DUMA-Bench is a new benchmark that evaluates the security of large language model agents in dual‑control settings, where both the agent and the user can modify the shared environment. It builds on the existing τ²‑bench by adding adversarial environments that cover eight vulnerability classes, such as RAG poisoning and unsafe output handling. The authors tested 14 models from five families and found that dual‑control interaction raises attack success rates from 26.9% to 41.1%, demonstrating that agent security depends on the interaction between model, user, and environment.
By Ivan Aleksandrov, German Kochnev, Sabrina Sadiekh, Yaroslav Rogoza
The paper titled "The Moral Check: Strategic AI Governance for the Pacing Problem" argues that technology cannot self‑steer and that strategy must guide AI development by ensuring purpose and judgment precede compute. It presents a dual contribution: a PRISMA 2020 review of 130 empirical studies and the Strategic AI Governance Ex‑Ante Framework (SAGE‑X), which operationalizes four strategic mindset pillars to mitigate velocity myopia, moral hazard, empirical hazard endpoints, and guardrail decay. The framework includes a calculable Moral Check Index and an Enterprise Lifecycle Audit Instrument to enforce that AI scaling does not outpace deliberative moral judgment, human agency, and societal trust.
By Zaid Amin, Rahma Santhi Zinaida, Nazlena Mohamad Ali
RiverVLN introduces the first benchmark for long‑horizon vision‑language navigation (VLN) of unmanned surface vehicles (USVs) in continuous riverine motion. The PGT‑NAV framework converts navigation instructions into an ordered sequence of visually verifiable semantic phases, maintaining an active phase online through grounded visual and motion evidence. This phase‑grounded approach reduces recursive position and heading drift, achieving a 0.79 success rate in Unity‑ROS closed‑loop tests and demonstrating transfer to real‑world USV deployment.
By Jieling Wu, Yuehao Huang, Jiajun Lv, Tao Huang, Yong Liu, Weiwei Liu
ActGov is a runtime enforcement framework that validates each action proposed by a large language model (LLM) agent before it interacts with external tools, ensuring that actions stay within task‑scoped authorization boundaries and comply with dynamically constructed policies. It builds policies from tool specifications, benign tasks, and failure traces, verifying updates via SMT‑based counterexample checking. In evaluations on AgentDojo and AgentDyn benchmarks, ActGov consistently reduces indirect prompt‑injection attack success while maintaining task utility, outperforming existing defenses.
By Kaiyuan Zhang, Yuke Peng, Ke Jiang, Yinqian Zhang
Visual Jev is a method that encodes an image and its public context once, then processes multiple independent forced‑choice questions in a single batch by reading candidate probabilities from the backbone’s language‑model head. Post‑training on four benchmarks improves macro accuracy from 70.6% to 76.1%, especially for the two task families seen during training. The batched approach is 8.9× faster than serial execution and 3.4× faster than a baseline that recomputes the prefix, though it uses more peak memory.
By Guanxu Yu, Yuhang Yao
The paper proposes three information‑theoretic criteria for selecting the most relevant basis functions in sparse Gaussian process regression, tailored to different levels of prior knowledge. Experiments on six UCI regression datasets and three basis families (HSGP, VFF, VISH) show that the no‑data criterion is a robust default, often outperforming simple truncation, while the data‑aware criteria yield significant improvements for HSGP. The study demonstrates that careful basis‑function selection can lead to better performance without increasing computational cost.
By Marnix Van Soom, Ivan De Boi
Agent0 is a fully autonomous framework that enables large language model agents to evolve without external data by using a multi‑step co‑evolution process. It pits a curriculum agent against an executor agent, both derived from the same base LLM, where the curriculum agent creates increasingly challenging tasks and the executor learns to solve them. By integrating external tools into the executor’s workflow, the system creates a self‑reinforcing cycle that continuously generates high‑quality curricula, leading to significant gains in reasoning performance—an 18% improvement on mathematical reasoning and 24% on general reasoning for the Qwen3‑8B‑Base model.
