🔥 GateMem: Benchmarking Memory Governance in Multi-Principal Shared-Memory Agents

💡 The paper introduces GateMem, a benchmark for evaluating the performance of multi-principal shared-memory agents. The problem addressed is that current memory agents are not reliable in shared institutional settings, such as hospitals, workplaces, and households, where multiple users with different roles and authorization contexts access and update a common memory pool. The challenge is to balance utility, access control, and forgetting in these settings.

The GateMem benchmark is designed to evaluate memory agents in terms of their ability to provide utility for legitimate requests, control access to sensitive information, and forget information that is no longer needed or has been explicitly deleted. The benchmark consists of a set of tasks and evaluation metrics that span multiple domains, including medical, office, education, and household settings.

The results of the paper show that current memory agents are not able to achieve strong performance on all three aspects of utility, access control, and forgetting. The authors evaluated several baseline models and backbone architectures, and found that no method is able to simultaneously achieve high utility, robust access control, and reliable forgetting. The results also show that long-context prompting can achieve good governance scores, but at a high computational cost, while retrieval-based and external-memory methods can reduce the cost but may leak unauthorized or deleted information.

Overall, the paper highlights the need for more research on developing reliable multi-principal shared-memory agents that can balance utility, access control, and forgetting in shared institutional settings. The GateMem benchmark provides a useful tool for evaluating the performance of these agents and identifying areas for improvement.

📅 Published on Jun 17

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2606.18829
• PDF: https://arxiv.org/pdf/2606.18829
• Project Page: https://rzhub.github.io/GateMem/project.html

📊 Datasets citing this paper:
• https://huggingface.co/datasets/Ray368/GateMem

🚀 Spaces citing this paper:
• https://huggingface.co/spaces/Ray368/GateMem-Submit

━━━━━━━━━━━━━━━━━━━━━━━━
📢 By: https://xn--r1a.website/PaperNexus

#MultiPrincipalSystems #SharedMemoryAgents #MemoryGovernance #AccessControlMechanisms #BenchmarkingArtificialIntelligence

🔥 Unlimited OCR Works

💡 The paper introduces Unlimited OCR, a model designed to efficiently transcribe long documents by addressing the growing memory consumption issue in existing end-to-end OCR models. The problem with current models is that as the output sequence length increases, memory consumption also increases due to the accumulated KV cache, leading to slower generation. To solve this, the authors propose Reference Sliding Window Attention, a new attention mechanism that reduces computation costs while maintaining a constant KV cache throughout the decoding process. This mechanism is applied to the decoder of the DeepSeek OCR model, replacing all attention layers. The resulting Unlimited OCR model can transcribe dozens of pages of documents in a single forward pass, making it more efficient. The proposed attention mechanism is also general-purpose and can be applied to other tasks such as speech recognition and translation. The authors make their code and model weights publicly available, allowing others to build upon their work. Overall, the paper contributes to the development of more efficient OCR models that can handle long documents without a significant decrease in performance.

📅 Published on Jun 22

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2606.23050
• PDF: https://arxiv.org/pdf/2606.23050

🤖 Models citing this paper:
• https://huggingface.co/baidu/Unlimited-OCR
• https://huggingface.co/Yehor/Unlimited-OCR

🚀 Spaces citing this paper:
• https://huggingface.co/spaces/akhaliq/Unlimited-OCR
• https://huggingface.co/spaces/Dlcastro/composa-unlimited-ocr-bench

━━━━━━━━━━━━━━━━━━━━━━━━
📢 By: https://xn--r1a.website/PaperNexus

#OpticalCharacterRecognition #EndToEndOCR #DeepLearningForOCR #ReferenceSlidingWindowAttention #EfficientDocumentTranscription

🔥 PlanBench-XL: Evaluating Long-Horizon Planning of LLM Tool-Use Agents in Large-Scale Tool Ecosystems

💡 The paper introduces PlanBench-XL, a benchmark for evaluating the ability of large language model agents to plan and adapt in complex environments with limited visibility and dynamic disruptions. The problem addressed is that existing benchmarks rarely evaluate planning under retrieval-limited tool visibility, where agents must discover relevant tools and adapt to dynamic environments over long horizons. To address this gap, PlanBench-XL is an interactive benchmark consisting of 327 retail tasks and 1665 tools, testing whether agents can iteratively retrieve usable tools and invoke them to achieve a final goal. The benchmark also features a blocking mechanism that simulates real-world unpredictability by introducing missing, failing, or distracting tool functions, forcing agents to detect disrupted paths and adapt at runtime. The method used is to evaluate ten leading large language models on PlanBench-XL, with and without the blocking mechanism. The results show that massive-tool planning remains challenging, with the best performing model achieving 51.90% accuracy in block-free settings but collapsing to 11.36% under the most severe blocking condition. Further analysis reveals that agents are vulnerable when failures lack explicit error signals or when recovery requires longer alternative tool-use paths. The paper's contributions establish PlanBench-XL as a testbed for diagnosing agentic planning failures and highlight the need for robust adaptive planning in long-horizon tasks with large, imperfect tool environments.

