Title: End-to-end autonomous scientific discovery on a real optical platform

Source: http://arxiv.org/abs/2604.27092v1

Summary:

This work marks a milestone by demonstrating the first AI agentic system to autonomously identify and experimentally validate a previously unreported physical mechanism. It introduces a novel dual-layer architecture and Meta-Trace memory to maintain stable research trajectories over long-horizon, complex investigations.

Title: Parallel Prefix Verification for Speculative Generation

Source: http://arxiv.org/abs/2605.04263v1

Summary:

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: AGEL-Comp: A Neuro-Symbolic Framework for Compositional Generalization in Interactive Agents

Source: http://arxiv.org/abs/2604.26522v1

Summary:

This framework introduces a principled neuro-symbolic architecture that addresses systemic failures in compositional generalization within LLM-based agents. It integrates dynamic causal program graphs, inductive logic programming, and neural theorem proving to enable agents to build explicit and interpretable world models.

Title: Operating-Layer Controls for Onchain Language-Model Agents Under Real Capital

Source: http://arxiv.org/abs/2604.26091v1

Summary:

This paper identifies the "operating layer"—comprising prompt compilation, typed controls, and execution guards—as the critical architectural framework for ensuring agentic reliability in high-stakes environments. It provides a foundational blueprint for deploying agents that manage real capital through structured validation loops rather than relying solely on base model capabilities.

Title: ADEMA: A Knowledge-State Orchestration Architecture for Long-Horizon Knowledge Synthesis with LLMAgents

Source: http://arxiv.org/abs/2604.25849v1

Summary:

ADEMA establishes a foundational architecture for long-horizon agentic reasoning by introducing explicit epistemic bookkeeping and knowledge-state orchestration to prevent information drift. This framework provides a new primitive for complex knowledge synthesis that treats evidence-bearing artifacts and state transitions as primary, recoverable design commitments.

Title: Nautile-370M: Spectral Memory Meets Attention in a Small Reasoning Model

Source: http://arxiv.org/abs/2604.24809v1

Summary:

This paper introduces a hybrid architecture that alternates linear-time spectral operators with transformer layers, providing a formal proof that such 'spectral memory' can match the expressiveness of full self-attention. It demonstrates a significant breakthrough in creating small, efficient reasoning models that maintain long-context performance under strict parameter budgets.

Title: Kwai Summary Attention Technical Report

Source: http://arxiv.org/abs/2604.24432v1

Summary:

Kwai Summary Attention (KSA) introduces a novel architectural primitive that compresses historical context into learnable summary tokens, enabling a O(n/k) complexity for long-context sequence modeling. This approach provides a foundational new path for scaling next-generation LLMs by trading minimal memory for interpretable, semantic-level retention of long-range dependencies.

Title: Structural Enforcement of Goal Integrity in AI Agents via Separation-of-Powers Architecture

Source: http://arxiv.org/abs/2604.23646v1

Summary:

This research introduces the Policy-Execution-Authorization (PEA) architecture, a foundational system-level primitive that decouples agent intent from execution to structurally prevent agentic misalignment. By moving beyond probabilistic model-level safety to cryptographically constrained system enforcement, it provides a novel framework for autonomous agent governance.

Title: DLM: Unified Decision Language Models for Offline Multi-Agent Sequential Decision Making

Source: http://arxiv.org/abs/2604.23557v1

Summary:

This paper proposes a unified framework that treats multi-agent decision-making as a dialogue-style sequence prediction problem, enabling robust zero-shot generalization across heterogeneous environments. It establishes a foundational approach for scaling multi-agent reinforcement learning by leveraging the flexible architectural interface of large language models.

Title: From Coarse to Fine: Self-Adaptive Hierarchical Planning for LLM Agents

Source: http://arxiv.org/abs/2604.23194v1

Summary:

This paper introduces AdaPlan-H, a novel agentic reasoning framework that enables LLM agents to dynamically adjust planning granularity based on task complexity. It provides a foundational primitive for long-horizon task execution by mimicking human progressive refinement to optimize the balance between planning detail and execution efficiency.