Tool use, function calling, orchestration and the protocols that let models act rather than only answer.
arXiv:2604.08567v3 Announce Type: replace Abstract: Large language models (LLMs) and LLM-based agents are increasingly deployed as assistants in planning and decision making, yet most existing system...
arXiv:2605.09252v3 Announce Type: replace Abstract: Tool-augmented LLM agents tend to call tools indiscriminately, even when the model can answer directly. Each unnecessary call wastes API fees and l...
arXiv:2606.18902v3 Announce Type: replace Abstract: Context engineering has emerged as a primary lever for improving AI systems without parameter updates. Recent work showing that textual gradients d...
arXiv:2611.06083v1 Announce Type: new Abstract: Omni models answer questions about videos by jointly processing sampled frames and audio in one turn, with progress driven largely by scaling model siz...
arXiv:2611.06302v1 Announce Type: new Abstract: Fine-grained action understanding requires vision-language models to ground what an agent does, how it acts, and why it acts in spatio-temporal evidenc...
arXiv:2611.06635v1 Announce Type: new Abstract: Unmanned aerial vehicle (UAV)-based collaborative perception enables multiple agents to share complementary perceptual information, overcoming limitati...
arXiv:2611.06701v1 Announce Type: new Abstract: Deep learning methods for pathology are constrained by the limited availability and high cost of large-scale pixel-level annotations. Synthetic data ca...
arXiv:2611.06453v1 Announce Type: cross Abstract: Closed-loop driving benchmarks in CARLA rank agents by a driving score that multiplies route completion by an infraction penalty. This score does not...
arXiv:2511.17481v2 Announce Type: replace Abstract: World models learn to predict the temporal evolution of visual observations given a control signal, potentially enabling agents to reason about env...
arXiv:2609.34378v2 Announce Type: replace Abstract: Humans naturally possess the ability to work persistently toward long-term goals. Given a challenging task, humans can continuously work for months...
The paper investigates the role of thought supervision in distilling large language model agents. By independently perturbing teacher thoughts and actions during distillation, the authors find that removing or altering thoughts has little impact on student performance across various benchmarks and model scales. However, thought supervision becomes important when the action does not reveal the information contained in the thought, suggesting that including thoughts in the training data may be unnecessary for many cases.
The paper introduces GraphRail, a verification framework for search-augmented reasoning systems that checks whether evidence covers the question, reasoning addresses the question, and reasoning is grounded in the evidence. GraphRail represents these components as structured claims and compares them without relying on a generative judge or gold answer. Experiments on three multi‑hop QA benchmarks and three reasoning frameworks show that GraphRail improves error recall over generative verifiers and remains stable under stochastic extraction, while also providing diagnoses that can guide online reasoning to detect and correct intermediate errors.
EmoReward is a self‑evolving reward calibration framework for emotional video generation using reinforcement learning. It distills reliability from reward traces during training to adaptively weight multiple reward primitives—such as visual quality, motion, and text alignment—based on emotional evidence. The method employs grounding cues, heterogeneous signal primitives, reflective reward gating, and a RewardMemory module to calibrate fusion weights across training runs, demonstrating improvements on HunyuanVideo 1.5 and transferability to SD 3.5 and FLUX.
The paper presents a VR exergame that gamifies High‑Intensity Interval Training (HIIT) using 11 predefined combo moves and a standard smartphone camera on a tripod for full‑body pose estimation. Real‑time biomechanical data are processed by a backend built on the Model Context Protocol and GPT‑4o to provide personalized feedback on form, efficiency, intensity, and injury risk. A within‑subject study with 20 participants showed higher enjoyment, immersion, and recommendation likelihood in VR sessions compared to non‑VR, while the MCP system’s feedback agreed strongly with expert evaluation.
The paper introduces COL‑Tree, a compact landmark‑based tree structure designed to accelerate hierarchical object search in spatial networks. It supports diverse query types—such as aggregate k‑nearest neighbor (AkNN), k‑farthest neighbor (kFN), and others—by enabling efficient graph traversal with a more accurate heuristic than Euclidean‑based methods. Experiments on real‑world and synthetic road, terrain, and physics mesh networks show that COL‑Tree outperforms existing approaches by up to three orders of magnitude, with only modest preprocessing overhead.
SecureGARD is a dynamic ontology‑guided governance framework designed to secure tool‑using large language model agents. It grounds heterogeneous tool calls in a shared semantic representation, using an executable ontology combined with episode‑level task state to make deterministic ready, waiting, and forbidden decisions. The system evolves its ontology offline from verified training executions, ensuring logical consistency and protecting entailments, and achieves a 0% attack success rate across multiple benchmarks while improving task utility under attack.
LongBanana is the first expert-designed and expert-verified benchmark for long-context multi-reference image synthesis, featuring 90 tasks each with about 32 reference images, 20 constraints, and 50 evaluation items, totaling 4,670 atomic evaluation items. The benchmark uses a graph-grounded evaluation design that converts entity-relation graph obligations into checklist items for Content Fidelity, Relational Binding, and Global Constraint Satisfaction. An accompanying agent-harness framework, LongBanana-Gen, demonstrates that minimal harnesses improve generation quality, with a ReAct-Plan-Exec-Reflect workflow further enhancing evidence grounding and repair.
The paper presents a comprehensive evaluation of 11 memory substrates for long‑horizon LLM agents, using a unified set of 26 performance and efficiency metrics across three backbone models and four benchmark suites. It finds that no single substrate consistently outperforms others; retrieval‑heavy substrates aid long‑context factual QA but can impair sequential decision‑making, and scalability can shift performance trade‑offs. The study demonstrates that selecting a memory substrate tailored to the operating regime yields better results than any fixed choice, offering empirical guidance for designing efficient, reliable, and regime‑aware long‑term memory systems.
LITMUS is a benchmark for testing the safety of large language model agents in real operating system environments, focusing on behavior jailbreaks that can cause irreversible OS-level harm. It includes 819 high-risk test cases organized into a harmful seed subset and six attack-extended subsets covering jailbreak speaking, skill injection, and entity wrapping, and employs a semantic-physical dual verification mechanism with OS-level state rollback. Evaluation of frontier agents shows that even strong models can execute dangerous operations, exhibit execution hallucination, and are vulnerable to skill injection and entity wrapping attacks.
MKD2World introduces Typed Spatial Markdown (TSM), a Markdown-style textual representation that serializes desktop web interface elements by type, visual semantics, location, and hierarchy. Using the AtomBlock-WebUI dataset, a screenshot-to-TSM renderer is trained, enabling a two-stage world model that predicts current and next states for GUI agents. The resulting MKD2World-9B model outperforms GPT-5.6-Sol on EleF1 and boosts the success rate of a Qwen3.5-35B web agent on Online-Mind2Web tasks.