The paper introduces AURA, a framework that supervises intermediate reasoning states to better separate relevant documents from hard distractors and trains a lightweight utility estimator to adapt reasoning depth during inference. On the BRIGHT benchmark, AURA-4B achieves an average nDCG@10 of 34.7 while using only 2.8 reasoning steps and 96 generated tokens per query, and adaptive stopping reduces latency from 4.6 to 0.76 seconds compared to full-trajectory reasoning. Experiments on MS MARCO indicate that this specialization does not significantly harm conventional in-domain retrieval performance.
By Yingtao Luo, Kangyi Zhao, Shiyi Du, Yubo Li
VoteGEO-Bench is a benchmark that examines how generative engine optimization (GEO) affects the visibility of election-related web content in large language model responses. It covers the 2024 U.S. presidential election and the 2026 Japanese House of Representatives election, using election queries, historical webpages, and voter personas. The study evaluates 11 rewriting strategies, finding that target‑specified rewrites and those that improve page organization or strengthen arguments increase exposure more than other tactics, while partisan identities can create larger baseline exposure gaps than a single rewritten page.
By Shiqi Zhang, Xiaokui Xiao
The study identifies a specific attention head (L8H10) in Vision Transformers that reads annotation text on chest X‑ray images instead of lung pathology, affecting 62% of COVID‑19 images. Causal inpainting experiments confirm that the head’s behavior is driven by annotation content, and removing the head increases lung attention by 1.52×. This spurious annotation‑attending behavior is consistent across multiple ViT architectures and datasets, persists after full fine‑tuning, and is invisible to standard output‑level evaluation, highlighting the need for mechanistic auditing.
By Joshua Muhumuza, Cornelius Kalenzi
MBBQ (Multilingual Bias Benchmark for Question-answering) is a curated extension of the English BBQ dataset that includes Dutch, Spanish, and Turkish, designed to measure stereotypes across these languages while controlling for cultural differences and task accuracy. The dataset also provides a parallel control set to assess question-answering performance independently of bias. Experiments with several open-source and proprietary LLMs show that non‑English languages can exhibit greater bias than English, and cross‑lingual bias differences persist except in the most accurate models.
By Vera Neplenbroek, Arianna Bisazza, Raquel Fern\'andez
EnsembleLink is a record‑linkage method that achieves high accuracy without requiring any training labels. It uses pre‑trained language models to capture semantic relationships, enabling it to match records across diverse domains such as city names, person names, organizations, multilingual political parties, and bibliographic entries. The approach runs locally on open‑source models, avoiding external API calls, and completes typical linkage tasks in minutes.
By Noah Dasanaike
BASIS is a benchmark that tests large language model agents on their ability to infer users’ mental states, detect belief‑reality mismatches, and use this insight to craft better solutions in interactive settings. The study shows that current models often ignore users’ mental states, leading to poorer outcomes, while explicitly eliciting and incorporating such information improves solution quality. Additional experiments compare BASIS with Chain‑of‑Thought prompting, targeted training, and multi‑agent collaboration to demonstrate the practical benefits of mental‑state awareness.
By Minyuan Ruan, Ziyue Wang, Kaiming Liu, Yunghwei Lai, Peng Li, Yang Liu
MIXAR is the first generative multilingual pixel‑based language model, pretrained on eight high‑resource languages across Latin and CJK scripts with up to 0.5 B parameters. It replaces unstable adversarial training with a GRPO‑based reinforcement‑learning stage, improving free‑generation accuracy by up to 2.3× on multilingual benchmarks. MIXAR generalizes better to unseen low‑resource languages than a tokenizer‑based model trained on the same corpus, matches tokenizer models on discriminative tasks, and resists orthographic attacks more gracefully.
By Chen Hu, Yintao Tai, Antonio Vergari, Frank Keller, Alessandro Suglia
SimCT is a method for on‑policy distillation that addresses the problem of mismatched tokenizers between teacher and student models. By comparing teacher and student predictions over short multi‑token continuations that both tokenizers can produce, SimCT restores supervision that would otherwise be lost when only exact shared tokens are used. Experiments on mathematical reasoning and code‑generation tasks show consistent improvements over shared‑vocabulary OPD and other cross‑tokenizer baselines, with ablations confirming the gains come from recovering discarded supervision.
