Researchers introduce GUARD to stabilize recursive self-improvement in AI agents
Researchers introduced GUARD, a framework for stabilizing recursive self-improvement in AI agents by using pre-execution self-checks, directional regret and a promotion gate. It aims to let agents update prompts, skills, tool rules and memory without external graders while avoiding regressions and memory bloat. The paper reports gains across memory, skill evolution, continual task and domain reasoning benchmarks, with reduced deployment token usage.
The arXiv paper addresses autonomous AI agents that keep a frozen base model but modify an editable harness of prompts, skills, tool rules and memory. The authors say recursive self-improvement can adapt this harness from experience without weight updates, but self-editing without external graders often backfires. Reflection and scalar scores can mix newly broken behavior with unsolved tasks, leading to local regressions, skill forgetting and memory bloat.
GUARD decomposes each task into pre-execution self-checks and guides harness updates with directional regret. Each round grades trajectories against those checks, pinning mastered rules while prioritizing newly broken behaviors. It selects among parallel candidate edits by comparing them against one another through critic-free group-relative fitness. A Soft-HURT gate promotes an edit only when gains outweigh regressions, scoping specialized fixes to prevent over-triggering and guaranteeing non-regression on its gated buffer.
It addresses a core bottleneck in recursive self-improvement: how to let an agent change its own operating instructions without breaking what already works. Many agent stacks rely on self-reflection, but ungraded edits can optimize noisy signals and degrade stable skills. By separating mastered rules from newly broken behavior and using a gate before promotion, GUARD resembles regression testing for agent behavior. The paper says test time uses one frozen harness in a single pass.
Across benchmarks on memory, skill evolution, continual task streams and domain reasoning, the authors report consistent performance gains without catastrophic forgetting or memory pollution, while substantially cutting deployment token usage. The contribution is a stabilization method for evolving agent harnesses, with implications for teams that maintain AI systems through prompts, skills, tools and memory rather than retraining. The paper does not claim recursive self-improvement is generally safe or solved.
The reported method addresses regressions and memory bloat in AI agents that update prompts, skills, tools and memory over time. It is relevant to teams maintaining such systems because it reports lower deployment token usage and no catastrophic forgetting on tested benchmarks.
Sources (1)
- [1]GUARD: Stabilizing Recursive Self-Improvement with Regret Guidance(https://arxiv.org/abs/2611.23476)