ClusterFusion is a hybrid text‑clustering framework that uses a large language model (LLM) as the core clustering engine, guided by lightweight embedding methods. The process involves embedding‑guided subset partitioning, LLM‑driven topic summarization, and LLM‑based topic assignment, allowing direct incorporation of domain knowledge and user preferences. Experiments on three public benchmarks and two new domain‑specific datasets show that ClusterFusion achieves state‑of‑the‑art performance on standard tasks and significant gains in specialized domains, with the authors releasing a new dataset and benchmark results.
By Yiming Xu, Yuan Yuan, Vijay Viswanathan, Graham Neubig
CanLegalRAGBench is a Canadian legal QA benchmark that uses realistic queries and expert‑annotated answers grounded in case law to evaluate retrieval‑augmented generation (RAG) systems. The study finds that retrieval performance varies with design choices and that open‑source embedding models can compete with closed‑source ones. It also highlights limitations of automatic evaluations, noting that RAG systems often produce hallucinations or overly detailed responses, with 8‑29% of claims unsupported by retrieved documents.
By Ethan Zhao, Maksym Taranukhin, Wei Cui, Moira Aikenhead, Vered Shwartz
The paper investigates a new vulnerability called memory injection, where malicious prompts can persist across sessions in modern LLM‑as‑a‑service platforms. By demonstrating that all NIST attack goals can be achieved on ChatGPT Instant 5.3, Claude Sonnet 4.6, and Gemini 3 Flash, the authors show that these services can spread disinformation, generate harmful content, or cause offline unresponsiveness even in fresh sessions. The study also reveals that many proposed defenses are not yet implemented and provides datasets and artifacts for further research.
By Pedro H. Barcha Correia, Matheus Azevedo de S\'a, Paulo H. Gomes de Senna, Giovanni Mandel Martignago, Victor Takashi Hayashi, Charles Christian Miers, Marcos A. Simplicio Jr
The paper introduces post‑hoc kernel‑variance heads that estimate epistemic uncertainty from a single vision‑language‑action policy’s embeddings, eliminating the need for multiple policies. These heads, which compute the conditional variance of a Gaussian process, match ensemble performance on LIBERO‑10 in task‑level success correlation, active learning demonstration acquisition, and supervisor referral, while cutting uncertainty‑scoring time by up to 50×. The study also shows that controlled perturbations reveal coverage gaps and that modality‑specific heads can detect unfamiliar inputs, improving within‑task referral when paired with a supervisor proxy.
By Johannes Keller, Albrecht M. Blo{\ss}e, Adrian Lindenmeyer, Thomas Neumuth, Daniel Schneider
ThreadRAG is a retrieval‑augmented generation framework that preserves contextual consistency by structuring retrieved documents into Semantic Nodes and building a Semantic Topology Graph (STG). It uses Energy‑Driven Pathfinding (EDP) to identify query‑relevant subgraphs and linearize them into coherent textual evidence paths. Experiments on HotpotQA, 2WikiMultihop, and another benchmark show ThreadRAG outperforms strong baselines, achieving a 14.4% EM gain on HotpotQA and a 5.8% EM lead on 2WikiMultihop while cutting token cost to 23.3% of the strongest baseline.
By Yufeng Li, Zaijing Li, Fuchen Zhang, Zongzong Wu, Ming Zhao, Fengxiao Tang
PRISM is a training‑free framework for financial information retrieval that combines prompt engineering, in‑context learning, and a candidate refinement module, with optional multi‑agent coordination. On FiQA‑2018 it improves Recall@100 from 0.7783 to 0.8326, and on FinAgentBench it achieves a Cohere‑refined top‑50 pool that raises NDCG@5 while cutting per‑query cost by about 35%. The authors also release an expert‑annotated evaluation set, a label‑robustness sweep, and a Pareto frontier of quality‑cost‑latency trade‑offs, and PRISM ranks third on the private FinAgentBench leaderboard without fine‑tuning.
By Chun Chet Ng, Jia Yu Lim, Zhen Hao Chu, Yixi Zhou, Wei Zeng Low
The paper presents a retrieval‑augmented generation (RAG) system combined with prompt engineering for Japanese‑to‑Chinese translation, focusing on sentences with noun‑modifying clause constructions (NMCCs). It integrates linguistic analysis, embedding‑based retrieval, and prompt construction before generating translations with GPT‑4o, without altering the base model. Experiments on a 66‑sentence test set show that increasing the knowledge‑base size from 0 to 2,000 examples raises macro‑averaged BLEU scores from 24.28 to 29.96, indicating improved translation quality for NMCC‑rich sentences.
By Wenshi Gu
CALLIOPE is a new top‑down framework designed to uncover topic hierarchies in large unstructured datasets, such as LLM conversations, while ensuring differential privacy. Unlike conventional methods that depend on clustering vector embeddings, CALLIOPE directly employs LLM capabilities throughout its pipeline. Experiments show that CALLIOPE delivers higher utility than embedding‑based approaches, particularly in high‑privacy scenarios, and its outputs are evaluated using both standard clustering metrics and newly introduced utility metrics.
By Edith Cohen, Vadym Doroshenko, Badih Ghazi, Charlie Harrison, Pritish Kamath, Alexander Knop, Ravi Kumar, Ethan Leeman, Pasin Manurangsi, Mikhail Pravilov, Adam Sealfon, Yurii Sushko, Marika Swanberg, Chiyuan Zhang
The paper introduces Signals and Queries (S&Q), a lightweight first-stage relevance filter for LLM agents that removes at least 30% of candidate policies before downstream evaluation while maintaining high recall. S&Q uses offline-generated activating examples and hard negatives to train a logistic classifier that predicts policy relevance via embeddings and vector dot products, eliminating per-turn LLM calls. Evaluated on 9,144 relevance labels across 254 conversation prefixes and 36 airline policies, S&Q outperforms top‑K and cosine‑threshold retrieval in precision at every recall level, consistently across two embedding providers.
