Tag
HarnessRL 提出一个与策略共同进化的训练框架,通过自适应任务采样、rollout 引导和 rubric 标准更新来解决开放式强化学习中的奖励失效问题,并发布了五领域评测集 Rubric-Mix-10k,在 4B 和 9B 模型规模上均优于最强基线。
本文揭示了自我改进智能体中技能-策略共进化的脆弱性陷阱,提出部署导向的评估协议CODA来审计技能漂移的影响,发现最差情况下共进化增益会被削弱52-93%。
AgentSafeEvo 提出一种基于攻防博弈的对抗协同进化强化学习框架,联合训练攻击者与防御者 LLM 智能体,以提升多轮工具调用场景下的安全性,同时保持通用能力不下降。作者在七个安全与能力基准上进行了评测,并开源了代码、模型和数据集。
EAU 是一个元智能体框架,在训练多轮 LLM 智能体时协同演化用户模拟器、环境与强化学习配置,使训练系统随策略一同进化。在三个交互式智能体基准上,EAU 比最强基线整体提升 14.7–15.0%,在 ColBench 上任务完成率最高提升 36.7%。
HERA 提出了一种通过共同演化 agent harness 与任务环境来提升 LLM 智能体在不可行任务上弃权能力的框架,将弃权准确率从 61.7% 提升至 83.3%,且可跨 19 个模型迁移,并发布 HERA-Bench 评测基准。
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%).
本文评估了将协同演化的通信协议从二维仿真直接迁移到三维物理机器人环境(无重训练)的效果,发现迁移仅部分且不对称,表明涌现通信的成功迁移不仅依赖信号过程本身,还依赖其演化时的生态与导航条件。
KernelZero introduces a co-evolution framework with Proposer and Coder models to enhance GPU kernel generation, achieving superior performance on CUDA and Triton benchmarks.
MACE is a memory-agent co-evolution framework for LLM-based multi-agent systems that uses adaptive memory graphs to organize functional memory units and adapt through execution feedback, improving task performance over baselines.
This research explores combining harness evolution with model adaptation for AI agents, discovering that direct imitation from experts harms weaker models' performance and proposing an on-policy correction method to improve performance without breaking harness fit for enterprise tasks.
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.
EnvHarness introduces a programmable layer to dynamically reshape static environments for reinforcement learning, improving agent performance through automated targeting of weaknesses with EnvRigger.
This paper proposes model-harness co-evolution as a fundamental principle for recursive self-improvement in AI agents, introducing HELIX, a source-traceable system that improves both execution and learning by generating structured training signals from verified trajectories.
Researchers propose the Red Queen Gödel Machine, a framework for self-improving AI where agents and evaluators evolve together to overcome evaluation ceilings, showing enhanced performance in tasks like scientific paper writing and grading.
This paper introduces PIAC, a framework that improves LLM-based automatic construction of parallel algorithm portfolios by using a potential-gain metric that eliminates the need for reference solutions and by leveraging LLMs to generate diverse instance mutators. It consistently outperforms existing LLM-ACP baselines on TSP and CVRP, achieving up to 19.76% relative improvement.
A survey paper proposing a three-stage taxonomy of co-evolution in agentic systems, covering agent-agent, agent-environment, and meta co-evolution to enable open-ended improvement beyond fixed human-designed paths.
This paper proposes scaffold-mediated post-training, a paradigm where procedural scaffolds co-evolve with LLM parameters through discovery, distillation, and dynamic recompilation. On FeatureBench, automatically discovered skills improve pass rate by 8.1pp, with a 27.7% pass rate after distillation.
This paper introduces CoCoEvolve, a self-supervised method that improves cross-representation understanding across charts, tables, and code by enforcing one-to-one consistency between representations, with training-time and test-time co-evolution objectives.
Echoverse presents a method for generating deep, evolving synthetic environments to train computer-use agents, demonstrating substantial accuracy gains and releasing a benchmark with grounded graders.
Introduces DecoEvo, a score-decoupled co-evolution method for LLM optimization in text space that jointly improves solver and rubric-generator skills without gold rubrics, achieving 2.8–5.0% relative gains over baselines across five benchmarks.