Learning GenAI via SOTA Papers - Video

Title: MUSE-Autoskill: Self-Evolving Agents via Skill Creation, Memory, Management, and EvaluationSource: http://arxiv.org/abs/2605.27366v1Summary:This paper proposes a novel architectural framework for self-evolving agents that can autonomously create, store, and refine a library of reusable skills through a unified lifecycle management system. It introduces the concept of skill-level memory and unit-testable assets, representing a major advancement in building agents capable of continuous improvement and cross-task experience accumulation.

Title: Innovation: An Almost Characterization of HallucinationSource: http://arxiv.org/abs/2605.26808v1Summary:This work establishes a foundational probabilistic framework that formalizes hallucination as "innovation," providing a mathematical characterization of why LLMs produce outputs outside their training data. By deriving new lower bounds on hallucination rates based on "missing mass," it offers a critical theoretical breakthrough for understanding and mitigating the core reliability limits of generative models.

Title: Multi-Agent Coordination Adaptation via Structure-Guided OrchestrationSource: http://arxiv.org/abs/2605.25746v1Summary:The paper introduces a novel probabilistic framework for multi-agent coordination by casting orchestration as posterior inference over task-specific structures. This foundational approach balances structural stability with dynamic adaptability, significantly improving execution efficiency and performance while reducing token overhead in complex agentic workflows.

Title: How Should LLMs Consume High-Quality Data? Optimal Data Scheduling via Quality-Aware Functional Scaling LawsSource: http://arxiv.org/abs/2605.25698v1Summary:This paper establishes foundational quality-aware functional scaling laws that provide the first theoretical closed-form solution for scheduling high-quality data during LLM training. The introduced 'Drop-Stable-Rampup' schedule optimizes training dynamics across noise-limited and signal-limited regimes, yielding significant breakthroughs in mathematical reasoning performance.

Title: Efficient Agentic Reasoning Through Self-Regulated Simulative PlanningSource: http://arxiv.org/abs/2605.22138v1Summary:This paper introduces a foundational three-system reasoning framework—comprising reactive, simulative, and self-regulated components—that enables agents to autonomously manage their planning depth and horizon. By treating the LLM as a world model for future-state prediction, it demonstrates that structured deliberation can allow smaller models to match the performance of systems orders of magnitude larger.

Title: Compiling Agentic Workflows into LLM Weights: Near-Frontier Quality at Two Orders of Magnitude Less CostSource: http://arxiv.org/abs/2605.22502v1Summary:This work proposes the 'subterranean agent' paradigm, which replaces external orchestration frameworks by compiling agentic workflows directly into the model's weights via fine-tuning. This foundational shift addresses the cost and latency bottlenecks of frontier-model prompting while providing a more efficient and private alternative for procedural task execution.

Title: Mem-π: Adaptive Memory through Learning When and What to GenerateSource: http://arxiv.org/abs/2605.21463v1Summary:Mem-π presents a foundational shift in agent memory architectures by replacing static similarity-based retrieval with a dedicated generative model that produces context-specific guidance. This framework enables agents to dynamically adapt their memory usage, leading to substantial improvements in complex reasoning and long-horizon task execution.

Title: Agent JIT Compilation for Latency-Optimizing Web Agent Planning and SchedulingSource: http://arxiv.org/abs/2605.21470v1Summary:This paper introduces Agent Just-In-Time (JIT) compilation, a novel architectural primitive that transforms natural language task descriptions into optimized, executable code plans. It represents a significant breakthrough in agentic efficiency by replacing traditional sequential loops with a compiled, parallelized execution framework that drastically reduces latency.

Title: PEEK: Context Map as an Orientation Cache for Long-Context LLM AgentsSource: http://arxiv.org/abs/2605.19932v1Summary:This work introduces 'context maps' as a novel architectural primitive for long-context agents, enabling them to cache and maintain structured orientation knowledge about recurring external datasets. By implementing a programmable cache policy for distilling and translating inference-time signals, it significantly improves efficiency and accuracy across multi-turn reasoning workloads.

Title: Formal Skill: Programmable Runtime Skills for Efficient and Accurate LLM AgentsSource: http://arxiv.org/abs/2605.19604v1Summary:This work introduces a foundational architectural primitive for agents that replaces informal natural-language instructions with programmable, stateful runtime skills governed by hook policies and action schemas. This shift from prompting to executable state machines provides a more enforceable and token-efficient control surface for reliable agentic workflows in real-world environments.

