6615 papers
GTR-Turbo is a highly efficient training method for multi-modal agents that eliminates the need for costly external teacher models by using merged checkpoints from ongoing reinforcement learning as a "free" teacher, thereby improving accuracy by 10–30% while reducing training time and compute costs by 50% and 60%, respectively.
This paper introduces the Pretrained Battery Transformer (PBT), a foundation model that leverages battery-knowledge-encoded mixture-of-experts layers to overcome data scarcity and heterogeneity, achieving state-of-the-art universal battery life prediction across diverse chemistries and conditions.
This paper presents a framework using YOLOv8 and Finer-CAM to achieve 96% accuracy in classifying seven European tree species from TLS 3D point clouds while demonstrating that the model's decisions are interpretable and primarily rely on crown features for most species, with stems playing a more significant role for others.
This paper introduces "Bayesian wind tunnels" to rigorously demonstrate that small transformers perform exact Bayesian inference through a specific geometric mechanism involving residual streams as belief substrates and attention-based routing, a capability that capacity-matched MLPs fundamentally lack.
This paper provides a first-order analysis demonstrating that cross-entropy training in transformers induces a coupled specialization of attention routing and value updates—functioning as a two-timescale EM procedure—that sculpts low-dimensional Bayesian manifolds, thereby explaining how gradient-based optimization enables precise probabilistic reasoning.
This paper demonstrates that production-grade language models preserve a low-dimensional geometric substrate, specifically an entropy-aligned axis in their last-layer value representations, which encodes Bayesian posterior structures and serves as a privileged readout for uncertainty, even though it is not a singular computational bottleneck for inference.
This paper systematically evaluates the phenomenon of "over-searching" in search-augmented large language models, where unnecessary tool invocation harms efficiency and accuracy, and proposes the Tokens Per Correctness (TPC) metric along with mitigation strategies to address this issue.
This paper proposes a Secure Multi-Tenant Architecture (SMTA) combined with a novel Burn-After-Use (BAU) mechanism to effectively prevent data leakage in enterprise LLMs by enforcing strict instance isolation and ephemeral context destruction, achieving high defense success rates against both semantic leakage attacks and post-session persistence threats in experimental evaluations.
This paper introduces a stealthy, multi-turn denial-of-service attack on LLM agents via the Model Context Protocol (MCP) that exploits tool-calling chains to amplify computational costs and resource consumption by up to 658 times while evading standard detection mechanisms.
This paper proposes a framework for lifelong learning in complex real-time environments like Dark Souls III by decomposing control into a directed skill graph of reusable components, which enhances sample efficiency and enables rapid adaptation to new challenges through selective fine-tuning of only the necessary skills.
MemOCR is a multimodal memory agent that enhances long-horizon reasoning under tight context budgets by converting structured rich-text history into a visually compressed image, allowing the agent to prioritize crucial evidence through layout-aware information density while aggressively reducing low-value details.
This paper introduces MHDash, an open-source platform that enables fine-grained, multi-turn evaluation of AI assistants for mental health support, revealing that conventional aggregate metrics fail to capture critical safety risks and that models often struggle specifically with high-risk cases despite strong overall performance.
This paper establishes a rate-distortion theorem demonstrating that hallucinations in large language models are an inevitable consequence of information-theoretic optimal memory compression when storing sparse facts, forcing the model to confidently assign high scores to non-facts rather than abstain.
This paper introduces EverMemBench, the first benchmark designed to evaluate long-horizon memory in multi-party collaborative dialogues, revealing that current LLM systems struggle with multi-hop reasoning, temporal versioning, and implicit relevance retrieval in realistic, complex interaction scenarios.
This paper introduces DEFT, a diffusion-based trajectory generator that enables robots to achieve fail-active operation by successfully completing tasks under arbitrary actuation failures, outperforming classical methods in both simulation and real-world scenarios while demonstrating robust zero-shot generalization.
The paper proposes DMS2F-HAD, a dual-branch Mamba-based network that efficiently fuses spatial and spectral features to achieve state-of-the-art accuracy and significantly faster inference speeds for hyperspectral anomaly detection across multiple benchmark datasets.
This paper introduces Fine-grained Group Policy Optimization (FGO), a reinforcement learning algorithm that effectively compresses verbose Chain-of-Thought reasoning in Large Language Models while simultaneously addressing the data inefficiency and entropy collapse limitations of Group Relative Policy Optimization (GRPO).
UniWeTok is a unified binary tokenizer featuring a massive codebook, a convolution-attention hybrid architecture with SigLu activation, and a novel three-stage training framework that achieves state-of-the-art performance in image generation and multimodal understanding with significantly lower computational costs than existing models.
TikArt is a reinforcement learning-trained multimodal agent that stabilizes fine-grained visual reasoning by employing a Think-Aperture-Observe loop to sequentially acquire and linguistically encode evidence through zooming and segmentation, thereby overcoming the limitations of single-pass global image encoding.
The paper proposes GOT-JEPA, a model-predictive pretraining framework that learns to predict robust tracking models from corrupted observations to improve generalization, and introduces OccuSolver to enhance occlusion handling through iterative, object-aware visibility estimation.