From Reactive Control to Holistic Planning: Evolutionary Prompt Optimization for Hybrid Deterministic-Neural Agents in Long-Horizon Environments
King Wai Siu · supervised by Prof. Dr. Ricardo Usbeck and Prof. Dr. Burkhardt Funk
A language model plays Magic Tower, a deterministic game decided by the order of fights, doors and items. A deterministic engine lists the legal actions with their exact costs, the model chooses one, and an exact solver knows each floor's optimum. This site holds the evidence: every result file, every episode with the model's reasoning at each turn, and the route from the first experiments to the final configuration.
The thesis in three minutes
Captions are in the picture. The interactive version pauses at each demo so you can try it, and links the captions file and the voice-over script.
Main findings
15 / 15simple-floor episodes won by Qwen3.5-397B with the engine's annotations, each at 2 HP (optimum 13)the re-run cell
0/5 → 10/10after two annotation lines that state both sides of a trade-off; same model, same promptthe repair sequence
MT7 → MT18highest tower floor across three observation designs; a resumed run reached the normal ending on MT21the tower runs
Start here
The first decision on the simple floorFour legal actions; only one keeps the floor winnable. See what Qwen3.5-397B saw and why it chose the Black Slime.open the replay Where the winning trajectory leaves the optimal routeTurn 15: the Slime Lord before the Bat. The episode still wins, at 2 HP instead of the optimum 13.open the replay The tower run that reached the normal endingA stored run resumed on MT18 with blue keys for sale, played on to the Nether Demon King on MT21.open the replay
Explore
The thesis in three minutesOne animated story from the game to the tower, with demos to pause and try; also as a video.watch the walkthroughPresentationThe defense slides with live demos: the exact solver, real decisions with the model's reasoning, every result cell.open the slidesGuided tourEleven short chapters from the game to the conclusion, each with something to try.start the tourJourneyFrom single scored episodes to the final configuration, with the readings that did not survive.26 milestonesResults177 cells behind the thesis tables, each with its settings, episodes and the command that re-runs it.browse by thesis tableEpisodes1,905 replays: the map, the observation, the reasoning and the answer at every turn.filter by floor, model, outcomeCompareA win and a loss side by side, opened at the first turn where they act differently.compare two episodesBFS solverThe search that gives every floor its optimum, step by step, and why the published one reported 22 instead of 68.step through the searchPrompt optimisationEvery evolved prompt, the teacher's diagnosis, the validator's verdict and the held-out result.candidate by candidatePlayAny single floor on the thesis engine, with the exact solver's value of every action.play a floorTower GameThe original 24-floor game in English, in the browser, and every tower run with the highest floor it reached.play the towerThesisThe full text with linked references, the result cells behind each table, and the PDF.read the thesis