Title: A Measure-Theoretic Analysis of Reasoning: Structural Generalization and Approximation LimitsSource: http://arxiv.org/abs/2605.19944v1Summary:This paper establishes fundamental theoretical bounds for LLM reasoning, proving that scaling physical layer depth is a non-negotiable requirement for out-of-distribution generalization that cannot be bypassed by scaling width. It also formalizes why specific architectural choices, such as shift-invariant embeddings, are mathematically necessary to maintain reasoning equivariance across domain shifts.

Title: Learning to Hand Off: Provably Convergent Workflow Learning under Interface ConstraintsSource: http://arxiv.org/abs/2605.19140v1Summary:This research provides the first finite-sample guarantee for neural Q-learning in decentralized multi-agent settings, a foundational breakthrough for reliable agentic workflow learning. By formalizing handoffs as interface-constrained SMDPs, it enables provably convergent learning in complex LLM pipelines where agents have restricted observability.

Title: Generating Pretraining Tokens from Organic Data for Data-Bound ScalingSource: http://arxiv.org/abs/2605.17849v1Summary:This work addresses the transition of LLM pretraining into data-bound regimes by introducing a synthetic data generation framework that maximizes the utility of limited organic datasets. It represents a significant breakthrough in scaling laws, demonstrating how to unlock up to 5x more effective tokens through model-aware rephrasing and reformatting.

Title: EXG: Self-Evolving Agents with Experience GraphsSource: http://arxiv.org/abs/2605.17721v1Summary:This paper introduces the first experience graph framework for self-evolving agents, providing a structured relational representation for successes and failures that enables real-time experience reuse. It establishes a principled foundation for scalable agent behavior by allowing behaviorally static agents to systematically improve through structured memory.

Title: Parallelizing Counterfactual Regret MinimizationSource: http://arxiv.org/abs/2605.14277v1Summary:This work introduces a generalized framework that reframes counterfactual regret minimization as linear algebra operations, allowing for massive parallelization on modern hardware. By achieving a four-order-of-magnitude speedup, it provides a foundational efficiency breakthrough for the reasoning algorithms central to strategic decision-making in complex environments.

Title: Orchard: An Open-Source Agentic Modeling FrameworkSource: http://arxiv.org/abs/2605.15040v1Summary:Orchard provides a scalable open-source framework for agentic modeling, introducing reusable environment primitives and training recipes that enable LLMs to achieve state-of-the-art performance on complex tasks. It addresses critical gaps in agent infrastructure by standardizing sandbox management and introducing credit-assignment SFT for learning from unresolved trajectories.

Title: Beyond Individual Intelligence: Surveying Collaboration, Failure Attribution, and Self-Evolution in LLM-based Multi-Agent SystemsSource: http://arxiv.org/abs/2605.14892v1Summary:This work introduces the LIFE progression framework, which formally characterizes the causal dependencies between agent foundation, collaboration, failure attribution, and autonomous self-evolution. It establishes a foundational conceptual roadmap for building self-organizing multi-agent systems that can continuously diagnose and refine their own collective intelligence.

Title: Agentifying Patient Dynamics within LLMs through Interacting with Clinical World ModelSource: http://arxiv.org/abs/2605.14723v1Summary:This work presents a novel world-model-augmented agentic reasoning loop that utilizes a 'propose-simulate-refine' workflow to ground LLM decisions in action-conditioned dynamics. It demonstrates how integrating world models with agentic reinforcement learning can significantly improve decision-making safety and efficacy in complex environments.

Title: MAP: A Map-then-Act Paradigm for Long-Horizon Interactive Agent ReasoningSource: http://arxiv.org/abs/2605.13037v1Summary:MAP proposes a paradigm shift for interactive agents by establishing environmental understanding through structured cognitive mapping before task execution. This approach overcomes the epistemic bottlenecks and inefficient failure cycles inherent in traditional reactive, goal-conditioned stepwise planning.

Title: Harnessing Agentic EvolutionSource: http://arxiv.org/abs/2605.13821v1Summary:AEvo introduces a meta-editing framework that treats the evolution context as a process-level state, allowing agents to iteratively refine their own procedures. This shifts agentic evolution from rigid hand-designed loops to a unified interface for actionable, long-horizon self-improvement.

