Drift Inspector is an open‑source system that extracts Atomic Contribution Claims (ACCs) from scientific abstracts using an LLM, then clusters these claims over time to map how a research field evolves. Applied to six years of EMNLP, the tool reveals a shift from classic NLP tasks toward LLM‑era capabilities such as reasoning and multimodality—trends that keyword or whole‑abstract counts miss. The pipeline has also processed the entire ACL Anthology, yielding 346,000 claims from 80,000 abstracts across 423 venues, with human‑validated extraction and clustering aligned to an external taxonomy.
By Vsevolod Karimov, Stepan Ostarkov, Anastasia Poroshina, Anatoly Frolov, Alexander Panchenko
PixelUMM is an encoder‑free model that unifies image and video understanding and generation directly in pixel space. It represents images as spatial patches and videos as spatiotemporal tubelets, feeding both through single‑layer linear projections into a shared multimodal backbone. The Mixture‑of‑Transformers architecture blends shared attention with task‑specific parameters, enabling autoregressive text prediction, pixel‑space flow matching, and clean‑pixel video generation, and experiments show competitive performance across tasks while providing design insights for future pixel‑space multimodal models.
By Cong Wei, Xuanchi Ren, Bryan Chu, Weiming Ren, Huan Ling, Jiahui Huang, Laura Leal-Taix\'e, Sanja Fidler, Wenhu Chen, Zian Wang, Jay Zhangjie Wu
The paper introduces Harness Primitives—reusable agent harness mechanisms mined from failed task trajectories—and a framework called STITCH that selects and composes these primitives into task‑specific harnesses at test time. This approach avoids generating or debugging harness code for each task, achieving up to 12‑point gains in task success over fixed harness baselines and outperforming human‑designed harnesses like Codex CLI. STITCH also demonstrates minimal test‑time overhead (2.7%) and scales efficiently with the size of the primitive library.
By Peng Kuang, Haibo Jin, Dehao Wu, Feiyang Deng, Xiaopeng Yuan, Jerry Wang, Haohan Wang
arXiv:2609.39049v1 Announce Type: cross
Abstract: A large language model (LLM) can rate depression severity directly from a social media post or mark which clinical criteria the post shows and let co...
By Xinkai Chen
arXiv:2609.38201v1 Announce Type: new
Abstract: Long-running tools can dominate coding-agent latency: compilers, test suites, and repository commands take seconds to minutes while the agent idles. Th...
By Jiangnan Yu, Ceyu Xu, Mengming Li, Shiyu Huang, Yiran Xia, Jian Weng, Hui Xue, Haohui Mai, Yuan Xie
arXiv:2609.38256v1 Announce Type: new
Abstract: Large language models (LLMs) are increasingly used to answer questions about politically contentious issues, yet evaluations typically treat a model's...
By Olivia Macmillan-Scott, Michael Jacobs, Nils Metternich, Mirco Musolesi
arXiv:2609.34330v2 Announce Type: replace
Abstract: Multimodal large language models (MLLMs) have demonstrated impressive performance in multimodal understanding, but processing large numbers of visu...
By Tinghao Wang, Yichen Guo, Qizhe Zhang, Yuan Zhang, Weimin Ouyang, Rui Huang, Jiajun Cao, Sixiang Chen, Hao Jiang, Jixian Wu, Zheng Lu, Bofan Zhu, Renyuan Li, Shanghang Zhang
arXiv:2609.38630v1 Announce Type: new
Abstract: Privacy redaction must remove personal information while preserving relationships expressed in text. We develop a multilingual named-entity tagger with...
By Jonathan Graehl
The paper introduces NFA-LM, a polynomial‑time method for generating language model outputs that satisfy nondeterministic finite automaton (NFA) constraints. It leverages recent results that the #NFA counting problem admits a fully polynomial randomized approximation scheme, providing theoretical guarantees under mild assumptions. Experiments demonstrate that NFA‑LM produces high‑quality outputs efficiently while keeping approximation error bounded.
By Jialiang Sun, Kuldeep Meel
HalluScoring 2026 is a shared task that evaluates hallucination detection and factual verification in Arabic question answering, focusing on generalization to unseen questions and LLMs. It comprises two main tasks with four subtasks: binary hallucination detection (Subtasks 1.1 and 1.2) and answer verification against six candidates in Islamic and general knowledge domains (Subtasks 2.1 and 2.2). Thirteen teams participated, with the top system achieving AUC‑ROC scores of 0.772 and 0.767 for detection, and 0.882 and 0.857 for verification.
By Aisha Alansari, Abdessalam Bouchekif, Ahmed Hasanaath, Salah Eddine Bekhouche, Malak Alkhorasani, Mohammed-En-Nadhir Zighem, Saad Ezzini, Hichem Telli, Hend Al-Khalifa, Muhammad Abdul-Mageed, Hadid Abdenour, Hamzah Luqman
ViLegalExpert is a large-scale Vietnamese legal benchmark built from real citizen–lawyer consultations, comprising over 172,000 questions across 34 legal domains with professional answers and expert-verified evidence. It supports legal information retrieval, extractive QA, and abstractive QA. Experiments show that while pretrained language models perform well on QA, hybrid retrieval methods achieve the best evidence retrieval, highlighting significant challenges in grounding legal answers to authoritative sources.
