Prompt Evolution vs Playbook Accumulation
Choose between evolving whole candidate prompts (GEPA-style) and accumulating itemized insights by delta updates (ACE-style) based on how many granular, long-lived rules the task needs to retain.
Both families are context adaptation: they improve behaviour by editing inputs, not weights. They differ in what is optimized and how it is updated, and so suit different tasks.
| Prompt evolution (e.g. GEPA) | Full-rewrite memory (e.g. Dynamic Cheatsheet) | Playbook accumulation (e.g. ACE) | |
|---|---|---|---|
| Artifact | One instruction prompt | One cheatsheet document | Itemized bullet context |
| Update | Propose whole-prompt variants from rollouts + reflective feedback; select by evaluator score under a rollout budget | Regenerate the whole cheatsheet each step, or write a fresh summary from retrieved examples | Curator writes only new insights; deterministic delta merge |
| Typical win | Short, transferable instructions | Short reusable heuristics and code snippets on independent single-turn queries | Many granular rules, procedures, edge cases retained over long horizons |
| Characteristic risk | Brevity bias | Context collapse | Unbounded growth; needs grow-and-refine |
Selection heuristics:
- Independent, single-turn reasoning (the AIME, Game-of-24, and GPQA setting where Dynamic Cheatsheet was mainly evaluated): reported gains there came from short reusable heuristics and code snippets. A compact evolved prompt or short cheatsheet may be enough. ACE did not test this setting head-to-head.
- Multi-turn agents with step-by-step procedures and tool-use rules, and knowledge-intensive domains with many specific rules and edge cases, favour accumulation. The evidence is AppWorld, FiNER, and Formula, plus single extra benchmarks for medical diagnosis and Text-to-SQL. The authors argue such knowledge is hard to compress into one instruction without losing detail.
- Need per-rule bookkeeping (which entry helped or hurt, targeted refinement, de-dup, stable earlier rules across a long run)? An itemized representation supports this directly. Whole-prompt candidates are optimized end to end, so there is no built-in per-rule attribution.
Agent-memory systems sit near the accumulation end with different units: reusable workflows induced from trajectories (Agent Workflow Memory), cached plan templates for cheap re-execution (agentic plan caching), linked Zettelkasten-style notes (Zettelkasten agent memory notes), and composed skills (skill library). Playbook accumulation generalizes the idea beyond memory to system prompts and evidence.
Mechanics of the prompt-evolution side. GEPA (Genetic-Pareto) collects execution traces, uses natural-language reflection to diagnose errors and assign credit, and proposes prompt updates. A genetic search keeps a Pareto frontier of high-scoring prompts to avoid local optima. MIPROv2 instead jointly optimizes instructions and demonstrations by Bayesian optimization (see instruction vs demonstration optimization).
Measured gap (DeepSeek-V3.1, offline, labels available). On AppWorld, ACE beat ICL by 12.3 and GEPA by 11.9 average points; MIPROv2 was not run there. On FiNER and Formula, it beat ICL, MIPROv2, and GEPA by 10.9 points on average. The authors attribute the gap to tasks that need precise domain knowledge and tool rules beyond what fixed demonstrations or one optimized prompt hold. On BIRD-SQL the gap was small (+5.1 vs GEPA +4.4), and GEPA gained more on harder queries.
When a detailed playbook is not worth it:
- Multi-hop QA such as HotPotQA often benefits more from concise, high-level instructions (how to retrieve and synthesize evidence) than from a long context.
- Fixed-strategy tasks such as Game of 24 may need a single reusable rule. Extra context there is just redundancy.
Accumulation pays off when a task needs detailed domain knowledge, complex tool use, or environment-specific strategies beyond what the model's weights or a simple system prompt already provide.
What the artifacts look like in practice (AppWorld). All three methods share one agent scaffold: a supervisor framing, a Python REPL, API-discovery calls, and standing rules such as "read the API spec before calling", "verify before irreversible changes", and "loop over page_index". They differ in what gets added:
- Evolved prompt (GEPA): the scaffold followed by hard-coded "Domain-Specific Strategy for … Tasks" sections (bill splitting, file organization, playlists, alarms). These are task-family procedures baked into one static prompt, with no per-rule identity. Presumably, covering a new task family would take another whole-prompt evolution round.
- Cheatsheet (Dynamic Cheatsheet): the scaffold plus a delimited
CHEATSHEETblock and fixed guidance on using it (analyze and pick strategies, explain before concluding, write self-contained code without hard-coded local paths). The cheatsheet content is regenerated wholesale. - Playbook (ACE): the scaffold plus a delimited playbook of ID-tagged bullets, with an instruction to use only the relevant parts. It grows by delta.
Sources
- ACE (Zhang et al.), §2.2, §4.2–4.4, §5, App. B–C, App. F (Figs. 6–9)