Title: SAGE: A Self-Evolving Agentic Graph-Memory Engine for Structure-Aware Associative MemorySource: http://arxiv.org/abs/2605.12061v1Summary:SAGE introduces a self-evolving graph-memory engine that couples a memory writer with a Graph Foundation Model-based reader to create a dynamic, self-improving long-term memory substrate. This framework is foundational for its architectural move beyond static RAG, enabling agents to autonomously refine their structure-aware associative memory through downstream feedback.

Title: On-Policy Self-Evolution via Failure Trajectories for Agentic Safety AlignmentSource: http://arxiv.org/abs/2605.11882v1Summary:FATE establishes a foundational framework for on-policy self-evolution by transforming agentic failure trajectories into high-density repair supervision without human demonstrations. By employing Pareto-Front Policy Optimization, it provides a scalable architectural primitive for agents to autonomously balance safety and utility across long-horizon tool-use tasks.

Title: Goal-Oriented Reasoning for RAG-based Memory in Conversational Agentic LLM SystemsSource: http://arxiv.org/abs/2605.12213v1Summary:This paper presents Goal-Mem, a framework that employs backward chaining and Natural Language Logic to create a goal-oriented reasoning loop for agentic memory systems. It provides a foundational advancement in how agents can systematically decompose complex queries and retrieve missing intermediate facts for robust multi-hop reasoning.

Title: The Bystander Effect in Multi-Agent Reasoning: Quantifying Cognitive Loafing in Collaborative InteractionsSource: http://arxiv.org/abs/2605.10698v1Summary:This study formalizes the 'Bystander Effect' in multi-agent systems, identifying a critical failure mode where agents subjugate independent reasoning to social compliance. It introduces the Interaction Depth Limit and Sovereignty Gap as foundational architectural constraints for designing robust and independent multi-agent reasoning topologies.

Title: PIVOT: Bridging Planning and Execution in LLM Agents via Trajectory RefinementSource: http://arxiv.org/abs/2605.11225v1Summary:PIVOT introduces a novel self-supervised framework that treats agent trajectories as optimizable objects refined through iterative environment feedback, bridging the gap between high-level planning and execution. This methodology establishes a principled approach to trajectory optimization that enhances both constraint satisfaction and computational efficiency in autonomous systems.

Title: DeepRefine: Agent-Compiled Knowledge Refinement via Reinforcement LearningSource: http://arxiv.org/abs/2605.10488v1Summary:DeepRefine establishes a general reinforcement learning framework for the autonomous refinement of agent-compiled knowledge bases using abductive diagnosis and a novel Gain-Beyond-Draft reward. It provides a foundational reasoning loop for maintaining persistent, high-fidelity external knowledge, which is essential for long-term agentic performance in knowledge-intensive tasks.

Title: LEAD: Length-Efficient Adaptive and Dynamic Reasoning for Large Language ModelsSource: http://arxiv.org/abs/2605.09806v1Summary:LEAD establishes a foundational reinforcement learning mechanism for reasoning models that dynamically calibrates the balance between correctness and verbosity at each training step. It solves the critical issue of 'overthinking' in modern reasoning models by introducing online, per-problem length estimation, paving the way for more efficient and scalable reasoning architectures.

Title: Do Self-Evolving Agents Forget? Capability Degradation and Preservation in Lifelong LLM Agent AdaptationSource: http://arxiv.org/abs/2605.09315v1Summary:This paper introduces the 'capability erosion' framework to quantify how autonomous self-evolution can degrade an agent's prior knowledge across workflows and models. It proposes Capability-Preserving Evolution (CPE) as a necessary architectural constraint for building stable, lifelong learning agents that can adapt to new tasks without catastrophic forgetting.

Title: Tools as Continuous Flow for Evolving Agentic ReasoningSource: http://arxiv.org/abs/2605.07339v1Summary:FlowAgent reconceptualizes agentic reasoning by replacing discrete, step-wise tool orchestration with continuous trajectory generation using conditional flow matching. This foundational framework provides theoretical guarantees for error attenuation and global planning, representing a significant shift in how agents execute long-horizon reasoning tasks.

Title: Memory-Efficient Looped Transformer: Decoupling Compute from Memory in Looped Language ModelsSource: http://arxiv.org/abs/2605.07721v1Summary:This paper introduces a novel architectural primitive that decouples reasoning depth from memory consumption in looped language models, enabling constant-memory iterative reasoning. By sharing a single KV cache across loops via a learnable gating mechanism, it provides a foundational efficiency breakthrough for models performing multi-step computation in embedding space.

