PACE: Two-Timescale Self-Evolution for Small Language Model Agents

Self-evolving agents improve by repeatedly proposing changes to their own prompts, skills, or workflows and keeping those that score higher on a small held-out set. Almost all effort has gone into the proposer that generates candidates; we argue the weak point is the acceptor, the rule that decides whether to commit a change. Applied hundreds of times against the same noisy dev estimate, the ubiquitous "keep it if the score went up" rule is uncontrolled adaptive multiple testing: the agent effectively p-hacks itself, accumulating false commits that make it churn and drift rather than improve. We recast committing as a sequential hypothesis test and propose PACE (Paired Anytime-valid Commit Evaluation), a training-free, anytime-valid commit gate. Each candidate is compared to the incumbent on identical instances and committed only when a testing-by-betting e-process accumulates decisive evidence, stopping early to save evaluations and controlling each candidate's false-commit probability at a user-set level even under optional stopping (a per-decision guarantee). On Qwen2.5 agents (0.5B-3B) self-evolving at the prompt level on GSM8K, SVAMP, and ARC-Challenge, greedy acceptance commits 30-42% false and 10-33% harmful edits when a genuine improvement is hidden among noisy proposals, while PACE commits the real one and essentially nothing else, matching greedy's held-out accuracy at sharply lower variance and about 18% lower evaluation cost. With no real gain available, greedy commits 13-21 spurious self-modifications per run (72-100% false) and degrades the most fragile agent by 4.9 points, while PACE holds at baseline. Reliability of self-evolution depends on the acceptor, not only on the proposer.

Modern LLM agents are often improved by modifying prompts, tools, or workflows manually, while the executable scaffold surrounding the model---the \emph{harness}---is typically treated as a fixed artifact after deployment. This work studies an alternative where the harness is \emph{task-specific and continuously evolvable}: each task family maintains its own harness, which is hot-swapped across iterations through a fixed task-injection seam and rewritten using environment feedback. We introduce \textbf{Hierarchical Self-Improvement (HSI)}, a framework in which a single frozen LLM $M$ operates across three hierarchical scopes: a task harness $H$ that executes tasks, an evolver that rewrites $H$, and a meta-evolver that rewrites the evolver's strategy code under a frozen outer anchor. A thinking-on/off design isolates the contribution of harness evolution by disabling reasoning during task execution while enabling it during self-modification. HSI is bounded by two factors: a \emph{feedback-fidelity bound}, since evolution requires informative reward signals to guide selection, and a \emph{backbone capability bound}, since harness redesign cannot overcome limitations of the frozen model. On BALROG with DeepSeek-V4-Flash-Preview as the frozen backbone, HSI achieves consistent gains over the initial harness on moderate-difficulty tasks ($+39.3$ on BabyAI, $+33.0$ on Crafter, $+25.0$ on TextWorld, and $+15.0$ on MiniHack, all in raw % Progress), while obtaining strong held-out generalization on BabaIsAI sub-suites ($0.98$ best-test on BreakStop and $1.00$ on GoTo from a $20\%$ unseen split). On tasks beyond the backbone's capability (NLE), harness evolution provides no improvement. These results demonstrate task-specific harness evolution as a viable axis for improving frozen LLM agents under clear empirical limits. Code is available at https://github.com/TailinZhou/hsi.

An LLM agent's real-task performance is shaped as much by the harness around its model as by the frozen model itself: its prompts, injected knowledge, runtime control, and configuration. In deployment the harness is often the only lever available, so improving it automatically is the natural way to raise performance without touching the weights. The hard part is not generating changes but knowing which one truly helped. Self-generated feedback is noisy, and an apparent gain can be a measurement artifact or an edit that merely overfits the tasks it was tuned on. We present a self-evolving agent-harness framework that separates proposing changes from crediting them: a language model diagnoses failures and proposes patches, while all sampling, measurement, and significance testing are owned by deterministic code, so every credited improvement is trustworthy by construction. Patches populate a gated, categorical quality-diversity archive (GSME) keyed on the (WHERE x WHY) pathology an edit addresses rather than the tasks it fixes, an anti-overfitting inductive bias; generalization is measured on a sealed test scored only after evolution. Across seven domains with a frozen open-weight model, the harness is train-selected and scored once on a sealed test; its credited gains there are +9 to +15.5pp and retain 86-147% of the training gain, evidence they generalize rather than overfit. The winning patch tracks the model's dominant pathology, not its size or family: changing the model can change the pathology and the patch, while the same pathology-to-patch match recurs across two model families. What transfers is the diagnose-and-credit loop, not any specific harness.