Tag
This paper introduces SkillMisevo-Gym and SkillMisevo-Bench to study how self-improving LLM agents can evolve unsafe skills from compromised experience, plus SafeEvolve as a mitigation wrapper. Experiments across 25 agent-method configurations show skill misevolution is widespread and can persist across sessions, though SafeEvolve reduces fresh-session harm significantly.
This paper presents a self-improving 'researcher agent' for Text-to-SPARQL question answering over knowledge graphs, which iteratively refines its own prompts and tools. Evaluated on DBpedia, it achieves 0.22 accuracy and identifies predicate selection as the main bottleneck.
Stanford University has released the full recordings of the graduate course CS329A: Self-Improving AI Agents on YouTube, with 9 lectures covering topics such as test-time compute, verifiers, reinforcement learning, tool feedback, and multi-step reasoning, systematically organizing the direction of agent self-improvement.
The paper 'Memory Reward Inflation in Self-Improving LLM Agents' shows that self-improving agents with frozen weights can still degrade by trusting flawed LLM-generated memory scores, with models endorsing 31–54% of their own wrong answers. This 'Echo Gap' persists across stronger LLMs.
A tweet discusses the missing piece in self-improving agents: a strict review gate where an experienced engineer must approve each skill. It explains how a growing library improves discovery and extraction quality, with humans approving all merges to keep the system trustworthy.
This paper proposes a method for co-evolving evaluation metrics and skills in self-improving LLM agent systems, demonstrating that metrics can be evolved and that a co-evolution approach recovers most of the performance of a ground-truth-driven oracle across code generation, text-to-SQL, and report generation tasks.
Prime Intellect Lab is out of beta, offering a platform to train models with support for various architectures and modalities, enabling self-improving agents.
A curated list of five notable AI articles from the past week, covering self-improving agents, Bun's migration to Rust, vLLM architecture, coding evaluation benchmarks, and agent autonomy levels.
Anthropic and Google released three free workshops on building self-improving agentic systems, covering multi-agent pitfalls, skill gaps, and building autonomous agents.
MetaSkill-Evolve introduces a recursive two-timescale framework for LLM agents to evolve both task skills and the improvement procedure itself, achieving notable accuracy gains on OfficeQA, SealQA, and ALFWorld benchmarks.
Introspection, a new AI startup founded by ex-xAI engineers, introduces 'autoresearch' – a feedback loop system where agents maintain and improve themselves using signals, evals, and human input, moving beyond traditional agent harnesses.
A practical guide explaining three levels of building self-improving AI agents, from manual loops to automated design, with recommended tools and frameworks.
A one-person company runs entirely with 7 AI agents, 10 cron jobs, and no human employees. The agents self-evaluate and improve, operating through Telegram.
This paper presents a novel blueprint for self-improving agents that combines scaffold editing and weight training through a meta-agent and feedback-agent, achieving a 14x speedup on a CUDA kernel for AlphaFold.
This paper introduces the Red Queen Gödel Machine (RQGM), an evolutionary framework for recursive self-improvement under non-stationary utilities, where agents and evaluators co-evolve, improving performance on coding tasks, scientific writing, and Olympiad-level proof grading.
This paper introduces Regimes, an auditable, held-out-gated improvement loop built on the ActiveGraph runtime for self-improving agents. It demonstrates modest improvements on the LongMemEval dataset by autonomously discovering prompt repairs that pass static checks, sandbox execution, and held-out validation.
EEVEE is a novel test-time prompt learning framework for LLM agents that handles heterogeneous data streams through task clustering and co-evolving router-prompt optimization, achieving significant improvements over existing methods across multiple benchmarks.
A prominent AI paper from the week addresses whether self-improving agents are truly discovering new knowledge or merely remixing existing information.
This paper introduces a categorical framework for distinguishing genuine scientific discovery from mere retrieval or search in self-improving AI agents, using category theory to formalize regime transitions. The authors demonstrate the framework with a protein mechanics example where an agent's accuracy drops as it tackles harder problems, but its theory compresses more data, indicating real discovery.
This paper disentangles the roles of evolver and agent in self-improving LLM agents, showing that a small evolver can write sufficiently good updates, while a mid-tier agent benefits most from using them. It recommends using the strongest model as the task executor, not the update writer.