By Jie Sun, Mao Zheng, Mingyang Song, Qiyong Zhong, Yilin Cheng, Bichuan Feng, Pengfei Liu, Junfeng Fang, Xiang Wang
The paper introduces Convex Gate (C‑Gate), a bridge that maps each speech frame to a convex combination of frozen token embeddings, enabling a single LLM decoder to process both speech and text. Using a Whisper‑large‑v3 encoder and a partially adapted Qwen2.5‑7B‑Instruct decoder, the authors jointly train on speech recognition, emotion recognition, and spoken reasoning, achieving accurate transcription and emotion detection while maintaining strong performance on text‑only benchmarks. Analyses reveal that the fixed text embeddings can support both semantic and acoustic readouts through continuous, speech‑conditioned mixtures.
By Ming-Hao Hsu, Yuxuan Hu, Shujie Liu, Jinyu Li, Yan Lu, Zhizheng Wu
The paper introduces Adaptive Data Scheduling (ADS), a dual-level framework that improves reinforcement learning for large language models by tailoring data sampling to both semantic clusters and the policy’s current capability. ADS first reorganizes training samples into semantic clusters and adjusts the inter-cluster sampling distribution to focus on clusters the policy can currently handle. Then, within each cluster, ADS selects prompts near the policy’s capability boundary to provide informative relative advantages. Experiments on three LLMs and seven reasoning benchmarks show ADS boosts average accuracy by 5.5% over Group Relative Policy Optimization (GRPO) and consistently benefits various RL objective designs.
By Zicheng Xu, Ruixuan Zhang, Yu-Neng Chuang, Xiuyi Lou, Hoang Anh Duy Le, Oren Gal, Alexander S. Szalay, Zhaozhuo Xu, Guanchu Wang, Vladimir Braverman
The paper introduces SetwiseEvalKit, a benchmark with nine dimensions and 28K rubrics for evaluating document sets, highlighting that current rerankers only achieve up to 45% coverage and struggle with cross‑document coordination. It proposes Rubric4Setwise, a training‑free method that transforms rubric criteria into selection signals, outperforming other methods in downstream generation tasks across both short‑form and long‑form scenarios. The study demonstrates the importance of evaluating and optimizing document sets as a whole rather than independently.
By Kailin Jiang, Lei Liu, Jian Xi, Hui Xu, Junlin Liu, Baochen Fu, Bin Li, Vichwang, Yu Lu, Haibo Shi
The paper introduces Archer, a training‑free caching method that enables efficient rollback in diffusion language models by reusing cached hidden states within a bounded neighborhood. Archer recomputes mutable responses at each update while reusing prompt key/value pairs, thereby amortizing prompt computation without limiting rollback capability. Experiments on code‑generation and non‑code benchmarks show significant speedups—up to 4.33× over LLaDA‑Confidence—while maintaining or improving accuracy.
By Xuning He, Qingyuan Shi, Zinan Sheng, Yongding Tao, Huanyu Liu, Ge Li, Xue Jiang, Yihong Dong
The paper introduces MIDA, a multimodal benchmark that tests whether Multimodal Large Language Models can detect deception in multi‑party social interactions. It evaluates 12 state‑of‑the‑art MLLMs, finding that even powerful models like GPT‑4o perform poorly at distinguishing truthful from deceptive statements. The authors propose a Social Chain‑of‑Thought reasoning pipeline and a Dynamic Social Epistemic Memory module, showing that structured epistemic memory improves performance but that multimodal social‑cue reasoning remains a significant challenge.
By Caixin Kang, Yifei Huang, Liangyang Ouyang, Mingfang Zhang, Ruicong Liu, Yoichi Sato
Orbia is a new benchmark that systematically evaluates 3D consistency in interactive video world models. It organizes evaluation into three dimensions—input preservation, generated‑content persistence, and 3D self‑consistency—using controlled exploration and revisit protocols. The benchmark covers 720 instances across indoor and outdoor scenes and tests 27 different models, providing a comprehensive framework for diagnosing and improving 3D consistency.
