Tag
Space-X is a self-evolving spatial-reasoning framework where an Execution Agent builds evidence chains with role-specific memory and a Control Agent diagnoses breaks to jointly revise tools, skills, and decision rules, yielding large accuracy gains across benchmarks with four base models.
An empirical arXiv study examines /retro, a skill that lets coding agents review completed sessions and propose harness changes, finding that while many proposals are generated, only some reduce repeated work on later tasks — showing harness self-evolution needs mechanisms to verify and revise changes.
FilingLoop introduces an interactive benchmark and self-evolving LLM agent framework for corporate disclosure review and revision, grounded in 380 SEC IPO review tasks with requirement-level evaluation that improves fulfillment from 24.28% to 27.96% via procedural learning without weight updates.
PhysEvolve 提出一个验证器驱动的自演化框架,利用物理引擎验证模型自生成的可执行世界与反事实仿真,以 GRPO 训练视觉语言模型的因果物理推理能力,在 PhysBench、CausalPhys 和 ContPhy 上取得显著提升且无需人类干预数据。
Google Cloud AI Research, in collaboration with UIUC, proposes OO-Harness, which introduces object-oriented design (inheritance, encapsulation, polymorphism) into the automatic evolution of agent harnesses. Through a class hierarchy, it enables code reuse and safe fallback for specialized scaffolds. On three long-horizon benchmarks, the main metric improves from 61.62% to 77.41%, and it significantly outperforms traditional self-evolution methods on out-of-distribution tasks.
VibeGame is a framework that builds complete 2D games from natural-language prompts using an AI-native game engine with frame-synchronous runtime control and an eight-agent adversarial team that separates implementation from review. It introduces a training-free self-evolution mechanism distilling accepted projects into reusable components, outperforming other agent harnesses on the GameCraft-Bench subset.
The paper presents RiseHarness, a framework from Tencent and multiple universities that iteratively improves agent harnesses for long-horizon terminal tasks while keeping model weights fixed, analyzing successful and failed runs to progressively integrate better execution mechanisms.
论文提出 DeepEvolve,一个基于未来事件预测的可持续进化智能体系统,通过在动态开放环境中与现实世界结果对齐实现递归自我改进(RSI),其自我改进的预测器在 FutureX 排行榜连续三周排名第一,评级提升最高达 195%。
论文提出 SERA(Self-Evolving Robotic Agent),通过可编辑的 harness 维护任务上下文并利用执行反馈自我进化规划,配合固定的 VLA 执行器来改进长时程机器人操作任务,在 RoboMME 上取得 51.38% 的成功率。
华为openJiuwen团队提出Reward Engine,通过构建由测试用例和评分细则组成的冻结RewardPack,为开放式工具智能体提供证据驱动的奖励信号,用于自修订与强化学习,在WorkBuddy Bench Office和Claw-Eval上显著提升通过一致性。
提出了EvolvingBrain,一个用于开放世界多模态智能体自我进化的自适应技能系统,通过分层技能自进化和状态感知自适应执行,在VAB-Minecraft和DiscoveryWorld上提升了任务成功率与泛化能力。
华为开源了多智能体协作框架「九问Swarm」(JiuwenSwarm),采用无内嵌模型的纯Harness设计,支持E2A统一协议、Leader/Teammate分布式执行、确定性工作流SwarmFlow,以及带人工审核的技能自进化与AutoHarness机制。
该论文提出一种将LLM代理的自进化限制在运行时harness(指令文本、工具调用逻辑与原语组合)的框架,模型权重保持固定,使每次适应都成为可审查的diff,并通过哈希链记录的准入门进行管控。模拟实验显示受控双循环仅放行144/7449个候选变更且未恶化误差,而无约束系统放行了309个有害变更。
本文提出 SAGA 框架,通过将交互轨迹逐步抽象为可复用的原则与程序,并在执行中实例化和校验这些原则,形成执行-抽象反馈循环,使 LLM 智能体能在不微调参数的情况下持续从经验中进化,在 ScienceWorld 和 ALFWorld 上取得性能提升。
MIT 团队开源了 JAZ 框架,用纯提示和极简的 invoke 循环替代复杂记忆附加系统,在长程召回基准上以更低成本超过 Letta (MemGPT) 和 ACE。
The paper introduces VERSE, a verified self-evolving optimizer that improves both the executor agent's harness and its own prompts, tools, and procedures using execution-based verification, achieving 42.3% and 37.7% accuracy on held-out SWE-rebench tasks versus 39.2% and 29.3% for the strongest baselines.
The paper identifies why self-evolving reasoning models suffer performance collapse — invalid questions and mathematically-equivalent question repetition — and proposes R-Quest, which uses validity and novelty feedback to sustain self-evolution, outperforming R-Zero by 17.32 points across 12 benchmarks over ten rounds.
FloWright proposes a hierarchical, structure-aware reward paradigm enabling LLM agents in multi-agent workflows to self-evolve and co-evolve without extra models or labels, while DataWright adaptively hardens datasets into workflow-level tasks. Small open models trained with FloWright gain up to +7.41% performance, with co-evolving multiple roles (+5.03%) outperforming optimizing a single role (+2.83%).
Professor Li Hongyi from National Taiwan University has updated the 2026 spring machine learning course, focusing on AI Agents and model self-evolution in the large model era. This marks a complete transition of the course from classical machine learning to cutting-edge large model technologies.
This paper introduces VQS, a method for self-evolving vision-language models that uses program verification to generate and label questions from images, significantly improving accuracy and model performance over existing baselines.