A research paper introduces PoS, an inference-time framework for constructing and maintaining explicit belief states in large language model (LLM) agents, enabling them to undertake complex tasks with a coherent understanding of the current world.

ActiveSaddler, a new automated curriculum learning method, optimizes LLM agent harnesses by dynamically updating prompts, tool interfaces, and control logic from execution feedback, improving test Pass@1 by 4.4-7.5 percentage points.

AutoSynthData is a tool for generating training data for enterprise AI agents, allowing for more efficient and effective agent development. This can improve the reliability and governance of AI agents.

Researchers propose a co-evolving framework for AI agents to learn from failures by generating hard negatives. This framework improves average task reward by 5.7% across three domains. The code will be made publicly available.

Researchers investigate whether large language models (LLMs) can automate software optimization, finding that LLM-based agents can achieve substantial performance improvements in scientific software, potentially making manual optimization obsolete for well-scoped, verifiable problems.

The authors propose VACE, a method for co-evolving agent models and their harnesses through alternating reinforcement learning and trajectory-driven refinement. This approach is shown to improve performance on OfficeQA and AutomationBench benchmarks, outperforming weight-only reinforcement learning and ungated alternation.

This paper proposes PrimeSeeker, a capability-oriented framework for constructing web-grounded anchor structures to improve deep search agents. It achieves strong performance on five deep-search benchmarks and reduces retrieval redundancy.

Researchers propose using large language models (LLMs) to replace the conventional optimizing and lowering pipeline in compiler backends, demonstrating an LLM agent can translate Triton kernels directly into PTX with improved performance (0.83x-3.34x) compared to autotuned Triton.

This paper presents an architectural mediation approach using the Model Context Protocol (MCP) to enable controlled interaction between large language model (LLM) agents and data space services, allowing for interoperable and standards-aligned integration of AI agents into data space ecosystems.

Cloudflare releases EmDash 1.0, a stable and secure CMS built on Astro, with a decentralized plugin registry and support for MCP, CLI, and API. The CMS is free and open-source, and has been tested with real-world deployments and security measures.