SkillOpt 2.0: A Leaner, Faster Self‑Evolving Agent

Overview of SkillOpt‑Lite

PaperAgent introduces SkillOpt‑Lite, a minimal version of the original SkillOpt pipeline that removes most engineered components. Despite being simpler, the Lite version forms a lighter pipeline, converges more quickly, and reaches a higher performance ceiling.

Benchmark Radar

On a radar chart covering six benchmarks, SkillOpt‑Lite (blue) outperforms both the full SkillOpt (red) and the initial skill baseline (gray). The convergence curves show the blue line pulling ahead within the first 2–3 steps on every task.

What Made the Original SkillOpt Heavy

The first SkillOpt incorporated a full suite of “brakes”: mini‑batch tree merging, learning‑rate decay (L from 4 to 2), a reject‑edit buffer, and epoch‑level slow updates with meta‑skill reflection. The author asks: what is the minimal viable pipeline, and does each component have a theoretical or empirical justification?

Mapping to Zero‑Order Optimization (ZO)

All skill‑optimization methods are unified under a zero‑order optimization framework:

Reflexion – 1‑point gradient estimate

SkillCat – central‑difference estimator

SkillAdapter – coordinate descent

Edit budget – trust‑region analogue

Reject buffer – control‑variates analogue

The author notes that recent methods essentially reinvent numerical optimization using natural‑language interfaces.

Language‑Mediated Program Compilation

Classic ZO relies on blind perturbations, but an agent’s rollout leaves a readable execution trace. Consequently, skill optimization becomes language‑mediated program compilation: the skill is a program, the LLM acts as a compiler, and the trace serves as a debug log.

PAC‑Learning Principles

From PAC learning the author derives two principles:

Cross‑trajectory consensus mining – avoiding over‑fitting to single anomalies, which would inflate the stability coefficient β_exp and cause generalisation collapse.

Independent validation‑set gating – a held‑out validation set can completely remove β_exp, but it must be strictly independent; the paper cites Reflexion, SkillCat, and Trace2Skill as violating this rule.

Pilot Experiment: Copilot 1 Batch vs. SkillOpt 4 Epochs

The author stores raw GPT‑5.4‑nano trajectories as text files and lets GitHub Copilot use a shell tool to parse the logs and directly modify the skill, with no merging, learning‑rate, or validation. Results (relative accuracy, best = 100): LiveMath 50.9 vs SkillOpt 30.3, DocVQA 83.7 vs 81.2 – a single batch outperforms four epochs of the full pipeline. However, on Spreadsheet the score drops to 15.3 (initial skill scores 29.9), demonstrating that closed‑loop validation cannot be omitted.

Insight 4

Providing the agent with a raw shell tool to directly parse logs yields more stable performance than a heavily engineered baseline.

SkillOpt‑Lite Design: Keep Only the Debug Loop

Delete : mini‑batch pooling, learning‑rate schedule, reject buffer, epoch‑slow update.

Retain : a four‑step disk‑based debug loop:

Dump each trajectory to a separate file.

Explore trajectories – the optimizer lists directories, clusters failures, and reads selected files.

Consensus mining + minimal diff patch generation.

Independent validation‑set gating – overwrite best_skill.md if the new skill exceeds historical best; rejected versions are archived.

/skillopt-loop rounds=10 batchsize=40 target=gpt5.4-nano

Main Experiments: Inference Gains, Semantic Gains Flat

Across six benchmarks and five model scales, the most striking numbers are:

LiveMath: GPT‑5.5 improves from 36.6 to 73.6 (+37.0), surpassing full SkillOpt by +8.8.

nano: improves from 29.9 to 55.7 (+25.4), overtaking the standard GPT‑5.4 optimized by SkillOpt (54.0).

Spreadsheet: GPT‑5.4 scores 79.4 vs. SkillOpt 61.5; GPT‑4o LiveMath jumps from 31.2 to 58.8.

Mechanism explanation: mini‑batch pooling averages multiple text updates, diluting gradient signal; Lite’s local file editing removes logical deadlocks. Semantic tasks (SearchQA, OfficeQA) see only modest gains (+0.1 ~ +1.5) because capacity is already saturated, but Lite achieves a markedly lower compute cost.

Convergence Speed

On the best‑validation curves for the first ten batches, Lite pulls ahead within the first 2–3 steps on complex tasks such as LiveMath and Spreadsheet. SkillOpt lags early due to splitting and damping overhead, and its ceiling is lower. Removing the buffer and pooling does not cause premature convergence or oscillation; the final ceiling is higher.

HarnessOpt Easter Egg: Nano Beats GPT‑5.5

Extending the “everything is a file” philosophy, HarnessOpt treats the harness as a code file, relaxes path restrictions, and optimises the execution scaffold within the same loop. Engineering adds round‑0 human approval and three loop invariants (whitelist, sandbox smoke test, git rollback). Results show nano (0.7758) surpasses the standard harness + full SkillOpt GPT‑5.5 (0.7620); combined optimisation reaches 0.8577. Failure modes differ by model: mini/5.4 suffers data blindness, 5.5/nano suffers tool failure leading to reasoning loops; HarnessOpt expands preview range and intercepts repeated states to trigger fallback.

Conclusion

When the base model is sufficiently strong, the optimal optimiser is no longer a sophisticated algorithm but a coding agent capable of reading logs and editing files.

SkillOpt-Lite: Better and Faster Agent Self‑evolution via One Line of Vibe
https://arxiv.org/abs/2607.03451
https://github.com/EvolvingLMMs-Lab/SkillOpt-Lite