Title: Hive: A Multi-Agent Infrastructure for Algorithm- and Task-Level Scaling

Source: http://arxiv.org/abs/2604.17353v1

Summary:

Hive provides a foundational infrastructure for scaling agentic systems at the algorithm and task levels, introducing key mechanisms like Logits Cache and Agent-Aware Scheduling to optimize multi-agent resource allocation. It represents a significant breakthrough in efficiency for complex agent workflows, enabling the systematic scaling of inference-time computation and task decomposition.

Title: Compiling Deterministic Structure into SLM Harnesses

Source: http://arxiv.org/abs/2604.17450v1

Summary:

This paper introduces Semantic Gradient Descent (SGDe), a novel PAC-learning-based framework that compiles complex agentic workflows into optimized, deterministic execution plans for small language models. It establishes a new architectural primitive for agent design by leveraging frontier teacher models to iteratively refine student workflows via discrete semantic gradients.

Title: The Topological Trouble With Transformers

Source: http://arxiv.org/abs/2604.17121v1

Summary:

This research identifies a fundamental topological limit in the Transformer architecture's ability to track dynamic states due to its feedforward nature, which exhausts model depth as states evolve. It advocates for a shift towards recurrent or continuous-thought architectures as essential primitives for achieving the temporally extended cognition required for advanced AI agents.

Title: The Consensus Trap: Rescuing Multi-Agent LLMs from Adversarial Majorities via Token-Level Collaboration

Source: http://arxiv.org/abs/2604.17139v1

Summary:

This paper introduces Token-Level Round-Robin Collaboration, a novel multi-agent interaction paradigm that interleaves token generation to overcome the brittle aggregation of majority voting. It provides a mathematical proof that this non-linear approach creates a dynamic chain of logic more robust to adversarial corruption and intermediate reasoning failures than traditional response-level methods.

Title: Neurosymbolic Repo-level Code Localization

Source: http://arxiv.org/abs/2604.16021v1

Summary:

This work presents LogicLoc, a neurosymbolic agentic framework that integrates LLMs with Datalog for deterministic structural reasoning in codebase analysis. It represents a foundational shift toward verifiable agentic workflows by offloading complex structural traversals to a symbolic engine, drastically reducing token overhead while improving accuracy.

Title: Structured Abductive-Deductive-Inductive Reasoning for LLMs via Algebraic Invariants

Source: http://arxiv.org/abs/2604.15727v1

Summary:

This paper introduces a symbolic reasoning scaffold that operationalizes tripartite inference (abduction, deduction, induction) to enforce logical consistency in LLMs. By using algebraic invariants like the "Weakest Link" bound, it provides a foundational mechanism to prevent error propagation in complex multi-step reasoning chains.

Title: SGA-MCTS: Decoupling Planning from Execution via Training-Free Atomic Experience Retrieval

Source: http://arxiv.org/abs/2604.14712v1

Summary:

This work presents a novel framework that amortizes the high cost of inference-time search by casting LLM planning as non-parametric retrieval of symbolic 'SGA atoms.' By enabling System 2 reasoning depth at System 1 speeds without task-specific fine-tuning, it establishes a new efficiency-reasoning Pareto frontier for agentic planning.

Title: Autogenesis: A Self-Evolving Agent Protocol

Source: http://arxiv.org/abs/2604.15034v1

Summary:

This paper introduces the Autogenesis Protocol (AGP), a foundational framework for self-evolving multi-agent systems that standardizes the lifecycle and evolution of agentic resources. It provides a structured approach to decoupling agent evolution from execution, enabling scalable and auditable improvements in autonomous systems.

Title: AIBuildAI: An AI Agent for Automatically Building AI Models

Source: http://arxiv.org/abs/2604.14455v1

Summary:

This paper introduces a hierarchical multi-agent framework that automates the full lifecycle of AI model development, achieving human-level performance on the MLE-Bench. It represents a foundational shift towards autonomous, self-improving agentic systems capable of complex engineering reasoning.

Title: Bridging MARL to SARL: An Order-Independent Multi-Agent Transformer via Latent Consensus

Source: http://arxiv.org/abs/2604.13472v1

Summary:

The provided text introduces the Consensus Multi-Agent Transformer (CMAT), a novel framework designed to improve cooperative multi-agent reinforcement learning (MARL) by reformulating it as a hierarchical single-agent problem. While traditional models often struggle with action-generation order sensitivity and unstable training, CMAT utilizes a Transformer-based decoder to iteratively generate a latent consensus vector. This shared strategy allows all agents to select their actions simultaneously and independently while remaining highly coordinated. By treating the collective of agents as a unified entity, the system can be optimized using standard Proximal Policy Optimization (PPO). Extensive testing across benchmarks like StarCraft II and Google Research Football demonstrates that this consensus-driven approach consistently outperforms existing centralized and sequential baselines. Ultimately, the research offers a more robust method for reaching optimal joint decisions in complex, multi-agent environments.