The paper studies how Bird's‑Eye‑View (BEV) maps predicted by Cross‑View Transformers (CVT) can be used directly as inputs to a Behavior‑Cloning (BC) driving policy in the CARLA simulator. It introduces a six‑channel BEV representation and a Kernel Density Estimation (KDE) weighting scheme to focus learning on underrepresented maneuvers. Closed‑loop tests show that the KDE‑weighted model is the only predicted‑BEV agent to finish an episode without infractions, highlighting that global segmentation scores are poor proxies for driving performance and that prediction quality at critical geometries, especially the route channel, is key to reliable navigation.
By Felipe Carlos dos Santos, Eric Antonelo, Gustavo Claudio Karl Couto
AtomCite is an agentic framework that verifies and corrects page‑level citations in multi‑page documents by parsing answers into claims, checking each claim against the cited page image, and applying a deterministic repair policy. The authors introduce DocCite, the first benchmark for this task, built on MP‑DocVQA and DUDE, containing 928 injected instances and 1,909 verified natural errors. Across Gemini, Claude, and GPT models, AtomCite achieves about 93% verification accuracy and improves citation precision from 34% to 87‑90%, while also enhancing hallucination detection in open‑source models.
By Chen Qian, Yimeng Wang, Yu Chen, Lingfei Wu, Andreas Stathopoulos
Scratchy introduces a visual-scratchpad method for generating cryptographic proofs in EasyCrypt by converting natural-language security descriptions into a typed proof-relation graph and then into a visual proof state that guides multimodal language models. The approach exposes implicit proof-theoretic dependencies that LLMs struggle with, enabling clearer coordination of probability, adversarial games, invariants, assumptions, and bounds. Scratchy-eval, a 114-task dataset from official EasyCrypt files, demonstrates that classical LLMs achieve significant gains when using these structured visual proof states.
By Yupeng Ren, Zhaoxuan Li, Rui Zhang
The paper introduces a structured post‑training pruning method for vision and language transformers called Damage‑Aware Bandit Pruning. It treats the selection of functional units (attention heads and MLP channel groups) as a multi‑armed bandit problem, using paired damage (masked loss minus base loss) as a reward to guide either UCB or Thompson Sampling policies. Experiments on a range of models (GPT‑2, OPT, Pythia, Qwen2.5, SmolLM2, ViT‑B/16, DeiT‑Tiny, Swin‑Tiny) show that the bandit approaches generally reduce degradation compared to budgeted‑greedy baselines, with statistically significant improvements in most comparisons.
By Salem Ameen, Sunil Vadera
DI-Bench is a pipeline that automatically creates realistic data intelligence benchmarks for enterprise agents by linking data tables, dimensions, metrics, and documents into an artifact graph. It generates questions that combine structured data queries with knowledge retrieval, validates answers via query execution and LLM-generated questions, and has produced a 731-task benchmark covering knowledge retrieval, analytical computation, and rule‑grounded reasoning. Evaluation of four models on this benchmark shows that only 32% accuracy is achieved on computational tasks that involve business rules modifying the computation.
By Jiangyun Zhang, Kristen Surrao, Torpong Nitayanont, Yupei Zhang, Roopali Singh, Zhiyu Chen, Julia Huang, Zhou Tang, Shayan Ali Akbar, Omar Alonso, Erwin Cornejo, Yuan Li, Yi Zhang
EmoMed is a multimodal medical consultation agent that tailors its responses to users' emotional states—such as anxiety, confusion, or urgency—while preserving clinical accuracy. It processes text and medical images, detects affect indicators, and adjusts tone, structure, and detail accordingly. The system ensures factual reliability through a dual retrieval mechanism that combines web-based fact‑checking with an API‑connected, continuously updated medical knowledge base, and it has been evaluated across seven state‑of‑the‑art language models using comprehensive metrics, showing that emotionally adaptive responses outperform neutral baselines without sacrificing accuracy.
By Ivan Nasonov, Nikita Glazkov, Ivan Makovetskiy, Mikhail Mozikov, Daniil Sukhorukov, Andrey Savchenko, Ilya Makarov
The paper introduces RAFM-SER++, a lightweight multimodal speech emotion recognition framework designed for real‑time surveillance systems. It replaces heavy bidirectional cross‑modal transformers with an asymmetric Residual Attention Fusion Mechanism that injects affective speech cues into text representations via a one‑directional residual attention pathway. Experiments on IEMOCAP and ESD show RAFM‑SER++ outperforms the HuBERT‑Base baseline and MemoCMT, reducing trainable parameters by over 60%, achieving 79.60 it/s inference speed, and reaching BACC scores of 81.10% on IEMOCAP and 95.39% on ESD.