📅 Published on Jun 21

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2606.22388
• PDF: https://arxiv.org/pdf/2606.22388
• Project Page: https://planbench-xl.github.io/

━━━━━━━━━━━━━━━━━━━━━━━━
📢 By: https://xn--r1a.website/PaperNexus

#LongHorizonPlanning #LLMToolUse #LargeScaleToolEcosystems #PlanningUnderUncertainty #ToolVisibilityEvaluation

🔥 MeshFlow: Mesh Generation with Equivariant Flow Matching

💡 MeshFlow is a method for generating triangle meshes directly using equivariant optimal-transport flow matching models. The problem of generating meshes is challenging due to the symmetries present in the representation, including permutation invariance of faces and vertices. Traditional autoregressive methods serialize meshes into long sequences, which can be slow and inefficient.

MeshFlow addresses this problem by learning to generate triangle meshes as triangle soups, which are unordered collections of triangles. The method uses equivariant optimal-transport flow matching models that respect the symmetries of triangle soups, including arbitrary permutations of faces and permutations of vertices within each face.

To achieve this, the authors propose a modification to the Diffusion Transformer architecture, resulting in a scalable network that can model a velocity field while maintaining the desired equivariance. The authors also introduce an optimal-transport-based training objective that improves convergence by eliminating supervision signals that violate these symmetries.

The results show that MeshFlow achieves mesh quality comparable to state-of-the-art autoregressive mesh generators, but provides a significant speedup of about 18 times during inference. This makes MeshFlow a more efficient and effective method for generating high-quality triangle meshes. Overall, the contributions of MeshFlow include a novel method for generating triangle meshes, a scalable and equivariant network architecture, and an optimal-transport-based training objective that improves convergence and mesh quality.

📅 Published on Jun 22

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2606.23489
• PDF: https://arxiv.org/pdf/2606.23489
• Project Page: https://qiisun.github.io/MeshFlow/

━━━━━━━━━━━━━━━━━━━━━━━━
📢 By: https://xn--r1a.website/PaperNexus

#EquivariantFlowMatching #MeshGeneration #OptimalTransportModels #TriangleMeshes #GeometricDeepLearning

🔥 DataClaw0: Agentic Tailoring Multimodal Data from Raw Streams

💡 The paper introduces a new paradigm called Agentic Data Tailoring which aims to structure high entropy multimodal data streams using learnable data processing. The problem addressed is that existing methods for processing unstructured multimodal data are costly and inefficient, failing to unlock the deep procedural logic embedded in the data. The proposed method uses a two stage pipeline, first generating a large scale dataset using generative semantic synthesis and deterministic factual anchors, and then training a model called DataClaw0-9B using supervised fine tuning and group relative policy optimization. The DataClaw0-9B model is able to achieve robust alignment with complex refinement and tailoring intents. The paper also introduces a new benchmark called DataClaw0-val for evaluating data refinement capabilities. The results show that the DataClaw0 model is able to deliver high information density tailored data, facilitating efficient model adaptation to new tasks with limited training data. The evaluations are done on tasks such as video generation, real world visual question answering, and GUI navigation, and the results confirm the effectiveness of the proposed method. Overall, the paper proposes a new paradigm for data processing and provides a method and benchmark for evaluating data refinement capabilities, with results showing the potential of the proposed method for efficient model adaptation and high quality data processing.

📅 Published on Jun 19

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2606.21337
• PDF: https://arxiv.org/pdf/2606.21337
• Project Page: https://czjdsg.github.io/MakeAnyData/#cases

━━━━━━━━━━━━━━━━━━━━━━━━
📢 By: https://xn--r1a.website/PaperNexus

#MultimodalDataProcessing #AgenticDataTailoring #LearnableDataProcessing #MultimodalStreamAnalysis #DeepProceduralLogicExtraction

🔥 EnterpriseClawBench: Benchmarking Agents from Real Workplace Sessions

💡 The paper introduces EnterpriseClawBench, a benchmark for evaluating enterprise agents based on real-world sessions. Enterprise agents are increasingly used in workspaces to read files, invoke tools, and deliver business artifacts. However, existing evaluation metrics are limited, focusing on single performance scores. To address this, the authors created EnterpriseClawBench, which consists of 852 reproducible tasks derived from a large archive of workplace sessions. Each task is paired with relevant information such as fixtures, prompts, and semantic rubrics.

The benchmark is not publicly released due to the proprietary nature of the data, but the construction and evaluation protocol is made available. The authors used this benchmark to evaluate the performance of various agent configurations and found that the best configuration achieved a score of 0.663, indicating that there is still significant room for improvement.

