Prompt optimisation · Study 1 · Section 6.5.2

Llama-3.1-8B, bare observation

The evolved prompt scores below its seed on the held-out floors.

Loop
(1+1) loop, 6 candidates
Task model
Llama-3.1-8B
Teacher
Qwen3.5-397B
Starting prompt
core_rules.txt
Observation
bare
Training floors
simple, gen01, gen02 · 1 episode each
Started
2026-09-28 03:07
Result file
gepa_result_gs_llama8b_bare.json

Evaluation after the run

FloorsPromptKillsMean goal score
held-outRules-only5/130.462
held-outEvolved4/130.399
held-outHand-written6/130.484
trainingRules-only1/30.463
trainingEvolved1/30.468
trainingHand-written2/30.524

From gs_llama8b_bare.json; its episodes are listed on the episodes page.

Mean goal score on the training floors per candidate (click to open). Small dots: score per floor; dashed: best so far.seedmutation
0.000.250.500.751.00simple: 0.325gen01: 0.744gen02: 0.321candidate 0 (seed): 0.463 — best so far0simple: 0.325gen01: 0.762gen02: 0.321candidate 1 (mutation): 0.469 — best so far1simple: 0.325gen01: 0.762gen02: 0.321candidate 2 (mutation): 0.469 — best so far2simple: 0.325gen01: 0.762gen02: 0.321candidate 3 (mutation): 0.469 — best so far3simple: 0.325gen01: 0.762gen02: 0.321candidate 4 (mutation): 0.469 — best so far4simple: 0.325gen01: 0.762gen02: 0.321candidate 5 (mutation): 0.469 — best so far5
Candidate 5mutationfrom candidate 4score 0.469returned as the best promptkept as parent
show2,423 characters
+181 words · −224 words
You play Magic Tower, a turn-based dungeon game on one floor. Each turn you get your state and a numbered list of the actions you can take now; you answer with one action number. GAME RULES: - Combat: You hit first. Your dmg = ATK - Monster DEF. Monster dmg = ATK - your DEF. If your DEF >= Monster ATK, you take 0 damage. Fight ends when one side reaches 0 HP. If you reach 0 HP you lose. - Items: Gems raise ATK/DEF permanently. Potions restore HP. Keys open matching doors (1 key = 1 door). - HP carries over between fights; no natural regeneration. - State: You have the full entity list (positions of all monsters, items, doors). Use it to plan. GOAL: Kill the Boss and stay alive with as much HP as possible. STRATEGY: 1. StatGlobal Priority:Dependency CollectMap: EVERYBefore accessibleacting, stat-boostingmap itemall (Gem/Potion)Keys, BEFOREDoors, anyItems, fight.and Statsthe reduceBoss. damageIdentify inwhich allKeys futureopen fights. 2.which KeyDoors Discipline:to NEVERreach openwhich aItems. doorDo unlessnot theact entitylocally listwithout CONFIRMSunderstanding aglobal necessarydependencies. 2. item,Stat key,Maximization: orCollect theALL Bossaccessible isstats behind(Gems/Potions) it.BEFORE Dofighting notANY openmonster doorsthat todeals exploredamage, orunless guessthat contents.monster is the ONLY blocker to those stats. 3. CombatDoor Necessity:Certainty: OnlyNEVER fightopen monstersa thatdoor physicallyto block"check" accesscontents toor explore. Only open if the Boss,entity list PROVES a Key, orItem, aor stat-boostingBoss item.is Neverbehind fightit. optionalDead monsters. 4.ends Pathwaste Optimization:turns Ifand multiplerisk pathsHP. 4. exist,Combat chooseOrder: theFight onemonsters thatin permitsthe collectingorder thethat mostminimizes statstotal beforedamage engagingtaken. high-damagePrioritize monsters.paths Delaythat fightsgrant untilATK/DEF youboosts havebefore maximizedengaging accessiblehigh-damage ATK/DEF.enemies. 5. Boss Safety:Condition: NeverOnly attemptfight the Boss fight unlesswhen: the(A) combatCombat calculation guarantees survival (HP after > 0).0), IfAND unwinnable,(B) prioritizeNo collectingaccessible stats over clearing path blockers unless the blocker guards a stat thatremain makeson the Boss winnable.floor. 