By Dongyue Lu, Guian Fang, Hur Lim, Yonggang Wu, Yanrong Wang, Hongshan Chen, Shengju Qian, Yue Liao, Wei Chow, Lingdong Kong, Tianxin Huang, Yu Yang, Hongji Yang, Junchao Huang, Ao Liang, Yihao Wang, Zihan Wang, Rong Li, Hanlin Chen, Xin Wang, Wei Tsang Ooi, Mike Zheng Shou, Shuicheng Yan
The paper introduces TextDART, a defect-aware reward model that evaluates individual text regions for semantic correctness and glyph quality, and SCANNER, a method that uses these regional judgments to provide spatially differentiated policy feedback. This approach addresses limitations of previous reinforcement learning methods that overlook glyph defects and obscure regional quality. Experiments on existing benchmarks and the new PolyText-Bench show state‑of‑the‑art performance in text accuracy and glyph fidelity.
By Sujie Hu, Huaisong Zhang, Pujing Yang, Qian Wang, Xiaomin Li, Xinhang Leng, Chen Li, Jing Lyu, Xiu Li, Ling Pan
The paper proposes a data selection method that reallocates training samples across visual concepts such as objects, attributes, color, texture, and backgrounds. By using sparse autoencoders to extract concept activations and applying diminishing returns to concept support, the method down‑weights well‑represented concepts and focuses on the tail, while also incorporating local support to strengthen under‑supported concepts. Experiments across fine‑grained, spurious‑correlation, general image classification, and multimodal pretraining show consistent improvements over strong baselines and reveal a link between enhanced concepts and robustness to out‑of‑distribution and spurious contexts.
By Ziheng Zhang, Neha Hulkund, Colin Lee, Zheda Mai, Hui Shen, Ying Ding, Tanya Berger-Wolf, Sara Beery, Wei-Lun Chao, Jianyang Gu
RoboCloth is a new dataset comprising 500 real cloth materials captured with robotically controlled cameras and lighting. Each material is imaged in about 580 high‑dynamic‑range frames across a wide view‑light hemisphere, with camera poses refined via structure‑from‑motion and aligned to the robot frame to provide per‑material sparse geometry and structured reflectance observations. The authors use this data to train a two‑stage neural reconstruction pipeline, demonstrating that a decoder trained on RoboCloth outperforms those trained on existing datasets and analytic PBR fitting, and transfers well to unseen acquisition systems.
By Zhen Li (Simon Fraser University), Haoran Qin (Westlake University), Francesco Di Sario (University of Torino,Simon Fraser University), Daniel Rebain (University of British Columbia,Wayve), Andrea Tagliasacchi (Simon Fraser University, Wayve,University of Toronto)
ControlFlow3D is a framework that upsamples point clouds by distilling knowledge from pretrained foundation models into a lightweight generation network. It performs conditional flow matching in a latent space defined by a frozen 3D encoder, using unsupervised soft KMeans to cluster multi‑view image tokens and 3D latent tokens into shared prototypes that guide the flow trajectory. A Mamba‑based geometry decoder then maps latent features to dense coordinates with linear complexity, and the authors introduce ShapeNetPU, a 35K‑object multi‑modal benchmark 31 times larger than previous datasets.
By Yuang Liu, Zhi Zuo, Zhengkai Zhao, Lirui Zhang, Pan Gao, Hao Feng, Zhengzhe Liu
ReState is a training‑free method for object removal in few‑step diffusion models that corrects the masked clean prediction toward background evidence using a noise‑level‑aware weight that fades out at the final forward pass. It complements attention rerouting, requiring no additional denoiser evaluations, and outperforms existing training‑free techniques on two paired benchmarks and two distilled backbones, improving PSNR by an average of 3.17 dB. On the OBER‑Test set, ReState also surpasses a 50‑step attention‑forcing reference while running 6.7× faster.
By Donghwan Lee, Sanghyun Jo, Sihyun Kim, Ziseok Lee, Kyungsu Kim
VLA‑Δ introduces a framework that explicitly learns physical state transitions for vision‑language‑action models, using changes between current and future states as learning targets. The authors build a large‑scale pretraining dataset with physical state‑transition annotations that cover task‑relevant semantics and visual motion changes, providing direct supervision for continuous action generation. The resulting model performs competitively on LIBERO and RoboTwin benchmarks and outperforms π₀.₅ in real‑robot experiments, while requiring no extra inference overhead.
By Qi Lu, Jiarun Li, Ce Zhang, Miao Pan, Yi Pan, Xu Wang, Wenqi Zhang, Xuhong Zhang