By Peng Xia, Kaide Zeng, Jiaqi Liu, Can Qin, Fang Wu, Yiyang Zhou, Caiming Xiong, Huaxiu Yao
The paper investigates the trade‑off between sparsity and complexity in high‑dimensional asset pricing models. By separating capacity sparsity (restrictions on effective model capacity) from factor sparsity (parsimonious structure of priced risks), the authors use nonlinear feature expansions, basis pursuit, column generation, and GPU acceleration to estimate models with up to 432 million candidate factors. Their empirical results show that while sparse portfolios underperform dense ridgeless benchmarks at lower complexity, they achieve higher Sharpe ratios and lower pricing errors when the candidate set is large, indicating that capacity expansion and factor sparsity can complement each other.
By Nima Afsharhajari, Jonathan Yu-Meng Li
The ICDAR2026 Competition on Multimodal Reasoning over Documents in Multiple Domains introduced a new Visual Question Answering benchmark that tests reasoning over documents from eight distinct domains such as business reports, scientific papers, and engineering drawings. Twenty valid submissions from eight teams were evaluated, featuring approaches ranging from zero‑shot vision‑language models to multi‑agent ensembles and fine‑tuned multimodal systems. Results indicate that the most effective systems employ structured evidence extraction, retrieval, verification, and orchestration across multiple components rather than single‑pass prompting.
By Artemis Llabr\'es, Marc Serra Ortega, Tom\`as Ockier, Samuel Ortega Cuadra, Amritpal Singh, Christos Georgakilas, Andrey Barsky, Ernest Valveny, Dimosthenis Karatzas
The paper introduces modality‑gated deep adapters, a parameter‑efficient method for adding new modalities to a frozen multimodal embedding language model without altering its existing outputs. These adapters are bottleneck modules attached to each decoder layer, grouped into modality‑specific packs that activate only during encoding of their own modality, ensuring exact preservation of the base model’s computation graph. Experiments on a 2B base model show significant gains in audio‑to‑text and thermal‑to‑text retrieval metrics, and the authors release the audio and thermal packs along with training and evaluation code.
By Abdul Basit Tonmoy, Kazi Fardinul Hoque, Md. Shahrier Islam Arham, Arman Luthra
Orthrus is a serving system that performs heterogeneous batching of embedding and generative models within a single inference loop. It uses chunked embedding with incremental pooling and workload‑aware batch composition to unify conflicting computational patterns. Experiments on four A100 GPUs show that Orthrus improves throughput by 1.28×–4.52× and reduces p99 latency by up to 55.8% compared to baseline deployments.
By Dohyun Park, Hubertus Franke, Daniel G. Waddington, Swaminathan Sundararaman, Yongjoo Park
The paper introduces Dynamic Tool Output Compression (DTOC), a framework that manages context in large‑language‑model agents by storing full tool outputs in external memory and inserting compact placeholders into the active context. DTOC treats context updates as explicit, reversible operations within the agent’s reasoning loop, allowing selective reconstruction of compressed outputs when needed. Experiments on the DeepSWE benchmark show that for responsive models such as Sonnet 4.6 and GPT‑5.4, DTOC reduces input tokens and agent steps while significantly improving solve rates and lowering cost per solved task, with ablation studies confirming the importance of reversibility for maintaining performance.
By Abhay Chaturvedi, Shreya Bhattacharya, Rashmika Gopalkrishnan, Peter van der Putten
AT3D-AD introduces a unified framework for detecting, localizing, and classifying 3D point‑cloud anomalies. It uses a Physics‑Driven Parametric Anomaly Synthesis module to generate synthetic defects for explicit supervision, a Hierarchical Global‑Local Anomaly Alignment module to refine representations, and a Semantic‑Geometric Anomaly Classification module to achieve precise, type‑discriminative localization. The method sets new state‑of‑the‑art results on four benchmarks, achieving high AUROC and Macro‑F1 scores.
By Jingyu Zeng, Haoquan Lu, Can Gao
The paper evaluates seven graph database engines, including Corvic AI, on a synthetic biomedical property graph with 1.02 million nodes and 5.34 million rows. It benchmarks query latency, bulk‑ingest throughput, point‑update latency, and correctness across a twenty‑query workload that covers neighborhood lookups, bounded paths, set intersections, anti‑joins, aggregation, ranking, temporal filters, full scans, and relational joins. The study finds that no single engine is universally fastest; performance depends on query shape, and the largest cost difference arises from bulk‑ingest throughput, which varies by three orders of magnitude and dominates total cost for workloads with fewer than about 10⁵ queries per data refresh.