The key contribution of this paper is the introduction of a comprehensive evaluation protocol that goes beyond single performance scores. The authors argue that enterprise agent evaluation should consider multiple factors, including harness-model combinations, artifact delivery, visual quality, cost, runtime, and skill-transfer behavior. This approach provides a more nuanced understanding of an agent's capabilities and limitations, allowing for more effective evaluation and development of enterprise agents.

Overall, the paper highlights the need for more comprehensive evaluation metrics for enterprise agents and provides a benchmark and evaluation protocol to support this goal. The results demonstrate the challenges of developing effective enterprise agents and the importance of considering multiple factors in their evaluation.

📅 Published on Jun 22

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2606.23654
• PDF: https://arxiv.org/pdf/2606.23654
• Project Page: https://frontisai.github.io/EnterpriseClawBench/

━━━━━━━━━━━━━━━━━━━━━━━━
📢 By: https://xn--r1a.website/PaperNexus

#EnterpriseAgents #WorkplaceAutomation #BenchmarkingAI #ArtificialIntelligenceInBusiness #EnterpriseArtificialIntelligence

🔥 UniverSat: Resolution- and Modality-Agnostic Transformers for Earth Observation

💡 The paper introduces UniverSat, a new approach to applying Vision Transformers to Earth Observation data. The problem with current Vision Transformers is that they rely on rigid patch projectors, which makes it difficult to transfer them to Earth Observation tasks where the input data can vary widely in terms of modality, scale, and resolution. To address this issue, the authors propose a Universal Patch Encoder that can map patches from arbitrary spatial, spectral, and temporal resolutions, and from both optical and non-optical sensors, into a shared embedding space using a shared set of weights. This allows a single model to be trained on heterogeneous multimodal data using self-supervision, resulting in robust and sensor-agnostic spatial features. The authors validate their approach by achieving strong results on classification and segmentation tasks using standard Earth Observation benchmarks. The key contribution of UniverSat is its ability to enable resolution- and modality-agnostic spatial feature extraction, making it a versatile and effective tool for Earth Observation tasks. The authors make their code and models available for further research and development.

📅 Published on Jun 22

🔗 Links:
• GitHub: https://github.com/huggingface
• Project Page: https://huggingface.co/papers?q=patch%20projectors
• arXiv: https://arxiv.org/abs/2606.23503
• PDF: https://arxiv.org/pdf/2606.23503

🤖 Models citing this paper:
• https://huggingface.co/g-astruc/UniverSat

━━━━━━━━━━━━━━━━━━━━━━━━
📢 By: https://xn--r1a.website/PaperNexus

#EarthObservation #VisionTransformers #MultimodalLearning #RemoteSensing #GeospatialAnalysis

🔥 World Action Models: A Survey

💡 The paper World Action Models A Survey provides a comprehensive overview of World Action Models, which are predictive action systems that generate future states for decision making. These models balance representational richness against computational constraints, and recent developments have led to a blurring of boundaries among various related models. The survey aims to clarify these boundaries and provide a common account of the field.

The authors organize existing works into two complementary views. The first view examines what each method is required to generate, including rendered futures, latent futures, and video generation free action reasoning. The second view decomposes each method into its predictive substrate, backbone, action coupling, and deployment regime. This anatomy allows for a unified discussion of key aspects such as interactability, causality, persistence, physical plausibility, and generalization.

The survey reveals a consistent design pattern in World Action Models, where design choices trade representational richness against compute, memory, latency, and action label cost. The authors find that the field is moving towards methods that generate less of the future while preserving what is required for control. The survey provides a clear and unified account of the field, covering data, evaluation, and open challenges, and provides a foundation for future research in World Action Models.

The main contributions of the paper are to clarify the boundaries and definitions of World Action Models, to provide a comprehensive overview of existing works, and to identify a consistent design pattern in the field. The survey also highlights the key challenges and open issues in World Action Models, including the need for more efficient and effective methods that balance representational richness against computational constraints. Overall, the paper provides a valuable resource for researchers and practitioners in the field of World Action Models, and helps to advance the state of the art in predictive action systems.

📅 Published on Jun 18

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2606.20781
• PDF: https://arxiv.org/pdf/2606.20781
• Project Page: https://world-action-models.github.io/

━━━━━━━━━━━━━━━━━━━━━━━━
📢 By: https://xn--r1a.website/PaperNexus

#WorldActionModels #PredictiveActionSystems #DecisionMakingModels #ActionReasoning #ArtificialIntelligenceModels

🔥 Tmax: A simple recipe for terminal agents

💡 The paper presents a novel approach to training terminal agents using reinforcement learning, called Tmax. Terminal agents are a popular application of language models, but their training has been hindered by the lack of simple and effective methods, limited data, and challenging benchmarks. The authors address these issues by introducing a simplified recipe for training terminal agents, which achieves superior performance with fewer parameters than previous methods.

