ViSAR is a training‑free, adaptive‑k retrieval method for Visual Document Question Answering that operates directly in the embedding space to build a query‑conditioned page‑level similarity matrix. By dynamically selecting the number of pages to retrieve based on query relevance, ViSAR reduces Retrieval‑Augmented Generation latency by up to 58.7% while maintaining or improving answer accuracy across multiple encoders and Large Vision‑Language Models. The structure of the similarity matrix also correlates with answer accuracy, indicating potential for retrieval quality‑aware document understanding.
By Adrien Mialland, Marc Plantevit, Julien Gallois, C\'eline Robardet
BioELX is a retrieve‑rerank framework for cross‑lingual biomedical entity linking that tackles two key problems: the English‑biased UMLS alias training data and the degradation caused by naïvely adding context. It fine‑tunes SapBERT_multi with Wikidata‑derived cross‑lingual alias supervision to create shared concept neighborhoods, and then reranks candidates using pretrained LLMs with mention‑anchored prompting to focus on the target mention. Experiments demonstrate state‑of‑the‑art performance on four benchmarks, improving Recall@1 by 4.8–18.2 percentage points without task‑specific annotations.
By Yi Wang, Corina Dima, Liangyu Zhong, Steffen Staab
The paper surveys Transformer-based large language models (LLMs) with a focus on efficiency, reviewing 312 articles that cover data curation, model design, downsizing, and dynamic inference. It also examines efficiency in adaptation strategies such as pre‑training, fine‑tuning, prompt‑engineering, and Retrieval‑Augmented Generation (RAG). A statistical analysis and evaluation of over 30 prominent NLP models on 13 benchmarks provide insights into both efficiency and efficacy, highlighting trends toward sustainable NLP practices.
By Wazib Ansar, Saptarsi Goswami, Amlan Chakrabarti
The paper proposes a response‑free method for estimating difficulty of reading‑comprehension multiple‑choice items by fine‑tuning a transformer on item wording. It introduces two extensions to a baseline joint‑encoding model: a component‑wise variant that encodes passage, question, and options separately, and a multi‑task variant that adds a question‑answering auxiliary task. Experiments on a corpus of nearly 30,000 items show that both extensions outperform the baseline, especially the multi‑task variant across all metrics and the component‑wise variant in rank ordering, even with limited training data.
By Jan Net\'ik, Patr\'icia Martinkov\'a
The paper investigates how different link prediction models for knowledge graphs produce varying predictions and explores the extent of their complementary knowledge. By evaluating an oracle that selects the best prediction from a set of models, the authors show a significant performance gap between individual models and the oracle, indicating substantial complementarity. However, this complementarity quickly saturates as more models are added, leaving many queries unsolved even with many models.
By Guillaume M\'erou\'e, Fabien Gandon, Pierre Monnin
NE‑R1 is a framework that improves Named Entity Recognition by using a retrieval‑on‑demand mechanism and a two‑stage training process that includes instruction tuning and reinforcement learning with chain‑of‑thought. It balances the use of internal model parameters and external knowledge through a multi‑dimensional reward that considers accuracy and retrieval benefit. The approach achieves state‑of‑the‑art results, improving in‑domain F1 by 2.52% and zero‑shot cross‑domain F1 by 1.18%.
By Meixuan Chen, Hehan Li, Ruizhi Zhao, Xin Lu, peizhi xu, Liwei Qian, LI Meifang, shuanglong li, Hanmeng Liu, Xin Pei, Yanbiao Ma
The paper investigates whether consensus among large language model (LLM) judges truly reflects human alignment. By treating each judge’s scores as vectors, the authors measure spread, effective rank, and angles to human scores across 42 judges on Indic benchmarks, revealing that inter‑judge agreement often mirrors shared blind spots rather than human judgments. They find that while judges agree as much as humans, they only reach 58‑66% of human agreement and frequently focus on axes humans do not weight, indicating that ensemble agreement alone is insufficient evidence of alignment.
By Sourabrata Mukherjee, Hamna Hamna, Kalika Bali, Sunayana Sitaram
The paper investigates language-of-study (LoS) bias in NLP peer reviews, defining and distinguishing negative and positive forms of bias. Using a new dataset, LOBSTER, and an LLM-based detection pipeline, the authors analyze 15,645 reviews and find that non‑English papers experience significantly higher bias rates, with negative bias outweighing positive bias. They further identify four subcategories of negative bias, noting that demanding unjustified cross‑lingual generalization is the most common.
By Ehsan Barkhordar, Abdulfattah Safa, Verena Blaschke, Erika Lombart, Marie-Catherine de Marneffe, G\"ozde G\"ul \c{S}ahin
The paper introduces CARPAS, a new task that dynamically refines user-provided aspects for aspect-based summarization in large language models (LLMs). It presents three new datasets and evaluates four prompting strategies, finding that LLMs tend to over-generate aspects, leading to overly long and misaligned summaries. To address this, the authors propose a two-stage framework that first generates lightweight scope guidance before aspect refinement and summarization, which improves focus, reduces over-generation, and enhances performance across all datasets.
By Yong-En Tian, Yu-Chien Tang, An-Zi Yen, Wen-Chih Peng
The paper presents a semester-long deployment of the VideoPoints platform, featuring a retrieval‑augmented chatbot that answers questions using only the active course’s lecture materials and provides clickable, timestamped citations. Across 833 interactions, 70.5% of messages included citations, no cross‑course references were made, and the bot declined to answer when no lecture evidence matched. The study also shows that the system improves correct‑lecture retrieval by 6.3 percentage points over dense‑only retrieval on the EduVidQA benchmark.