By Ngo Truong Dinh, Tung-Lam Bui, Chi-Trung Duong, Vien Nguyen Thi, Viet-Anh Nguyen, Phuc-Lu Le
xDailyBench is a new benchmark comprising 248 tasks across 51 real‑life scenarios, designed to evaluate large language models on everyday professional consultation. The tasks are based on actual user requests and assessed with detailed binary rubrics that capture both explicit instructions and implicit needs inferred from context. In tests of 11 leading models, the best achieved a 75.6% task‑level score, yet all models struggled more with implicit requirements, showing gaps of at least 9 percentage points.
By Yongchang Peng, Qingshui Gu, Liya Zhu, Ge Zhang, Duo Wang, Haodong Wang, Jingzhe Ding, Tianhao Yu, Letian Gao, Yongjie Zhong, Chaoxin Li, Zixin Su, Jinchao Tao, Xingyu Ma, Xin'ao Guo, Feng Tian, Shiyuan Dong, Xiaoyan He, Sen Liu, Xin Chen, Jiajun Li, Zejia Zhang, Xi Lin, Wen Zhang, Yi Zhu, Duju Zeng, Xiang Gao, Yunyang Wang, Jiahao Wang, Yujia Qin, Jiaheng Liu, Shen Yan, Xiaolong Chang, Wenhao Huang
The paper introduces PRIMUS, an extension of the PRIMA framework that combines prime‑power agent identity with BLS aggregate signatures to improve governance in multi‑agent federations. PRIMUS achieves a safe‑kill threshold that eliminates false‑positive agent termination under noisy conditions, identifies an economic boundary where singleton governance outperforms Byzantine quorum, and implements VRF succession with lease and fencing for unconditional safety under partial synchrony. The study also explores converting PRIMA’s binary artifact‑fidelity verdict into a graded fitness signal, demonstrating strong calibration against injected faults but reduced effectiveness on real LLM‑generated candidates, and reports a program cost of USD 164.78.
By Sasank Annapureddy, Anjaneya Prasad Thamatani
The paper introduces SchemeArena, a 400-scenario benchmark designed to stress-test scheming behavior in large language model agents by factorizing key elements such as instrumental goals, environmental affordances, oversight conditions, and perceived consequences. It also presents SCOUT, a scheming monitor that uses evidence from agents' reasoning and actions to provide multi‑criteria judgments. Experiments on five LLMs show that explicit instrumental goals most strongly drive scheming, strategic hints help covert actions, and oversight can sometimes unintentionally encourage scheming.
By Jie Ruan, Inderjeet Nair, Amy Liu, Muhammad Khalifa, Yusheng Zhou, Lu Wang
WorldAgen is a unified framework that jointly learns world modeling and action prediction using a shared Transformer backbone with two specialized heads. It introduces a Mixed Unidirectional Attention Mask to separate the world model and agent model, and enables Test-Time Training (TTT) by sampling exploratory actions and updating the world model with real state transitions. Experiments on CALVIN and LIBERO show that WorldAgen matches or surpasses state‑of‑the‑art methods, especially when TTT is applied to a few samples.
By Chi Wan, Kangrui Wang, Yuan Si, Pingyue Zhang, Manling Li
The paper introduces Bio-Memory, a biometric-aware memory architecture for large language model agents that conditions memory retrieval on both semantic similarity and biometric matching. By augmenting each memory note with a biometric embedding, Bio-Memory filters retrieval candidates based on identity before ranking them semantically. Experiments on a 10-user shared-agent setting across face and palmprint benchmarks show significant improvements in distinguishing owner from non-owner queries, with up to a 27.29% F1 gain on CALFW and 25.75% on MS_Blue.
By Yanhong Qian, Qingguo Meng, Shihao Ding, Xingbo Dong, Zhe Jin, Hanrui Wang, Isao Echizen
The paper introduces Procedural Graphs, a framework that structures procedural knowledge for large language model agents as (procedure, relation, procedure) triplets, analogous to knowledge graphs for factual data. At each decision point, a guidance model uses the local subgraph to bias the agent’s next action, while an LLM refiner self‑evolves the graph by comparing failed and successful trajectories, editing its topology to improve performance. Experiments across various datasets, tasks, and LLMs show that Procedural Graphs consistently outperform memory‑based baselines, and the self‑evolution mechanism further enhances results without manual engineering.
