Daily updates, analyzing cutting-edge AI research papers
This article provides a comprehensive overview of the "Code as Agent Harness" paradigm, detailing its three‑layer architecture, the roles of code in reasoning, acting, and environment modeling, the mechanisms that enable reliable long‑term execution, and how multi‑agent systems scale the harness through shared code and feedback loops.
SkillOpt reframes LLM agent skills as trainable external state, applying a deep‑learning‑style optimizer to systematically improve skill documents, and demonstrates across six benchmarks, seven models, and three execution modes that this approach yields consistent, large gains and robust transferability.
This free collection gathers 127 top‑conference papers covering 17 large‑model research directions—from perception and decision to safety—providing PDFs, GitHub links, and a web interface to help AI engineers, researchers, and students stay up‑to‑date.
The article introduces GRAM (Generative Recursive Reasoning Models), a parallel multi‑trajectory inference framework that replaces deterministic single‑track recursion with stochastic latent transitions and width scaling, achieving state‑of‑the‑art results on Sudoku‑Extreme, ARC‑AGI, N‑Queens and Graph Coloring benchmarks.
The article summarizes two recent DeepSeek papers—one presenting an L1–L5 taxonomy and four architecture patterns for autonomous research agents, the other proposing a three‑dimensional taxonomy for continual learning, detailing method families, a self‑improvement phase diagram, experimental comparisons, an impossibility theorem, and the production statistics of the Deli AutoResearch framework.
Anthropic’s Claude Opus 4.8 upgrades agentic reliability and honesty, while its new Dynamic Workflows turn hundreds of agents into a hierarchical, parallel, verifiable pipeline that can orchestrate large‑scale code migrations such as React‑to‑Solid.js or a 750k‑line Rust rewrite in days.
The article analyzes Microsoft’s AgenticRAG, a tool‑based RAG framework that lets LLMs control retrieval, showing up to a 5.9× recall improvement over standard methods, reduced need for fine‑tuning, and practical design insights from pre‑production deployment.
The author reviews the 商汤办公小浣熊桌面端 2.0 agent, showing how it moves beyond chat‑only assistants to directly manipulate local files, browsers, and enterprise tools, automating a weekly competitive‑analysis report and embodying the OPC (One Person Capability) concept.
The INTRA paper shows that using a decoder’s cross‑attention as an internal retrieval mechanism eliminates the need for a separate retriever, achieving state‑of‑the‑art multihop QA performance with only 164 K trainable parameters and shared pre‑encoded representations.
The article surveys DeepSeek’s Harness initiative, presenting the Binding‑Constraint Thesis, three‑stage evolution from prompt to harness engineering, the ETCLOVG seven‑layer architecture, and concrete benchmark evidence that harness‑only improvements far outweigh model upgrades, while detailing security, observability, and governance considerations for reliable LLM agents.