By Yam Marcovitz
The paper investigates how information-theoretic and geometric compression unfolds over time in recurrent neural networks, specifically discretized LSTMs. It finds that compression is strongest at the final timesteps of many datasets and attributes this to three mechanisms: class separability, temporal stationarity of shared dynamics, and limited representational capacity. Through synthetic and real-world experiments, the authors demonstrate that these factors jointly determine the extent of temporal compression and its role as a regularizer when excess capacity exists, but can harm generalization when it consumes task-relevant memory.
By Nikolaus Kopp, Franz Pernkopf, Bernhard Geiger
The report introduces pplx-embed-v2-late, a pair of multilingual and multimodal late‑interaction models sized at 0.6B and 9B parameters. Both models achieve state‑of‑the‑art performance on text, image, and visual document retrieval benchmarks, with the 0.6B model matching larger models on visual document tasks. They are trained via representation distillation from an 18B teacher, enabling a shared embedding space that allows documents to be indexed with the 9B model and queried with the low‑latency 0.6B model.
By Markus Krimmel, Louis Milliken, Ekaterina Glazkova, Maximilian Schall, Antoine Chaffin, Sedigheh Eslami, Jonas Golde, Denis Bykov, Claire Beamer, Jeremy Yang, Bo Wang
Masked language modelling (MLM) remains the dominant pretraining objective for transformer encoders on longitudinal electronic health records (EHRs), but reconstructing individual events can overemphasise noisy observations and fails to encourage learning higher‑level patient states. Joint‑Embedding Predictive Architectures (JEPAs), which predict representations in latent space, were tested on MIMIC‑IV data and, while training stably, consistently underperformed standard MLM. The study found that larger target sizes increase prediction error, yet reducing target size does not improve downstream performance, indicating that stable latent prediction alone is insufficient for competitive EHR representation learning and underscoring the difficulty of defining clinically meaningful latent targets for EHR.
By Ashvin Gupta, Brendan Delaney, Alessandra Russo
arXiv:2610.29302v1 Announce Type: cross
Abstract: Language models are being explored as the interface through which space robots interpret operator commands and select among verified skills. Running...
By Jia Deng, Zizhe Wang, Xin Lu, Yi Huang, Zehao Jia
arXiv:2610.28957v1 Announce Type: new
Abstract: Scanning every passage in a long document seems thorough, but does it preserve more useful evidence or produce better answers? We compare 11 retrieval...
By Juan Terven, Diana-Margarita C\'ordova-Esparza, Julio-Alejandro Romero-Gonz\'alez, Edgar Chavez-Urbiola, Juan Irving Vasquez, Julian Andres Ramirez-Bautista, Silvia L. Chaparro-C\'ardenas, Gendry Alfonso-Francia
arXiv:2610.29792v1 Announce Type: new
Abstract: Retrieval-Augmented Generation (RAG) improves the performance of Large Language Models (LLMs) by retrieving external knowledge. Recently, some RAG meth...
By Xinze Li, Dingwen Zhang, Zhenghao Liu, Yukun Yan, Shuo Wang, Yu Gu, Ge Yu, Maosong Sun
arXiv:2604.00706v3 Announce Type: replace
Abstract: Assessing the veracity of a claim made online is a complex and important task with real-world implications. When these claims are directed at commu...
By Israel Abebe Azime, Jesujoba Oluwadara Alabi, Crystina Zhang, Iffat Maab, Atnafu Lambebo Tonja, Tadesse Destaw Belay, Folasade Peace Alabi, Salomey Osei, Saminu Mohammad Aliyu, Nkechinyere Faith Aguobi, Bontu Fufa Balcha, Blessing Kudzaishe Sibanda, Davis David, Mouhamadane Mboup, Daud Abolade, Neo Putini, Philipp Slusallek, David Ifeoluwa Adelani, Dietrich Klakow
arXiv:2610.22671v2 Announce Type: replace
Abstract: As assistants move from reactive answering toward proactive intervention, memory changes role: it no longer only supplies the answer to a query, it...
By Shenglai Zeng, Xiao Yang, Tiffany Ma, Xinyuan Zhang, Nanshu Wang, Deren Lei, Kai Sun, Xilun Chen, Tammy Stark, Xianhui Zhu, Yuetai Li, Rulin Shao, Tianxin Wei, Rakesh Wanga, Anuj Kumar, Jiliang Tang, Wen-tau Yih, Xin Luna Dong
arXiv:2610.29223v1 Announce Type: new
Abstract: Retrieval-augmented generation can expose unwanted concepts through retrieved evidence even when the generator is moderated. We study query-side concep...
By Yiyang Duan, Fan Wu, Tiantong Wu, Che Wang, Zhiren Gong, Wei Guo, Yang Cao, Wei Yang Bryan Lim
arXiv:2610.29830v1 Announce Type: new
Abstract: Deep learning models for chest X-ray anomaly detection remain vulnerable to subtle distributional shifts (e.g., acquisition technique, patient-related...
By Samuel Halimi, Loic Themyr, Arnaud Abreu
arXiv:2610.29881v1 Announce Type: new
Abstract: Learning across medical imaging modalities requires making use of images without paired text and accommodating new domains while retaining existing rep...
By DM BAPPY, Eunsu Kim, Hanjoo Lee, Sung Hak Lee, Sangjeong Ahn