Title: Logic-Regularized Verifier Elicits Reasoning from LLMsSource: http://arxiv.org/abs/2605.05893v1Summary:This work presents a novel reasoning framework that uses logical consistency rules to regularize unsupervised verifiers, eliminating the need for expensive supervised datasets. By treating verification as a binary latent variable problem, it achieves performance comparable to supervised models in eliciting complex reasoning from off-the-shelf LLMs.

Title: HaM-World: Soft-Hamiltonian World Models with Selective Memory for PlanningSource: http://arxiv.org/abs/2605.05951v1Summary:This paper introduces a foundational architectural primitive for world models by combining Hamiltonian geometric structures with Mamba-based selective memory to stabilize long-horizon planning. It provides agents with a structured latent state for dynamics, rewards, and action search, significantly improving robustness in out-of-distribution planning tasks.

Title: Uno-Orchestra: Parsimonious Agent Routing via Selective DelegationSource: http://arxiv.org/abs/2605.05007v1Summary:This paper introduces a novel orchestration policy that jointly optimizes task decomposition and agent routing, establishing a new frontier for efficiency and accuracy in multi-agent systems. It moves beyond rigid workflows by learning selective delegation from RL trajectories to achieve high performance at an order of magnitude lower cost.

Title: Gyan: An Explainable Neuro-Symbolic Language ModelSource: http://arxiv.org/abs/2605.04759v1Summary:Gyan proposes a breakthrough non-transformer architecture that decouples language modeling from knowledge representation to eliminate hallucinations and drastically reduce compute requirements. It introduces a neuro-symbolic framework that mimics human compositional context, offering a more trustable and efficient foundational primitive for AI development.

Title: ScrapMem: A Bio-inspired Framework for On-device Personalized Agent Memory via Optical ForgettingSource: http://arxiv.org/abs/2605.03804v1Summary:ScrapMem introduces a novel on-device memory architecture for agents that employs bio-inspired 'Optical Forgetting' to maintain long-term multimodal context with extreme storage efficiency. It establishes a foundational framework for personalized agentic reasoning on resource-constrained edge devices through structured episodic memory management.

Title: Parallel Prefix Verification for Speculative GenerationSource: http://arxiv.org/abs/2605.04263v1Summary:This paper introduces PARSE, a novel speculative generation primitive that enables semantic-level verification across multiple prefixes in a single forward pass. By eliminating sequential bottlenecks in speculative decoding, it achieves up to 4.3x throughput gains, representing a major efficiency breakthrough for frontier LLM inference.

Title: OpenSeeker-v2: Pushing the Limits of Search Agents with Informative and High-Difficulty TrajectoriesSource: http://arxiv.org/abs/2605.04036v1Summary:This paper establishes a high-efficiency paradigm for training frontier search agents using only supervised fine-tuning on high-quality synthesized trajectories, challenging resource-intensive industry standards. It provides a foundational methodology for achieving state-of-the-art agentic reasoning and search capabilities with significantly reduced computational requirements.

Title: Awaking Spatial Intelligence in Unified Multimodal Understanding and GenerationSource: http://arxiv.org/abs/2605.04128v1Summary:JoyAI-Image establishes a new foundational architecture for multimodal agents by tightly coupling a spatially enhanced MLLM with a Multimodal Diffusion Transformer through a shared interface. This unified primitive enables a bidirectional feedback loop between visual perception and controllable generation, advancing the development of spatially-aware world models.

Title: The Compliance Trap: How Structural Constraints Degrade Frontier AI Metacognition Under Adversarial PressureSource: http://arxiv.org/abs/2605.02398v1Summary:This work identifies the 'Compliance Trap,' a fundamental failure mode where alignment constraints cause catastrophic metacognitive collapse in frontier models under pressure. It provides a foundational framework for understanding and evaluating the structural limits of AI reasoning and stability.

Title: MAGE: Safeguarding LLM Agents against Long-Horizon Threats via Shadow MemorySource: http://arxiv.org/abs/2605.03228v1Summary:MAGE introduces the 'shadow memory' abstraction, a novel defensive framework that maintains a safety-focused agentic memory to counter long-horizon threats. It establishes a new paradigm for agentic safety by enabling models to proactively assess risk across extended execution trajectories.

