From Procedural Skills to Strategy Genes: Towards Experience-Driven Test-Time Evolution
Junjie Wang, Yiming Ren, Haoyang Zhang
cs.SE, cs.CL
2026-04-16
In 4,590 trials on 45 scientific coding tasks, a 230-token strategy gene lifts mean pass rate from 51.0% to 54.0%; a 2,500-token Skill pack falls to 49.9%.
A lot of agent work now treats reusable experience as a content object: verbal reflection, skill libraries, procedural memory files. The usual bet is that a more complete, better documented package will steer the next run more reliably.
This beta technical report from Tsinghua and EvoMap tests that bet. The question is not whether to store experience, but what form it should take when it re-enters a frozen model. They treat reusable experience as a test-time control signal: parameters stay fixed, only an external object is allowed to shift generation. Two packagings of the same per-scenario source are compared. Skill is a human-facing document bundle, about 2,500 tokens, with overview, workflow, pitfalls, errorhandling plus apinotes, examples, and scripts. A strategy gene is a model-facing control object, about 230 tokens, with keywords, summary, strategy, and AVOID.
The testbed is Gene-Bench: 45 scientific coding scenarios and 4,590 retained trials. The model writes Python, a sandbox runs it for 120 seconds, and scoring is checkpoint pass rate rather than binary success. Models are locked to Gemini 3.1 Pro Preview and Gemini 3.1 Flash Lite Preview at temperature 0.05 with a 16,384-token output cap. Control goes into systemInstruction; the task stays in contents.
A strategy gene is not a truncated Skill. Distillation keeps matching cues, a one-line intent, a short strategy list, and failure-aware AVOID items. Gene Evolution Protocol then canonicalizes the gene so it can be matched, replaced, and revised as an object. Two higher layers sit around it: a capsule records a validated execution path, an event is an immutable evolution log. The loop is SCAN, SIGNAL, INTENT, MUTATE, VALIDATE, SOLIDIFY.
The UV-Vis peak-detection scenario shows the packaging gap. Models often treat mindistance as a wavelength instead of a sample index, or report peakwidths output as FWHM without converting back. Skill dumps overview, workflow, and reference material into context. The gene only flags the high-risk decisions: detect peaks by prominence, convert distance to index units first, convert widths back to wavelength before reporting FWHM.
Three probes split the 4,590 trials. Skill Probe (1,440) asks whether the document package works as control. Gene Probe (1,890) asks whether the gene is merely a shorter prompt. Evolution Probe (1,260) asks whether the gene can carry accumulated experience. The experience source is held fixed per scenario; only the wrapper changes.
Against a no-guidance mean of 51.0%:
| Condition | Pro | Flash | Avg. | Δ |
| Full Skill | 50.7% | 49.0% | 49.9% | -1.1 |
| No guidance | 60.1% | 41.8% | 51.0% | 0.0 |
| Strategy gene | 59.9% | 48.2% | 54.0% | +3.0 |
Skill lifts Flash from 41.8% to 49.0% and drops Pro from 60.1% to 50.7%. The gene barely moves Pro (59.9% vs 60.1%) and raises Flash to 48.2%. The +3.0 average is almost entirely a weak-model gain.
Inside Skill, usable signal is sparse. Overview costs 4.7 points. Workflow is the only clear positive at +1.5. ErrorHandling is +0.7, QuickRef +0.5, Pitfalls about flat (+0.1). Truncating Workflow and Pitfalls to 230 tokens leaves +0.5 and +1.0, still below the gene at +3.0. Brevity explains part of the gap, not all of it.
Shorter is not automatically better. Keywords only reach 53.5% (+2.5). Keywords plus summary fall back to 51.0%. The three-field gene hits 54.0%. Reattaching API notes or examples drops the mean to 51.5% and 52.0%. Two complementary genes crash to 44.9% (-6.1); two conflicting genes still hold 53.2%. On precise scientific code, a blurred control focus is worse than a contradiction.
Wrong content hurts: wrong algorithm 48.8%, wrong domain 49.4%. Structural distortion is cheaper: inverted priority 52.8%, overconstrained 55.9%, a stale paradigm even 56.6%. Flattening the same gene into prose lands at 50.5%. The schema itself is worth about 3.5 points.
Failure history is not carrier-neutral. Skill plus failure is 47.8%, freeform plus failure 49.6%, gene plus failure 52.0%. Naive append still sits 2.0 points below a clean gene. Distilling failures into standalone warnings is the best of this set at 54.4% (+4.6; this subset's no-guidance baseline is 49.8%). Mixing strategy and failure weakens both.
On CritPt, two gene-evolved systems (OpenClaw host, Evolver engine, about two days each) beat their paired bases: Gemini 3 Pro Preview 9.1% to 18.57%, Gemini 3.1 Pro Preview 17.7% to 27.14%. The March run covers 70 tasks, 210 gene slots, 36 unique IDs, with 148 arxiv-derived selections. Observable input/output cost for the February run is about $0.81, reported as 3.1% of a $26 gemini-3-pro-preview baseline; reasoning tokens are missing from that figure.
For anyone stuffing SKILL.md into an agent, the hard result is simple. A human-readable document in the system prompt is often a liability. Pro losing about 9.4 points under full Skill is more important than the +3.0 average.
The practical recipe is also specific. Keep experience short, fielded, and failure-aware. Compress updates into AVOID items instead of pasting logs. Inject one task-matched gene, not a bag of related ones.
The CritPt jump of about 9 points should not be read as a prompt-format win. That stack includes retrieval, validation, and solidification. The +3.0 on Gene-Bench is the evidence that belongs to packaging.
This is a beta technical report with no standalone limitations section. Only two Gemini Preview models are tested, at temperature 0.05. There is no Claude, GPT, or open-weight number.
All Gene-Bench scores are checkpoint pass rates on scientific Python tasks. Workflow helps Flash and hurts Pro, so control quality is model-dependent and the average Δ hides that split.
The CritPt comparison bundles an evolution engine, a host runtime, and the gene format. How much of 9.1% to 18.57% is representation versus multi-step tool use is not separated. The authors are at EvoMap, evaluating their own protocol and engine. Independent replication is not in the paper.
Overconstrained and stale-paradigm genes beat the clean template, so the "right" schema is not stable. Two complementary genes collapsing also means the claimed composability is not shown in this setup.