SWECCathon 2026

Devpost hackathon: SWECCathon 2026. Imported automatically.

Winners

RhetBench1st Place OverallRhetBench — a benchmark testing whether AI agents can persuade an NPC on a real-world topic. The character's hidden personality determines which persuasion straFastAPIPython

Sokoban2nd Place OverallCan an LLM reason about physical space and consequences? Sokoban: push boxes, climb ramps, reach the goal, but one wrong move soft-locks the level. Understand yJavaScriptPython

Efficient Scientific DiscoveryFuturistic Track WinnerEvaluate whether agents can identify hidden scientific laws by selecting informative experiments under budget and observation noise.GeminiJavaScriptPython
JengaBenchGames Track WinnerEvaluates visual-spatial reasoning and strategic decision-making of LLMs in a deterministic 3D Jenga environment.DockerFastAPIJavaScript
Orbital PlannerAGI & Real World Track WinnerPlans impulsive burns to rendezvous with moving satellites or moons in 3D Earth orbit; the Mesocosm AI agent explains strategy and outcomes; a deterministic phyJavaScriptPython
UnderlieFuturistic Track WinnerA benchmark to test AI's ability to create emotions.Python

Citation TrapHonorable Mentions!Right answer, invented sources. Citation Trap is a benchmark that scores LLM citation faithfulness vs. answer correctness across 500 questions, catching models AnthropicFastAPIGemini
Clue-LessHonorable Mentions!Benchmarking LLM's abilities and approaches towards solving a non-deterministic cryptic puzzle called MinuteCryptic.Python
PikachuHonorable Mentions!N/APython