Title: GRAIL: A Deep-Granularity Hybrid Resonance Framework for Real-Time Agent Discovery via SLM-Enhanced IndexingSource: http://arxiv.org/abs/2605.02489v1Summary:GRAIL introduces a novel SLM-enhanced indexing and resonance framework that solves the foundational scaling bottleneck of agent discovery in large-scale ecosystems. By achieving a 79x reduction in latency, it provides a critical architectural primitive for enabling a real-time 'Internet of Agents.'

Title: ARISE: A Repository-level Graph Representation and Toolset for Agentic Fault Localization and Program RepairSource: http://arxiv.org/abs/2605.03117v1Summary:ARISE introduces a multi-granularity program graph and data-flow slicing as a first-class, queryable agent primitive for repository-level fault localization. This architectural primitive significantly improves agentic reasoning for complex program repair by allowing models to trace semantic dependencies across entire codebases.

Title: Talk is Cheap, Communication is Hard: Dynamic Grounding Failures and Repair in Multi-Agent NegotiationSource: http://arxiv.org/abs/2605.01750v1Summary:This research identifies 'dynamic grounding' as a foundational and critical axis for multi-agent coordination, revealing it as a primary bottleneck that exceeds individual reasoning or simple information exchange. It proposes a novel framework for multi-turn repair and joint plan formation, offering a new paradigm for designing robust agentic reasoning loops in collaborative settings.

Title: Sheaf-Theoretic Planning: A Categorical Foundation for Resilient Multi-Agent Autonomous SystemsSource: http://arxiv.org/abs/2605.01879v1Summary:This paper introduces Sheaf-Theoretic Planning (STP) as a transformative architectural primitive that replaces traditional monolithic logical models with a foundation in topos theory and sheaf semantics. It establishes a novel mathematical framework for resilient multi-agent coordination, specifically designed to handle divergent belief states and unobserved interventions in stochastic environments.

Title: SciResearcher: Scaling Deep Research Agents for Frontier Scientific ReasoningSource: http://arxiv.org/abs/2605.01489v1Summary:Introduces a novel paradigm for automated data construction to scale agentic reasoning, achieving state-of-the-art results with an 8B foundation model. It bridges the gap in frontier science by synthesizing conceptual and computational tasks grounded in academic evidence.

Title: Lifting Traces to Logic: Programmatic Skill Induction with Neuro-Symbolic Learning for Long-Horizon Agentic TasksSource: http://arxiv.org/abs/2605.01293v1Summary:Proposes a neuro-symbolic framework that lifts interaction traces into modular, logic-grounded programs to solve long-horizon planning challenges. This paradigm enables agents to induce skills from few-shot examples and adapt to dynamic environments by discovering the underlying logic of actions.

Title: SAGA: Workflow-Atomic Scheduling for AI Agent Inference on GPU ClustersSource: http://arxiv.org/abs/2605.00528v1Summary:SAGA represents a foundational breakthrough in agentic AI systems by transitioning from request-level to workflow-atomic scheduling for GPU inference. By capturing and optimizing for the chained structure of agentic tasks, it significantly reduces latency and resource overhead, enabling the scaling of complex, multi-step AI agents.

Title: Position: agentic AI orchestration should be Bayes-consistentSource: http://arxiv.org/abs/2605.00742v1Summary:This position paper establishes a novel theoretical foundation for agentic AI by arguing that the orchestration layer must be Bayes-consistent to handle uncertainty in complex decision-making. It provides a formal framework for belief-updating and utility-aware action selection, which are essential primitives for reliable agentic reasoning loops.

Title: PRTS: A Primitive Reasoning and Tasking System via Contrastive RepresentationsSource: http://arxiv.org/abs/2604.27472v1Summary:PRTS establishes a new foundation model paradigm by reformulating VLA pretraining as Goal-Conditioned Reinforcement Learning to learn a unified goal-reachability embedding space. This architectural primitive bridges the gap between semantic reasoning and physical execution, enabling more robust long-horizon planning for autonomous agents.

Title: ObjectGraph: From Document Injection to Knowledge Traversal -- A Native File Format for the Agentic EraSource: http://arxiv.org/abs/2604.27820v1Summary:ObjectGraph introduces a new native file format that reconceives linear documents as typed knowledge graphs specifically optimized for autonomous agent traversal rather than human reading. This foundational shift in data representation addresses the token-inefficient mismatch between agents and traditional formats, enabling a more scalable infrastructure for agentic AI.