By S M Masrur Ahmed, Jaspal Subhlok
The paper introduces Padyam2Gadyam, a dataset of 600 13th‑17th Century Telugu poems paired with human‑verified Telugu and English prose translations. It evaluates two traditional machine translation systems and five large language models on zero‑shot poem‑to‑prose translation, finding that general‑purpose LLMs outperform the MT systems but still exhibit systematic issues in generating and evaluating prose translations.
By Chalamalasetti Kranti, Sowmya Vajjala
The paper introduces Evidence Sufficiency Boundary Training, a framework that teaches models to abstain from answering until the supplied evidence is fully sufficient, and to remain stable when additional redundant evidence is added. By constructing ordered evidence chains from datasets such as HotpotQA, 2WikiMultiHopQA, and MuSiQue, the method applies level supervision, a boundary flip margin, post‑boundary stability, and answer recall protection. Experiments with Qwen2.5‑3B‑Instruct and LoRA adaptation show improved boundary localization (flip accuracy 0.807 vs 0.781) and a lower unsupported‑answer rate (0.095 vs 0.101) while maintaining competitive raw QA F1.
By Haruto Sato, Yuki Tanaka, Ren Nakamura, Aoi Kobayashi, Mei Ito
FDARxBench is an expert‑curated benchmark designed to evaluate document‑grounded question answering on FDA drug label documents, focusing on generic drug assessment. It features a multi‑stage pipeline that generates high‑quality QA examples covering factual, multi‑hop, and refusal tasks, and includes protocols for both open‑book and closed‑book reasoning. Experiments with various language models show significant gaps in factual grounding, long‑context retrieval, and safe refusal behavior, highlighting the challenge of regulatory‑grade label comprehension.
By Betty Xiong, Jillian Fisher, Benjamin Newman, Meng Hu, Shivangi Gupta, Yejin Choi, Lanyan Fang, Russ B Altman
The paper introduces BASSE, a multilingual meta‑evaluation dataset containing 2,040 human‑rated abstractive summaries produced manually or by five LLMs with four prompts. Annotators scored each summary on coherence, consistency, fluency, relevance, and 5W1H using a 5‑point Likert scale. Benchmarking shows proprietary LLM‑judge models best align with human judgments, followed by criteria‑specific automatic metrics, while open‑source judge LLMs perform poorly.
By Jeremy Barnes, Naiara Perez, Alba Bonet-Jover, Bego\~na Altuna
GEM is a Generative LiDAR world model that uses a deformable Mamba architecture to better handle the disorder of LiDAR point clouds and distinguish dynamic objects from static structures. The model tokenizes LiDAR sweeps, unsupervisedly disentangles dynamic and static features, and applies a tri‑path deformable Mamba for selective scanning and adaptive gating fusion, improving spatial‑temporal understanding. Experiments show GEM outperforms existing methods across multiple benchmarks, and it can be paired with a planner and BEV controller for autonomous rollout and "what‑if" scenario generation.
By Yang Wu, Zhaojiang Liu, Qiang Meng, Youquan Liu, Renliang Weng, Jianjun Qian, Jian Yang, Jin Xie
The paper introduces DLM-One, a score‑distillation framework that enables one‑step sequence generation with continuous diffusion language models (DLMs). By aligning a student model’s outputs with a pretrained teacher DLM’s score function in the forward‑diffused noisy space, DLM-One removes the need for iterative refinement. Experiments across various DLM architectures show up to ~2000× speedup in sampling steps and ~500× in wall‑clock time while retaining competitive performance, and the authors propose an adversarially‑regularized two‑stage training scheme to mitigate student degeneration.
By Tianqi Chen, Shujian Zhang, Mingyuan Zhou
SGD-KV is a head‑aware framework that compresses key‑value caches in large language models by using a chunk‑summarization diagnostic task to identify attention heads that specialize in hierarchical information aggregation. It prioritizes these heads during compression, achieving state‑of‑the‑art performance on long‑context benchmarks with up to 1M tokens while cutting KV cache memory usage by as much as 75%. Experiments on Qwen2.5‑7B‑1M and Qwen3‑32B confirm that allocating cache budget based on summarization scores yields a superior efficiency‑accuracy trade‑off for long‑context inference.
The paper examines how large language models resolve conflicts that arise within contextual knowledge, rather than between internal knowledge and external context. It introduces a taxonomy of six contextual conflict types and presents the ContextConflict dataset with 5,781 samples covering reasoning and summarization tasks. Experiments on nine LLMs reveal persistent shortcomings in conflict resolution, uncover a bias toward earlier evidence, and propose a training‑free steering method that improves accuracy and summary quality.
The paper introduces the Conformal Relevance framework, which leverages in-context learning example curation and ensembling to generate a score function that preserves coverage while enhancing conciseness for NLP tasks such as summarization and extractive question answering. Unlike traditional methods that rely on labor-intensive, hand-engineered LLM prompts to rate content importance, this approach requires minimal manual input. The authors validate the framework across seven NLP tasks and provide theoretical insights into how diversity in ensembled conformal scores can improve worst-case sentence scores, including a saturation bound on ensemble gains.
Loom is a generative consensus framework designed for real‑world root‑cause analysis (RCA) that combines open‑form hypotheses from modular heuristics with a lightweight large language model (LLM). It projects hypotheses into a continuous embedding space and uses an iterative centroid‑based reweighting algorithm to resolve conflicts, producing a single consensus that is then synthesized by one LLM call. On the OpenRCA benchmark, Loom achieves state‑of‑the‑art accuracy on Bank and Market‑2 while being significantly faster and more efficient than existing autonomous agents.