Unverified50% confidenceFactExact time
ImpossibleBench was designed in 2025 by Zhong, Raghunathan, and Carlini to measure how models trade off between specifications and tests when the two conflict.
1
Sources
50%
Confidence
Long-term
Relevance
10/11/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedCollision detection algorithms (such as GJK or SAT) can perform geometric intersection operations between the robot's own 3D model and the bounding box of the target space to calculate a feasibility confidence score.54% similarUnverifiedIn the Simulate step, three levels of test scenarios are written for each Agent skill: Nominal (happy path), Edge (edge cases), and Chaos (chaotic conditions).54% similarUnverifiedThe test task was to build a 3D dragon from scratch using Three.js.52% similarUnverifiedHaving multiple agents compete against each other is a more realistic stress test that produces edge cases that preset scripts cannot cover.52% similarUnverifiedThe tests designed three real-world scenario tasks: Kanban application development, Blender lantern scene modeling, and a Godot 3D platformer game, with the prompts for Blender and Godot not publicly disclosed and not having been trained on.51% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/998199API
curl https://kongchang.com/api/v1/knowledge/claims/998199MCP
get_claim(id=998199)