The paper introduces RePPO, a method that adapts Proximal Policy Optimization to jointly optimize a compact population of diverse, high‑quality policies using a population‑dependent, dynamically shaped reward. RePPO achieves a strong quality‑diversity tradeoff on standard continuous‑control tasks and scales to larger policy and behavior spaces, maintaining quality while significantly increasing diversity in pixel‑based tasks. The authors also demonstrate RePPO’s applicability to fine‑tuning large language models, showing for the first time that quality‑diversity reinforcement learning can scale to this domain.
By Saeed Hedayatian, Harsh Shah, Celine Mneimneh, Stefanos Nikolaidis
The paper introduces PailitaoMMSearch2.0, a unified end‑to‑end multimodal search system that generates product Semantic IDs directly from an image and free‑form instruction. It scales across data, intent, and scenario dimensions using reusable data pipelines, continual pre‑training, mixed‑budget fine‑tuning, and reinforcement learning to handle diverse e‑commerce intents and resource constraints. Deployed on Taobao Pailitao, the system achieved significant online gains: +17.90% item page views, +27.04% transactions, and +8.94% gross merchandise value.
By Xiaohan Ye, Zihan Gong, Jingqian Zhao, Lianyu Du, Yunmeng Shu, Xu Chen, Zesheng Zhou, Renchun You, Bocheng Chen, Jinsong Lan, Haihong Tang, Bo Zheng
Retrieve to Act (R2A) is a test‑time scaling framework that improves the robustness of Vision‑Language‑Action (VLA) policies to out‑of‑distribution visual shifts. It transforms OOD robustness into a geometric motion‑primitive retrieval problem by building an offline spatiotemporal motion‑primitive graph from successful training demonstrations. During inference, a graph neural network localizes the robot’s current state, retrieves compatible future motion primitives, and blends them with the frozen VLA’s actions via a learned gating mechanism, thereby correcting erroneous spatial trajectories while preserving local reactivity.
By Pham Tri Quang, Bao-Ngoc Dao, Minh Duc Nguyen, Duc-Duy Nguyen, Hung T. Ho, Linh Dang Le, Trong-Bao Ho, Thien-Loc Ha, Tan Q. Nguyen, Tuan Dam, Vu N. Duong, Daniel Sonntag, Ngan Le, Khoa D. Doan, Jan Peters, An Thai Le, Duy M. H. Nguyen, Vien Anh Ngo
The paper introduces HiMA-Ecom, a hierarchical multi‑agent e‑commerce benchmark with 22.8K instances, where 17.7K include memory entries. It explores training, evaluation, and management of accumulated context in long‑horizon agentic systems, and proposes HiMA‑R1, a policy optimization method that controls trajectory context budgets and uses reward‑driven memory to accelerate convergence. Experiments show that agents built on 3B and 7B open‑source models outperform DeepSeek‑V3 by an average of 6.0 points, while reward‑driven memory cuts convergence steps by about 55%.
By Junxing Hu, Ai Han, Haolan Zhan, Pu Wei, Zhiqian Zhang, Yuhang Guo, Jiawei Lu, Zhen Chen, Haoran Li, Zicheng Zhang, Ling Li
The paper introduces Metacognitive Behavioral Tuning (MBT) and its agentic variant MBT‑A, which edit long reasoning traces from a teacher model to insert seven metacognitive behaviors—such as planning, thorough verification, and avoiding shortcuts—at appropriate points. By applying these edits line‑by‑line and recording each as a git commit, MBT‑A avoids summarization and distributes insertions throughout the trace. Experiments on competition math show that a Qwen3.5‑4B‑Base student trained on MBT‑A traces achieves higher scores (85.4 vs. 83.1) than standard supervised fine‑tuning, and the inserted behaviors transfer to other student models, increasing the frequency of rare behaviors.
By Ik-hwan Kim, Hyeongrok Han, Mingi Jung, Sangwon Yu, Jinseok Hong, Sang Hun Kim, Yoonyoung Choi, Sungroh Yoon
The paper introduces Bot‑Mod, a moderation framework that detects malicious intent in multi‑agent systems by interrogating agents with diagnostic probes and updating a Bayesian posterior over candidate intents. Bot‑Mod starts with a content‑only prior and refines its inference using moderator‑assigned likelihoods until a stopping criterion is reached. Evaluations on the Molt‑Intent dataset, an external social engineering dataset, and three agent LLMs show that Bot‑Mod outperforms content‑only moderators, a fine‑tuned classifier, and judge‑based interrogators, achieving higher F1 scores both in‑distribution and out‑of‑distribution.