The method involves generating a large dataset of terminal environments using a novel taxonomy that combines difficulty control, personas, and verifier diversification. This allows for the cheap generation of large amounts of data, which is then used to train open-weight models using reinforcement learning with a simple outcome-only recipe.

The results show that Tmax achieves 27 percent on Terminal-Bench 2.0 with only 9 billion parameters, outperforming much larger models from prior work. The authors also release their terminal dataset, which is over 2.5 times larger than previously released terminal-agent datasets, as well as their models and code as a strong baseline for future academic work on terminal agents.

The contributions of the paper are threefold. First, it presents a simple and effective recipe for training terminal agents, which can be used as a baseline for future work. Second, it introduces a novel taxonomy for generating terminal environments, which allows for the cheap generation of large amounts of data. Third, it releases a large dataset of terminal environments, models, and code, which can be used by other researchers to advance the field of terminal agents. Overall, the paper provides a significant contribution to the field of terminal agents and reinforcement learning, and has the potential to advance the state of the art in this area.

📅 Published on Jun 22

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2606.23321
• PDF: https://arxiv.org/pdf/2606.23321
• Project Page: https://wai-org.com/blog/tmax/

🤖 Models citing this paper:
• https://huggingface.co/allenai/tmax-27b
• https://huggingface.co/allenai/tmax-9b
• https://huggingface.co/allenai/qwen35-9b-openthoughts

📊 Datasets citing this paper:
• https://huggingface.co/datasets/allenai/TMax-15K
• https://huggingface.co/datasets/allenai/tmax-15k-open-instruct
• https://huggingface.co/datasets/allenai/tmax-sft

━━━━━━━━━━━━━━━━━━━━━━━━
📢 By: https://xn--r1a.website/PaperNexus

#TerminalAgents #ReinforcementLearning #LanguageModels #TmaxAlgorithm #AgentTrainingMethods

🔥 Efficient Guided Generation for Large Language Models

💡 The paper presents an efficient method for guiding large language model text generation using regular expressions and context-free grammars. The problem addressed is that guided generation can be impractical due to significant overhead. The authors propose an approach that adds minimal overhead to the token sequence generation process. This method makes guided generation feasible in practice. The approach is implemented in the open source Python library Outlines, providing a practical solution for efficient guided generation. The results indicate that the method is effective, allowing for guided generation with little to no overhead, which is a significant contribution to the field of natural language processing.

📅 Published on Jul 19, 2023

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2307.09702
• PDF: https://arxiv.org/pdf/2307.09702

━━━━━━━━━━━━━━━━━━━━━━━━
📢 By: https://xn--r1a.website/PaperNexus

#LargeLanguageModels #GuidedTextGeneration #RegularExpressions #ContextFreeGrammars #EfficientGenerationMethods

🔥 OpenRath: Session-Centered Runtime State for Agent Systems

💡 The paper introduces OpenRath, a programming model for multi-agent systems that addresses the issue of fragmented runtime state. In current agent systems, various aspects such as transcripts, tool effects, and memory events are recorded separately, making it difficult to inspect or reproduce the system's behavior. OpenRath solves this problem by introducing a central runtime abstraction called Session, which is a first-class value that can be passed between agents and workflows.

The Session abstraction is designed to be branchable, inspectable, replayable, backend-aware, and composable, allowing it to record various execution state information such as conversation chunks, sandbox placement, and tool evidence. This enables explicit fork, merge, and replay operations as runtime operations rather than reconstructing states from external traces.

OpenRath also defines other key concepts such as Sandbox, Tool, Agent, Memory, Workflow, and Selector, which work together to provide a comprehensive programming model for multi-agent systems. The Selector is particularly important as it turns control flow into runtime-routed decisions.

The paper presents the programming model, architecture, and evidence protocol of OpenRath, and claims that the Session abstraction provides agent systems with a first-class runtime value for auditable composition. The results of this work are limited to controlled runtime properties, and further evaluation is needed to compare the performance of OpenRath with other systems and to assess its availability and quality.

Overall, OpenRath contributes a novel programming model for multi-agent systems that provides a unified and explicit way to manage runtime state, making it easier to inspect, reproduce, and debug the behavior of these systems.

📅 Published on Jun 17

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2606.19409
• PDF: https://arxiv.org/pdf/2606.19409
• Project Page: https://docs.openrath.com

━━━━━━━━━━━━━━━━━━━━━━━━
📢 By: https://xn--r1a.website/PaperNexus

#MultiAgentSystems #RuntimeStateManagement #AgentOrientedProgramming #SessionCenteredArchitecture #DistributedSystemDesign