last week i was playing around with pre-filling reasoning traces in deepseek to see if i could steer its thinking towards better responses anecdotally this seemed to work so i got with Dani Balcells to figure out how we could turn this into a more rigorous experiment...

TextArena is live on arXiv! We present a benchmark of 57+ competitive text-based games to evaluate and train LLMs on agentic behavior — including negotiation, deception, theory of mind and many more. Real-time TrueSkill. Multiplayer support. Human-vs-models. Model-vs-model.

Minimum vibable product.

Super cool work by León and the TextArena team! Very excited to have collaborated on this :)