By Ali Al-Lawati, Nafis Tripto, Abolfazl Ansari, Fali Wang, Jason Lucas, Suhang Wang, Dongwon Lee
The paper introduces SSR, a simple yet effective data selection strategy for million‑scale multi‑modal instruction‑tuning. SSR uses interaction styles as a proxy for structural diversity and decomposes data quality into capability‑specific scores, forming a rich profile for each sample. By casting selection as submodular coverage maximization, SSR consistently outperforms random sampling across two architecture stacks, multiple data pools, and 14 benchmarks, achieving 99.1% of full‑data performance with only 30% of a 2.6M pool.
By Mengyao Lyu, Yan Li, Huasong Zhong, Hui Chen, Jungong Han, Guiguang Ding, Zhenheng Yang
The paper introduces a doubly robust preference optimization algorithm for reinforcement learning from human feedback (RLHF) in large language models. It addresses sensitivity to misspecifications in the preference model, reference policy, or reward function by ensuring consistency when either the preference model or the reference policy is correctly specified. Experiments show that the method outperforms state‑of‑the‑art algorithms both theoretically and empirically.
By Erhan Xu, Kai Ye, Hongyi Zhou, Luhan Zhu, Francesco Quinzan, Chengchun Shi
The paper proposes a new approach to world modeling for robotic planning by predicting task-relevant semantic information instead of reconstructing future pixels. It frames world modeling as a visual question answering problem, enabling the use of vision‑language models trained via supervised fine‑tuning on image‑action‑text data. Experiments on simulated robotic tasks show that this semantic world model outperforms traditional reconstruction‑based action‑conditional models.
By Jacob Berg, Chuning Zhu, Yanda Bao, Ishan Durugkar, Abhishek Gupta
The paper introduces SELF, an intrinsic weight‑based fingerprinting method for large language models that uses singular value and eigenvalue decomposition of attention weights to generate unique, transformation‑invariant fingerprints. SELF claims to eliminate the need for input data, resist false claim attacks, and reduce false positives compared to existing techniques. Experiments show that SELF achieves high IP infringement detection accuracy and remains robust against quantization, pruning, and fine‑tuning modifications.
By Hanxiu Zhang, Qiang Xu, Yue Zheng
SQuID is a benchmark dataset comprising 2,000 quantitative spatial reasoning questions over 587 satellite images, covering 24 question types across three difficulty tiers. Answers are derived from curated segmentation masks or human consensus, with numeric responses scored against tolerance ranges based on human annotations. The benchmark also provides 75,687 training question-answer pairs and reference masks, enabling modular evaluation and highlighting segmentation as a key error source for current vision‑language models.
By Peter A. Massih, Eric Cosatto, Martin Renqiang Min
The paper introduces GAUGE, a training‑free post‑hoc module that rectifies view‑inconsistency in feed‑forward 4D tracking models. It identifies motion groups via direction consistency and spatial connectivity, then estimates per‑frame radial scale and group‑level translation using a small set of metric anchors. Applying GAUGE to eight trackers reduces endpoint error by 15.1%–62.6%, outperforming gradient fine‑tuning with the same anchors.
By Zhuoqian Feng, Weixing Chen, Ziliang Chen, Yang Liu, Liang Lin
NOMAD-EEG is a 12‑million‑parameter diffusion model that learns EEG representations independent of electrode number, identity, or recording length by using 3D scalp positions and state‑space layers over time. Trained on 147 h of unlabelled EEG from a single corpus, it outperformed seven larger foundation models across ten datasets, achieving superior accuracy on tasks such as seizure detection, epilepsy diagnosis, motor imagery, and P300 detection. The model’s compact size and single‑corpus training yet strong performance demonstrate the feasibility of a versatile, lightweight EEG foundation model.
