self-evolving

Presents NeSy-Spatial, a neuro-symbolic framework that self-evolves spatial reasoning skills by composing tool-use and geometry skills, improving accuracy on spatial reasoning benchmarks.

BigBang-v1 is a self-evolving 36B LLM from Endless Frontier Lab in Shanghai, trained with AI-generated frontier tasks and achieving strong performance with only 10K high-quality examples across science, coding, tool use, and long context.

Introduces EvoHarness-RL, a framework that learns runtime harness policies for long-horizon LLM agents, enabling them to construct and update external state (belief, progress, experience) during task execution. Using Qwen3-8B on ALFWorld, it achieves 96.9% success and reveals harness annealing and evolution dynamics.

A tweet expressing amazement at the concept of dynamic agent orgs—self-evolving multi-agent systems where the graph structure rewrites itself during execution.

This paper surveys self-evolving coding agents, which improve their future behavior by updating frameworks, memory, skills, tools, or models from prior coding interactions, and presents a taxonomy of what evolves, when, and what software-specific evidence drives it.

Presents SciToolAgent-Evo, an ontology-aware self-evolving LLM agent for open-world scientific tool acquisition, along with the OpenSciToolBench benchmark of 900 realistic tasks. The agent uses an evolving memory and LinUCB-based bandit gate to dynamically explore and acquire novel tools.

AgentStream introduces a unified framework to evaluate self-evolving LLM agents under streaming task scenarios, showing that self-evolution reliability varies across scenarios and is gated by model capability.

SERPO introduces a self-evolving rubric policy optimization framework for test-time reinforcement learning in open-ended generation, replacing answer voting with a closed loop that co-evolves response evidence, query-specific rubrics, and policy parameters, achieving significant improvements on health and research benchmarks.

This paper introduces a Self-Evolving Just-In-Time Memory framework for proactive embodied safety, combining a Risk-Sufficient Topological Belief Graph, Agency-Grounded Factual Memory, and Experience Memory to mitigate hazards without stalling task progress. Experiments on IS-Bench show significant Safe-Success rate improvements (e.g., +30.3% on Qwen3-VL-8B).

Microsoft's strategy for self-evolving agent skills, training them like neural networks with epochs, batch size, learning rates, and validation gates, fully open-source.

Microsoft open-sourced SkillOpt, a tool that enables AI agent skills to self-improve by automatically evaluating and rewriting their instructions, outperforming hand-crafted prompts and other optimizers with no model lock-in.

This paper proposes SEED, a method that internalizes post-hoc skills from trajectories into model parameters through self-evolving online distillation, solving the reward sparsity problem in long-horizon RL training, achieving significant improvements on benchmarks such as ALFWorld.

HealthClaw is an open-source agent architecture for longitudinal personal health management that uses self-evolving memory to improve support over repeated encounters, achieving higher accuracy and privacy compared to baselines across biomedical tasks.

Proposes SEED, a self-evolving on-policy distillation framework that converts completed trajectories into hindsight skills to improve reinforcement learning for interactive agent tasks, achieving consistent performance gains and sample efficiency.

Proposes EVOQUANT, a self-evolving framework that uses LLMs and a verifier pipeline to automate quantitative trading strategy optimization, achieving significant Sharpe ratio improvements across A-share and crypto markets.

Introduces Critic Experience Bank (CEB), a self-evolving critic framework for step-level confidence estimation in LLM agents that leverages a memory bank of past judgments and consequences to improve calibration without training.

This paper presents a method for compiling repeated standard operating procedure steps into validated, versioned tools before deployment, replacing inference-time code generation. In a fulfillment center alarm-triage system, this approach reduces p50 latency by 42% and end-to-end error rate by up to 53%.

This paper introduces EvoSOP, a framework that enables LLM agents to synthesize atomic actions into reusable Standard Operating Procedures (SOPs) and iteratively optimize their toolset, significantly improving task success rates and reducing interaction rounds.

HASE is a reinforcement-learning framework that co-evolves model weights, task solutions, and harness components (guidance and evaluation) in a unified agentic process, enabling a single 8B-parameter model to match the performance of much larger systems on text classification and alpha factor mining tasks.

ComfyClaw is an agentic skill evolution framework for ComfyUI image generation workflows, using typed graph editing and region-level VLM verifiers to translate visual failures into repair suggestions, outperforming baselines across multiple configurations.