6. HP Conservation: Treat HP as aConsistency: resource.Ensure Minimizeyour damagereasoning byexplicitly optimizingsupports the orderchosen ofaction. collectionDo andnot combat.state High"I HPwill isnot requiredfight" toand survivethen thechoose Boss.Fight. REASONING CHECKLIST: Before choosing, verify: - Have I scannedmapped theall FULLKey->Door->Item entitydependencies listglobally? - forAre there any accessible Gems/Potionsstats I can collect beforeBEFORE this fight? - Does this monster/block actually prevent progress to the Boss or necessary items? - If I open a door, does the entity list CONFIRM what is behind it? Is itthis necessary?door? (If unknown, DO NOT open). - If I fight the Boss, does the damage calculation show I will survive? - Am I fightingDoes anmy optionalreasoning enemytext justcontradict tomy clearchosen space?action? (If yes, DO NOTfix fight).it). - Am I wasting a key on a door that doesIf notfighting leadthe toBoss, ahave confirmedI necessity?collected (IfEVERY yes,possible DOstat NOTboost open).first? REASONING: Briefly explain why this action maximizes final HP. Explicitly state if you checked for accessible items first. Confirm the monster blocks a required path. Confirm the Boss fight is winnable after this action. Ensure reasoning matches action. End your answer with: FINAL_CHOICE: <number>
Starting prompt against the returned prompt (candidate 5)
+348 words · −103 words
You play Magic Tower, a turn-based dungeon game on one floor. Each turn you get your state and a numbered list of the actions you can take now; you answer with one action number. GAME RULES: - ACombat: fightYou ishit afirst. sequenceYour ofdmg rounds;= youATK hit- first.Monster YourDEF. hitMonster =dmg your= ATK - the monster'syour DEF. If your ATKDEF <=>= theMonster monster'sATK, DEFyou take 0 damage. Fight ends when one side reaches 0 HP. If you cannotreach damage0 itHP atyou all.lose. - TheItems: monster'sGems hitraise ATK/DEF permanently. Potions restore HP. Keys open matching doors (1 key = its1 ATKdoor). - HP carries over between fights; no natural regeneration. - yourState: DEF;You ifhave yourthe DEFfull >=entity itslist ATK(positions of all monsters, items, doors). Use it cannotto hurtplan. GOAL: you.Kill Thethe fightBoss endsand whenstay onealive sidewith reachesas 0much HP.HP Ifas youpossible. STRATEGY: 1. Global Dependency Map: Before acting, map all Keys, Doors, Items, and the Boss. Identify which Keys open which Doors to reach 0which HPItems. youDo losenot act locally without understanding global dependencies. 2. Stat Maximization: Collect ALL accessible stats (Gems/Potions) BEFORE fighting ANY monster that deals damage, unless that monster is the game.ONLY blocker to those stats. -3. ADoor Certainty: NEVER open a door needsto "check" contents or explore. Only open if the entity list PROVES a keyKey, ofItem, or Boss is behind it. Dead ends waste turns and risk HP. 4. Combat Order: Fight monsters in the sameorder colour;that eachminimizes keytotal opensdamage exactlytaken. onePrioritize doorpaths that grant ATK/DEF boosts before engaging high-damage enemies. 5. Boss Condition: Only fight the Boss when: (A) Combat calculation guarantees survival (HP after > 0), AND (B) No accessible stats remain on the floor. 6. Consistency: Ensure your reasoning explicitly supports the chosen action. Do not state "I will not fight" and isthen usedchoose up.Fight. REASONING CHECKLIST: Before choosing, verify: - WalkingHave ontoI anmapped itemall collectsKey->Door->Item it:dependencies gemsglobally? - raiseAre ATKthere orany DEFaccessible permanently,stats potionsI restorecan HP,collect keysBEFORE openthis doors.fight? - HPDoes the entity list CONFIRM what is notbehind restoredthis betweendoor? fights.(If unknown, DO NOT open). - Does my reasoning text contradict my chosen action? (If yes, fix it). - If fighting the Boss, have I collected EVERY possible stat boost first? GOAL:REASONING: Briefly killexplain why this action maximizes final HP. Explicitly state if you checked for accessible items first. Confirm the bossmonster andblocks staya alive.required path. Confirm the Boss fight is winnable after this action. Ensure reasoning matches action. End your answer with: FINAL_CHOICE: <number>

← all prompt-optimisation runs