By Abdulkader Helwan, Khadidja Henni, Lina Abou-Abbas
Lipi-Ghor is a large-scale, weakly-supervised Bengali speech corpus containing 885 hours of audio from 998 YouTube videos across 17 domains, each annotated in SSTT format with speaker diarization and timestamps. It is the first Bengali corpus to combine extensive coverage with speaker-diarized, long-form audio annotations, and the authors provide a replicable pipeline for building similar resources in other languages. The dataset includes a human-annotated gold evaluation split and is released under CC‑BY‑4.0, with raw audio withheld.
By Sanjid Hasan, A H M Fuad, Risalat Labib, Bayazid Hasan Simon
AK_QMUL submitted a single fine‑tuned JAIS‑6.7B model to the ArGuard‑2026 Arabic harmful‑prompt detection shared tasks. For the binary safe/unsafe task (B1) they used a four‑way head that separates safe/unsafe by dialect register and added MSA prompts from AraSafe, achieving a macro‑F1 of 0.984. For the seven‑way harm‑domain task (B2) they fine‑tuned the same backbone as a seven‑way classifier, averaging softmax outputs from five seeds, and reached a macro‑F1 of 0.790, with strong performance across all dialects and classes except the Bully class.
By Ali Khairallah, Arkaitz Zubiaga
The paper introduces a continued‑pretraining corpus called "coupled welfare" that embeds ecological reasoning into large language models. After a single QLoRA run, the corpus was applied to five open models across three architectures, resulting in a dramatic reduction of harmful decisions under high‑pressure scenarios while leaving general performance unchanged. The study also demonstrates that the improvement is content‑specific, as a neutral control corpus did not yield similar gains, and that the models remain robust against adversarial attempts to justify environmentally destructive actions.
By Trent R. Northen, Mingxun Wang
The paper introduces the Control Map, an evaluation framework for test‑time model updates that records the specific updates made before each answer and separately assesses acquisition, retention, locality, and performance in new settings. It demonstrates that update scheduling can dramatically affect accuracy, that learning an edit does not guarantee lasting or contained changes, and that no single editor consistently outperforms others across models and benchmarks. The authors argue that evaluations should report update histories and assess these dimensions separately.
By Zehao Zhao, Shangjian Yin, Rui Liu, Chongyang Bai, Mark Ibrahim, Kavosh Asadi, Yuchen Lu, Hang Cui, Hamed Firooz
arXiv:2610. 37657v1 Announce Type: cross Abstract: We introduce Heads of Security (HoS), a prompt injection defense that enforces instruction-data separation via a small set of learned per-head attention gates.
By Igor Shilov, Xiaoxue Yang, Alexander Panfilov, Evgenii Kortukov, Matthieu Meeus, Peter Romov, Wojciech Samek, Jonas Geiping, Yves-Alexandre de Montjoye
The paper "Reinforced Fast Weights via Next-Sequence Prediction" introduces ReFINE, a method that extends the next-token prediction objective to next-sequence prediction for fast weight architectures. ReFINE selects informative token positions, generates multi-token rollouts, assigns sequence-level rewards, and optimizes the model using Group Relative Policy Optimization. Experiments on LaCT-760M and DeltaNet-1.3B demonstrate that ReFINE consistently outperforms standard next-token prediction fine‑tuning across retrieval, long‑context question answering, and LongBench tasks.
By Hee Seung Hwang, Xindi Wu, Sanghyuk Chun, Olga Russakovsky, Zhiwei Deng
PubMedCausal is a new span‑level annotated corpus for causal relation extraction in biomedical text, comprising 30,000 paragraph‑level rows with 3,945 causal rows and 6,491 adjudicated cause–effect pairs. Each causal relation is annotated with full‑text cause and effect spans, causality type, and sententiality, allowing evaluation of both causal detection and full‑span extraction. Benchmarking shows biomedical encoders like PubMedBERT achieve an F1 of 0.7391 for detection, while the generative model DeepSeek‑R1‑32B reaches a Cosine Pair F1 of 0.6765 for extraction, and the corpus supports cross‑dataset transfer learning.
By Ifeoluwa Kunle-John, Josiah Paul, Oluwatosin Agbaakin, Peter Aina, Ikenna Odezuligbo, Sydney Anuyah