Tag
提出CoEvolve框架，通过反事实轨迹对比实现多智能体系统中记忆与协作架构的对齐协同进化，在ALFWorld和WebShop基准上取得最佳综合性能。
MemScope 是一个由中国科学院提出的框架，通过工具使用的 Patch Agent 检查原始执行轨迹、运行时内存状态和现有架构代码，实现对智能体记忆架构的证据驱动、局部化原位进化，在 GAIA、xBench 和 ALFWorld 上均取得显著提升。
The paper introduces EvoSim, a self-evolving multi-agent AI scientist that autonomously constructs and refines physics-based models using experimental discrepancies, evaluated on lithium-metal battery modeling where it surpasses human-expert accuracy.
EvoSteer proposes an online self-evolving graph orchestration paradigm for LLM-based multi-agent systems, using a reference-anchored flow-matching loss (AnchorTB) and validated skill admission to repair failing steps during execution, outperforming baselines across twelve datasets spanning QA, math, code, and decision-making.
FailBank, a four-stage self-evolving framework, converts runtime safety-shield feedback from a CBF-based teacher into persistent policy improvements via guarded LoRA updates, improving task success by up to 25.4 percentage points over runtime shielding on the VLA-Arena benchmark.
The paper proposes Physical Coding, representing robot tasks as executable code, and introduces HexaAnything, a self-evolving coding agent that enhances performance on benchmarks like RoboCasa365 and PhyBench, with implications for real-world intelligence.
A Google paper introduces Procedural Graphs, an editable workflow structure for LLM agents that improves performance on long tasks by evolving from execution feedback, outperforming baselines in most benchmarks.
This position paper proposes a Foundation Model Operating System (FMOS) to virtualize foundation model interactions, providing applications with dedicated, trustworthy instances and enabling self-evolving capabilities through orchestration and policy enforcement.
EvoOntology introduces a self-evolving ontology layer for data agents, encapsulated as an MCP server, to bridge the agent-data gap and improve performance on heterogeneous data tasks as shown in benchmarks.
This paper introduces Procedural Graph, a framework that organizes LLM agent actions into structured triplets for improved long-horizon tool use, with self-evolving topology to enhance performance.
Enhanced the TUI of ouroboros, a CLI tool for self-evolving auto-research systems, and deleted 2.79 million lines of vestigial code to improve usability.
CHIME is a credit-aware hierarchical memory framework that separates planning and execution memory banks to improve long-horizon agentic planning by accurately attributing task outcomes and outperforming baselines.
J-Zero is a unified framework for co-evolving Challenger, Solver, and Judge models from zero data, enabling self-improvement in language models across both verifiable and unverifiable domains with performance surpassing baselines.
JIT-Agent is a trainable model that synthesizes adaptive agent harnesses for off-the-shelf LLMs, improving performance across diverse models and tasks.
This paper surveys agent memory in AI agents, addressing the problem of context explosion and proposing a framework for self-evolving agents through various memory types and management strategies.
This article comments on the low barrier to entry for autonomous-evolving code agent tools, drawing an analogy to stand-up comedy performances, pointing out that in the current code agent field, everyone can claim to be skilled at complex tasks.
Introduces ERSkill, a retrieval-centric framework for self-evolving, skill-guided adaptive memory access in LLM agents. It co-evolves retrieval skills and a routing policy, substantially outperforming strong baselines across agent memory benchmarks.
The article introduces MindMemOS, a portable and self-evolving memory operating layer for AI agents that uses a unified entity-property-time structure, with algorithms for memory refinement and skill evolution. It achieves notable accuracy on LOCOMO and PersonaMem benchmarks and improves SpreadsheetBench performance by 9.2 percentage points.
The author comments on Deepseek Harness's plugin-based philosophy, arguing that it turns everything (GUI/TUI, etc.) into plugins and opposes hardcoding workflows. They compare it with Pi Agent and Prime-Agent, propose the idea of building a self-evolving Agent in a REPL, and have already started implementing it with Codex.
MEGA is a self-evolving infrastructure for coding-agent optimization that distills reusable wisdom from sessions, composes it via a typed Wisdom Graph, and uses operational evidence to continuously improve both the agents and the knowledge guiding their optimization.