Google's GUARD paper: frozen base model, self-editing agent harness gains +8.89 points
omarsar0 · x · 2026-11-19
Google Cloud researchers published 'GUARD: Stabilizing Recursive Self-Improvement with Regret Guidance', targeting the classic failure modes of self-improving agent harnesses: local overfitting, catastrophic forgetting, and memory pollution.
Core idea
- The base model stays frozen; the agent edits its own prompts, skills, tool rules, and memory.
- Each task is decomposed into pre-execution self-checks; each round grades trajectories against them, using directional regret to isolate what an edit newly broke.
- Mastered rules are pinned while newly broken behaviors are prioritized; candidate edits compete via critic-free group-relative fitness; a Soft-HURT gate accepts an edit only when gains outweigh regressions.
Key results
- 80.80% accuracy on 43 hosted SDK tasks, +8.89 points over the starting harness (p = 0.008).
- Beats Meta-Harness by 6.28 points while using 54.5% fewer tokens per task.
- 90.1% update correctness on procedural memory tasks with zero memory pollution; aggressive rewriters scored similarly but tripled pollution.
- Monotonic improvement across continual task streams with zero drops and zero forgetting.
At test time a single frozen harness runs in one pass, offering a practical grader-free RSI path where weight-space RL is impractical.
More from coding & agent
- Microsoft open-sources openclaw-dev: one-command hosted OpenClaw agent on Azure with sandboxed tool execution — adnan_hashmi · 2026-11-19
- 6-minute Hubble documentary made with Claude and Jimeng CLI workflows — xiaohu · 2026-11-19
- Coding agents gone wild: test-cheating, scope drift, and rogue DB migrations — Common_Dream9420 · 2026-11-19
- Regenerative software: keeping intent as the source of truth so code can always be regenerated — gamagamer · 2026-11-19
- Every on Claude Fable 5.2: Human Thinking at the Start Matters More as AI Builds More — every · 2026-11-19
- Semantic caching LLM answers by embedding similarity is unsafe — 138-pair test shows why — Narrow_Side_7778 · 2026-11-19