By Dat Tien Nguyen, Nghia Hieu Nguyen, Anh Thi-Hoang Nguyen, Dung Ha Nguyen, Kiet Van Nguyen, Ngan Luu-Thuy Nguyen
The paper introduces the QuantCode Model, a specialized language model for generating executable algorithmic trading code. It combines continued pretraining on trading framework code with supervised fine‑tuning on validated request‑to‑code pairs, achieving significant gains on the QuantCode‑Bench benchmark and improving both single‑turn and multi‑turn agentic success rates. The study also highlights the importance of retaining parsing and tool‑calling capabilities during domain specialization.
By Alexey Chernysh, Orkhan Ekhtibarov, Dmitry Zmitrovich
The paper investigates how synthetic data generated by large language models (LLMs) can be characterized using sample-level learnability derived from encoder training dynamics. It compares different LLM families and scales across tasks such as single- and multi-label classification, labeling, and tree prediction, and contrasts these synthetic datasets with human-written data. The study also examines the robustness of learned data distributions across encoders and evaluates how data selection strategies based on learnability signals impact the performance of both synthetic and organic data.
By Irene Lago, Ana Ezquerro, David Vilares
MetaSteer is a new method for steering large language models that learns nonlinear, context-dependent interventions applied to attention projection matrices. Unlike traditional linear, context-independent techniques, MetaSteer adapts its effects based on the input, requiring no linear concept-geometry assumption. Trained once on a pooled preference corpus, it transfers zero‑shot to unseen concepts and out‑of‑distribution contexts, matching or surpassing strong task‑specific baselines on multiple benchmarks and model families.
By Mehdi Jafari, Hao Xue, Flora Salim
The paper introduces NAMOH, a native sparse attention mechanism that activates only a subset of heads per token, allowing each head to attend to a limited subsequence of tokens. By scaling the number of heads while keeping the active heads per token fixed, the method shortens head histories and reduces key‑value access without increasing overall storage. Experiments demonstrate that NAMOH can outperform fully activated models with the same parameter count and enable more efficient long‑context inference than smaller dense models.
By Zizhuo Fu, Runsheng Wang, Meng Li
The paper investigates Training‑Inference Mismatch (TIM) in large‑language‑model reinforcement learning, where rollout generation and policy optimization produce differing token probabilities despite identical model weights. By creating a zero‑mismatch diagnostic setting called VeXact, the authors isolate TIM and demonstrate that even minor token‑level numerical disagreements can trigger training collapse. They further show that TIM alters the effective optimization problem and propose remedies to mitigate its impact.
By Tianle Zhong, Neiwen Ling, Yifan Pi, Zijun Wei, Tianshu Yu, Geoffrey Fox, Peng Wu, Xiao Yu
Cogentic is a multi‑agent system designed to automate proof discovery for open research problems. It uses an iterative prove‑verify loop where an orchestrator assigns independent provers to different proof directions, verifies their outputs with specialized components, and records confirmed intermediate results in a persistent ledger for future rounds. Built on Gemini, Cogentic has produced novel results on five open problems in online learning, auction theory, and mechanism design, each verified by domain experts and detailed in companion papers.
By Yang Cai, Vineet Gupta, Yanchen Jiang, Christopher Liaw, Aranyak Mehta, Grigoris Velegkas, Di Wang
The article examines how classical test theory statistics—Kuder‑Richardson coefficient, dependability index, and Livingston‑Lewis accuracy—can mislead when applied to large language model (LLM) judges that are evaluated with a single prompt and no gold labels. Using Claude Haiku 4.5 on 210 short‑answer items, the authors show that these metrics fail to isolate the judge’s performance because the judge’s single administration provides no variance component. They argue that reliable statements about an LLM judge require gold labels or varied scorer facets, and that bank design heavily influences reliability estimates.
By Louis Yiven Zhu
The paper reports that fine‑tuning large language models on aligned data can unintentionally cause misaligned responses in other contexts—a phenomenon termed *context confusion*. The authors demonstrate this effect in gender equality, privacy, and physical safety domains, showing that it differs from emergent misalignment and is not mitigated by general alignment data but can be reduced with domain‑specific alignment or in‑context examples. They provide a mechanistic explanation based on representational shifts during fine‑tuning that lead to behavioral feature transfer across contexts.
By Yavuz Bakman, Duygu Nur Yaldiz, Baris Askin, Swastik Roy, Morteza Ziyadi, Salman Avestimehr, Sai Praneeth Karimireddy
The paper introduces a multidimensional framework for classifying human‑AI interactions (HAIIs) in clinical trials, categorising interactions by AI tasks, human‑AI relationships, interaction configurations, and interacting human groups. It evaluates the framework by applying it to 15 trials, with categorisation performed by both human reviewers and large language model classifiers. The study demonstrates that LLMs can assist in categorisation while highlighting the continued need for human judgment when trial records are incomplete or ambiguous.
By Sandra Woolley, Tim Collins, Khalid Khattak, Illia Chernomorets, Ariane Arevalo, Chris Richardson