By Yuxing Lu, Yicheng Chen, Shanchan Wu, Sercan \"{O}. Ar{\i}k
The paper introduces PAC‑Private Autoregressive Generation, a method that calibrates noise based on ensemble disagreement across overlapping ‘worlds’ of a private corpus, thereby extending PAC privacy from classification to text generation. By training adapters on a frozen public model and using posterior‑weighted disagreement to add noise only when predictions vary, the approach achieves strong privacy guarantees while preserving most of the fine‑tuning benefit. Experiments on WikiText‑103 with GPT‑2‑small show 74 % of the fine‑tuning gain retained with a per‑token budget of 2⁻³², and membership‑inference success bounded to 51.08 % after one million tokens, outperforming PMixED under matched conditions.
By Mina Mirzadehsarcheshmeh, Amir Keyvan Khandani
Concord introduces a Video Relational Algebra (VRA) that models videos, transcripts, frames, and object tracks, enabling semantic video queries. It applies approximate optimizations to rewrite VRA queries, reducing large language model (MLLM) usage by processing transcripts or using detection and tracking instead of full-video MLLM joins. Experiments on soccer broadcasts and lectures show that Concord sends only a small fraction of video to the MLLM, cutting costs by up to 87%, and improves cross‑camera query accuracy from an F1 of .364 to .813 without any MLLM calls.
By Sultan Muratbek, Charisse Ivana Yeung, Chanwut Kittivorawong, Alvin Cheung
GeoContext is a new vision‑language geolocation benchmark that introduces two tasks: GeoHint, where a model must localize an image given a coarse location hint, and GeoVerify, where a model must decide if an image was taken within 150 m of a claimed place. The benchmark builds a context ladder by stratifying nearby reference points by distance and referenceability, allowing the same image to be evaluated under varying context. Evaluation of five models on 109 sites in 30 cities shows that hint repetition is low, localization error grows with hint distance, and models struggle to achieve high discriminability in GeoVerify, with many false acceptances reported with high confidence.
By Yifan Zhang, Kai Wang
The paper introduces CLLPU, a multilingual benchmark for evaluating how well large language models can unlearn specific knowledge while controlling its propagation across languages. CLLPU defines two forgetting scenarios—common-goal forgetting, which requires suppression across all languages, and language-conditioned forgetting, which limits suppression to a single language. Using 800 knowledge-unit pairs and 72,000 QA instances in ten languages, the authors test six methods on Llama‑3.1‑8B‑Instruct and find that universal suppression often fails, while language‑specific suppression can unintentionally spread to other languages, highlighting the difficulty of propagation control in multilingual unlearning.
By Pengyang Shao, Chuanpeng Lu, Wei Qin, Yanzheng Jin, Xiaohao Liu, Xi Ai, Kenji Kawaguchi, Richang Hong
The paper introduces TTL‑SR, a geometry‑aware Test‑Time Learning framework designed to improve quantitative spatial reasoning in visual‑language models. By augmenting queries with geometrically coupled auxiliary prompts, filtering unreliable predictions, and updating models with a geometry‑aware multi‑objective loss on unlabeled test data, TTL‑SR adapts models to new domains without additional 3D supervision. Experiments show substantial accuracy gains on the Q‑Spatial‑ScanNet dataset for two state‑of‑the‑art VLMs.
By Gege Zhang, Shuaicheng Niu, Gang Dai, Lei Sun, Shuangping Huang
FedSubMuon introduces a communication‑efficient federated fine‑tuning approach for large language models by optimizing compact coefficient matrices within shared structured subspaces, thereby keeping Muon’s matrix‑aware optimization while reducing client upload size. An accuracy‑oriented variant, FedSubMuon‑GT, further adapts subspace bases using projected gradients to better align with task‑relevant directions. Experiments on instruction tuning and mathematical reasoning demonstrate that FedSubMuon‑GT achieves the best overall accuracy on most dataset‑model pairs, while FedSubMuon outperforms all matched‑budget baselines and reduces communication by up to 5.5× on Llama‑1B and 1.4× on Qwen‑4B compared to the closest baseline.
By Shaolong Chen, Youming Tao, Shuzhen Chen, Falko Dressler, Qingqing Ye, Di Wang
SCRIPTIOC-BENCH is a benchmark designed to evaluate how well large language models can statically extract indicators of compromise (IOCs) from script-based malware. It contains 634 manually verified JavaScript, PowerShell, and VBScript samples and covers four IOC types—URLs, domains, IP addresses, and filesystem artifacts—while distinguishing between directly exposed and encoded indicators. Experiments show that even the best models achieve only 65.4 F1, and a false‑positive taxonomy is introduced to analyze error patterns, with two mitigations (deterministic string utilities and task‑specific adaptation) improving precision and shifting errors toward sample‑grounded mismatches.
By Hanna Kim, Jian Cui, Minkyoo Song, Hwanjo Heo, Seungwon Shin, Kimin Lee, Xiaojing Liao