By Donald Nguyen, Gurbinder Gill, Hadi Ahmadi, Christopher J. Rossbach
LIMIT (Less Is More for Instruction Tuning in Text-to-SQL) challenges the belief that large instruction corpora are necessary for effective Text-to-SQL models. The framework uses a four‑stage data‑centric process—difficulty‑aware filtering, chain‑of‑thought synthesis, LLM‑as‑judge quality scoring, and genetic algorithm optimization—to select a compact set of examples that still achieve full schema coverage. On the BIRD and Spider benchmarks, LIMIT’s 796 and 863 samples enable Qwen3‑8B to reach 69.1% and 88.9% execution accuracy, outperforming methods trained on twenty times more data and setting a new state‑of‑the‑art for open‑source approaches.
By Haoyuan Ma, Hengwei Liu, Linjuan Wu, Yongliang Shen, Weiming Lu
PreGS is a multi-expert graph neural network that uses parameter transfer from a pre‑trained multi‑head GAT to freeze GraphSAGE experts, creating complementary structural branches. The model fuses raw node features, GAT head outputs, and expert representations through an MLP, then combines the result with pretrained GAT logits. An extended version, PreGSv2, adds source‑level weighting and a structural gating mechanism for adaptive feature integration, and both variants outperform several baseline GNNs on eight public datasets.
By Zhicong Cai, Yinglong Zhang, Xiaoying Hong, Xuewen Xia, Xing Xu
The paper demonstrates that greedy decoding from large language models is not precision‑invariant: the same model, prompt, and decoding algorithm can produce different outputs when run in BF16 versus FP16 on identical hardware. Across six models (1.1B–7B parameters, four families, and 12B) and three benchmarks, 49–100 % of prompts diverge, with a single token flip often cascading into trajectory‑level divergence. The authors develop an empirical error‑propagation analysis that identifies the top‑two logit margin at the LM head as the key factor, and they propose a low‑overhead intervention—selective FP32 LM head recomputation—that improves exact agreement by 22–36 percentage points with less than 4 % latency overhead.
"whyItMatters":"The findings reveal that precision choices can fundamentally alter model outputs, challenging the assumption of deterministic greedy decoding and highlighting the need for precision‑aware inference strategies."
By Gaoyuan Du, Anam Nawaz Khan, Rex Zhou, Xiaoyang Liu, Deepayan Chakrabarti, Fnu Suya, Xueping Li
The paper investigates best‑policy identification in finite‑horizon, risk‑sensitive reinforcement learning using the entropic risk measure. It identifies a gap between known lower bounds ≥ η(e^{|eta|H}) and upper bounds ≤ O(e^{2|eta|H}) for sample complexity, attributing the excess factor to loose concentration bounds for exponential utilities. By employing a forward‑model algorithm with KL‑based exploration bonuses and a novel stopping rule, the authors achieve a sample complexity that matches the lower bound, closing the previously open exponential gap.
By Amer Essakine, Claire Vernade
SpanSynth-Edit is a flow‑matching model that enables MIDI‑guided synthesis and editing of multi‑instrument audio mixtures using low‑frame‑rate scalar‑quantised latents. It encodes instrument‑labelled note lifecycles as unordered event sets, pools them into a conditioning vector per audio‑latent frame, and uses contextual audio for instrument‑specific timbre guidance. The model supports editing by resynthesising target regions from revised MIDI and demonstrates competitive performance on single‑ and multi‑instrument benchmarks, including within‑frame onset control.
By Sungkyun Chang, Keshav Bhandari, Simon Dixon, Emmanouil Benetos
BELXTR is a new biomedical entity linking model that uses a multi‑vector (late interaction) architecture to preserve token‑level matching information, unlike traditional embedding‑based approaches that compress mentions into a single vector. By extending the XTR model with a task‑specific training objective and active query expansion, BELXTR achieves state‑of‑the‑art performance on half of ten evaluated corpora, with an average 5‑percentage‑point gain in recall@1. The model shows especially strong results on cross‑species gene disambiguation, outperforming an LLM‑powered retrieve‑and‑rerank pipeline and approaching a specialized rule‑based system.
By Samuele Garda, Ulf Leser