Tokenless

Tokenless provides tooling and expertise to find and fix flaws in reinforcement learning environments, evaluation tasks, and graders. As models become more capable, training and evaluating them requires increasingly complex tasks and longer interactions, which creates more opportunities for unexpected behavior and exploits. The company states that its goal is to end reward hacking, since training can amplify small flaws in an environment as models learn to exploit them, making it difficult to distinguish real capability gains from better exploitation.

The offering combines tooling with hands-on expertise to understand and prevent these failures. Correcting such flaws is described as making training more effective and model performance easier to interpret, and as important to solving the AI alignment problem. The company links to its research for further detail.

Tokenless is aimed at teams training models with reinforcement learning, building environments, or running agent evaluations who want more confidence in their environments. Interested parties are directed to contact the company; no free plan, trial, editions, or seat-based or usage-based packaging is described.