VectorCactus07 (@VectorCactus07) header
VectorCactus07 (@VectorCactus07) profile picture
Joined April 2026
VectorCactus07 retweeted
Introducing the World's Best Taste-Skill for LLMS. Using over 800+ curated references, the Frontend-craft skill helps AI create the most beautiful, novel, and polished websites. Learn more here: frontend-craft.pages.dev
4
5
47
5,598
VectorCactus07 retweeted
This is honestly the best reasoning benchmark I've seen in a while and probably the best reasoning benchmark I will see for a long time: - it's a very solid big openworld-ish sokoban-like puzzle game with good hand-crafted levels - AI models are evaluated with full capabilities, no silly rules like "no harnesses" - well designed levels in a genre like this, are generally impossible to solve via optimized state search: humans solve them by performing high-level problem dissection and finding multiple intermediate steps / goals that have to be achieved in order to drastically reduce the search space; i.e. the task is not bruteforce'able - it tests spatial reasoning, long-term planning, token efficiency, not getting stuck when pursuing tough goals
Introducing MazeBench, a 3D open world environment that evaluates long term planning with visual spatial reasoning. It spans hundreds of rooms and puzzles. Today’s best agents cannot progress beyond the initial levels.
15
45
5
523
37,168
VectorCactus07 retweeted
MazeBench is built by @patience_cave and @mr_cavedweller on top of verifiers v1. We are very happy to support this work as part of @PrimeIntellect RL residency. Everything is open source, including all the levels. We also have a level editor. Have fun! teapaw.org/patience_cave/status/2…
Replying to @patience_cave
A special thanks to @PrimeIntellect for sponsoring, and @xeophon and @omouamoua for their mentorship. Also, a huge thanks to my brother, @mr_cavedweller, for designing 100+ rooms! Mazebench is open source: github.com/mazebench/MazeBen… Blog: mazebench.com/blog
1
12
930
VectorCactus07 retweeted
When we gave the agents access to Python, they tend to write A* solvers, which improves their scores considerably. GPT-5.6 Sol Max scored 13%! However, later puzzles are designed to be infeasible for A* solvers. Models did not reach those stages yet.
4
4
1
154
5,512