaning decision.
Ayoub Kirouane, Christos Petrocheilos
The study shows that SFT teaches models to reason visibly in Greek, while reinforcement learning fixes formatting and leakage problems that accuracy benchmarks fail to detect.
Hyangsuk Min, Hwanjun Song
ReMEMBER helps dialogue summarizers retrieve the missing past context needed to understand and summarize the present conversation.
Lena Holzwarth, Rita González-Márquez, Dmitry Kobak
A large-scale study estimates that most biomedical papers now contain signs of LLM-assisted writing, especially in their Discussion sections.
Yasir Ech-Chammakhy, Oussama Azrara, Jaafar Chbili, Anas Motii
STINER turns noisy X posts into structured cyber-threat intelligence, showing that domain-specific models can extract strategic signals faster and more accurately than general-purpose or LLM-based systems.
Jie Zhu, Huaixia Dou, Shuo Jiang, Junhui Li, Lifan Guo, Feng Chen, Chi Zhang, Fang Kong
ESC-Skills turns emotional support chat into a set of explicit, self-improving skills so the assistant can respond more safely, consistently, and helpfully.
Renhao Pei, Yihong Liu, Sampo Pyysalo, Hinrich Schütze, Shaoxiong Ji
This paper shows that giving LLMs step-by-step grammatical reasoning can boost translation quality for very low-resource languages, especially when the reasoning is used at inference time rather than learned from scratch.
Liang Xue, Haoyu Liu, Cheng Wang, Pengyu Chen, Haozhuo Zheng, Yang Liu
STREAM turns live streams and short videos into a large task-oriented dialogue dataset, aiming to fill the gap in realistic domain conversations for training and evaluating LLMs.
Delip Rao, Weiqiu You, Eric Wong, Chris Callison-Burch
This paper turns millions of NSF award abstracts into a new dataset for extracting scientific claims and research plans, aiming to power better claim verification and science analysis.
Jinheon Baek, Soyeong Jeong, Sangwoo Park, Woongyeong Yeo, Minki Kang, Patara Trirat, Heejun Lee, Sung Ju Hwang
OmniRetrieval aims to be a universal translator for search, sending each query to the right mix of text, tables, and graph databases instead of forcing everything into one format.
Zhixin Cai, Jun Bai, Yang Liu, Jiaqi Li, Yichi Zhang, Taichuan Li, Zhuofan Chen, Zixia Jia, Zilong Zheng, Wenge Rong
Xetrieval explains why dense retrievers rank documents highly by turning opaque embeddings into sparse human-readable features and using them to trace retrieval decisions.
M. Ali Bayram, Banu Diri, Savaş Yıldırım
This paper shows how to cheaply adapt a multilingual embedding model to Turkish by redesigning its tokenizer and distilling it offline, achieving strong retrieval-quality gains at low training cost.
Maurice Kraus, Ruben Härle, Sebastian Sztwiertnia, Abbas Goher Khan, Mehdi Ali, Michael Fromm, Kristian Kersting
This paper builds a large, carefully translated German pretraining corpus and shows that better curated data can noticeably improve German language models.
Claire Hobbs, R. Thomas McCoy
This paper argues that children may learn subject-verb agreement by exploiting word-pair statistics in their input, and shows neural networks can learn the same way under the right data conditions.
Zhaomin Wu, Jiayi Li, Bingsheng He
EmoTrack uses LLM-derived clinical signals plus structured transcript embeddings to track depression severity across counseling sessions more robustly, especially when longitudinal context is available.
Yu-Chen Den, Yung-Yu Shih, Zhi Rui Tam, Kuan-Yu Chen, Pu-Jen Cheng, Yun-Nung Chen, Eugene Yang
ICICLE lets generative retrievers add new documents on the fly by using in-context document IDs instead of retraining the model each time.
Huanchi Wang, Zihang Huang, Yifang Tian, Kristina Dzeparoska, Hans-Arno Jacobsen, Alberto Leon-Garcia
FAME uses an LLM only once offline to help build a lightweight on-premise expert system that pinpoints anomalous log messages with far less labeling effort.
Tong Wang, Yiqing Xu, Leo Yang Yang
This paper turns text-model interpretability into an auditability test: a feature should be both understandable to humans and genuinely separate from the prediction label.
Jan Tempus, Philip Whittington, Craig W. Schmidt, Dennis Komm, Tiago Pimentel
This paper turns tokenization from a greedy heuristic into a convex optimization problem, producing a tokenizer that can be near-optimal and often slightly better for language models.